Most enterprises have never heard of Mercor, Applied Compute, or Fleet. But when Cursor spent 85% of its training compute on post-training and RL rather than pretraining, it showed the importance of this layer for modern AI teams. The good news is that you don’t need to be a $60B coding startup to access RL Gyms. In fact, there’s now a thriving ecosystem of companies that enterprises can readily use.

This is Part III of the Owning Your Intelligence series. Part I made the case: the era of renting intelligence from frontier APIs is ending, and the enterprises that own their own models will compound faster than those that don’t. Part II showed how to do it — the post-training playbook, the role of RL, and the domain model builders who’ve already run this experiment and won. Part III maps the companies you need to know. It introduces six key categories spanning AI infrastructure, training specialists, and bespoke AI training data providers that make intelligence ownership possible.

The market behind the map

A few statistics worth knowing before diving in:

• AI training data is a $3.4 billion global market in 2025, projected to more than double to $8.3 billion by 2030 — a category that barely existed a few years ago.

• Reinforcement learning infrastructure sits at $12–15 billion today and is expanding at 28–35% annually. The RLHF platform segment specifically — the pipelines that turn human feedback into reward signals — is projected to grow from $2.8 billion to $18.6 billion by 2034.

• Enterprise post-training adoption has crossed a meaningful threshold: 67% of Fortune 500 companies have already completed or initiated at least one domain-specific fine-tuning project. The question is no longer whether enterprises will post-train their own models. It’s who in this ecosystem they’ll use to do it.

For a recap of the benefits of owning your intelligence, check out Part 1 of my series below:

Owning Your Intelligence Market Map

Owning intelligence isn’t a weekend do-it-yourself project. It requires serious tools and infrastructure. That’s why a fast-growing ecosystem of startups provides the infrastructure, training environments, real-time knowledge systems, and foundation models that let enterprises post-train and deploy custom AI without building a frontier research lab.

Introducing: The Market Map

Breaking Down the Map: Six Key Categories

Menlo Ventures recently catalogued the AI training data and RL environments market: more than 50 companies generating roughly $8.5 billion in combined annual revenue. More than 75% of that revenue flows to just four players — Scale AI, Surge AI, Mercor, and Handshake — but the remaining $2+ billion is spread across dozens of specialized infrastructure companies building what the next decade of AI will actually run on. While the overall landscape is too large to capture perfectly, the six categories below encompass most of the major players.

1. Inference & Serving Infrastructure

The layer that runs deployed custom models at scale.

Fireworks AI — The infrastructure story of the year: $315M in annualized revenue as of February 2026, up 416% year-over-year. Reportedly in talks to raise at a $15B valuation, a 3.75x jump from its $4B Series C just seven months earlier. The company offers training and inference under one roof. CEO Lin Qiao’s thesis: “Not just consume intelligence. Own it.”

Baseten — Infrastructure hardening into a category. $300M Series E at a $5B valuation in January 2026; Q1 2026 revenue tripled in a single quarter to roughly $600M ARR. That round closed in June 2026 as a $1.5B Series F led by Altimeter Capital, Conviction, and Spark Capital, priced across two tranches at $11B and $13B. This tripled Baseten’s valuation in less than five months.

Altimeter’s Apoorv Agrawal, who led the investment, frames the thesis around inference demand: “Post-trained open source models deliver the best combination of capability, cost and control.” Built for high-throughput production serving, especially agentic workloads at scale.

Modal — $300M raised at a $5B valuation, with $300M in ARR. Developer-first and maximally flexible — the choice for engineering teams that want control over how their models run in production without trading that control for a managed service.

2. RL & Post-Training Specialists

The teams that run the training process.

Scale AI — The dominant platform for AI training data and RLHF pipelines: the layer that turns raw human feedback into the reward signal post-training runs on. $1.5B in annualized revenue at a $29B valuation as of 2026. Most frontier labs and a growing number of enterprise post-trainers run on Scale’s data infrastructure.

Surge AI — One of the largest human-in-the-loop training data providers, running contractors and annotators at the scale frontier labs need to run RLHF. ~$3B in annualized revenue at a ~$25B valuation as of mid-2026 — making it one of the fastest-scaling companies in the AI training stack.

Mercor — AI-native talent marketplace that matches and manages the human contractors who run reinforcement learning pipelines. $2B+ in annualized revenue, with the company in talks to raise at a $10–20B valuation. From $1M to $1B ARR in roughly 20 months — one of the fastest revenue ramps in the category.

Recursive and Thinking Machines — Building RLaaS (Reinforcement Learning as a Service) platforms, the abstraction layer that lets enterprise teams run post-training workflows without standing up their own RL infrastructure. Both early-stage but backed by investors who treat RLaaS as a distinct infrastructure category rather than a feature of the inference stack.

Judgment Labs — Agent evaluation and observability infrastructure for post-trained models: the tooling that lets a team build the kind of private eval rubric this piece argues every enterprise needs before it can trust a custom model in production. Raised a $32M Series A led by Lightspeed in May 2026.

3. RL Training Environments (”AI Gyms”)

Platforms that give agents a safe place to learn before production.

Deeptune — Builds high-fidelity simulation environments where AI agents practice complex digital workflows before touching production — a flight simulator for enterprise AI. Closed a $43M Series A led by Andreessen Horowitz in January 2026.

Fleet — Builds RL environments for messy enterprise workflows specifically: legacy systems, inconsistent data, and the edge cases that synthetic training environments tend to miss. Now at $100M+ ARR with a $725M implied valuation — the fastest-growing pure-play AI gym.

Mechanize — Builds reinforcement learning environments designed to automate complex knowledge work at scale, training agents on the kinds of tasks that have historically required human judgment.

Applied Compute — Embeds RL engineers directly with enterprise customers, turning product usage data and eval traces into continuously improving models. Decentralizes intelligence by training smaller, highly customized, cost-effective models on a company’s own data. $1.3B valuation.

Trajectory — Rumored to be raising a $30M Series A led by Sequoia to build a continual learning platform. Similar to Applied Compute, company’s central thesis is that post-training open source creates a path toward more performant, cost-effective and interpretable agent systems, as open-weight models can be hosted within a firm's own secure infrastructure and trained on a company’s own workflows.

4. Real-Time Knowledge (Retrieval & Search)

The core infra enterprise AI tools need to use the web.

Parallel — Powerful API for deep research, data retrieval, and querying the web like a structured database — built for agents that need to reason over live, structured external data at inference time. Founded by former Twitter CEO Parag Agrawal.

TinyFish — End-to-end platform for autonomous AI agents, providing fast full-stack search capabilities so agents can retrieve and act on current information without leaving the inference loop.

Exa — Neural search at web scale, purpose-built for LLMs rather than human searchers. Search API built for agents. Backed by a16z.

Firecrawl — Web crawling and scraping infrastructure for teams that need to pipe fresh external data into AI workflows continuously.

5. Domain Model Builders

The Cursors and Harveys pioneering the playbook.

Cursor (Composer) — Post-trained Kimi K2.5 into Composer, spending 85% of total compute on post-training and RL rather than pretraining. Composer 2.5 matches Claude Opus 4.7 on coding benchmarks at one-tenth the cost.

Harvey — Post-trained a 27B open-weight model against its proprietary Legal Agent Benchmark (1,200+ tasks, 24 practice areas, all-pass grading), working with Baseten Research. Landed inside the closed-source frontier band at a fraction of frontier cost.

Ramp — One of the first enterprise companies to run an own-vs-rent playbook at scale, applying custom post-trained models to expense categorization, financial document analysis, and workflow automation.

Genspark — RFT-trained a Kimi K2 variant that beat SOTA on quality and tool-calling accuracy: +12% quality at roughly half the inference cost of the closed-model baseline it replaced.

Vercel (v0) — RFT-trained an auto-fixer for AI-generated code, hitting a 93%+ error-free rate at roughly 40x the speed of the general-purpose model it used to lean on.

6. Open-Source Foundation Models

The starting points for post-training.

Kimi K2.5 — Moonshot AI’s one-trillion-parameter mixture-of-experts model (~32B active parameters per inference pass). Cursor’s foundation of choice; strong on coding.

Qwen 3.5 — Exceptional reasoning, 1M-token context window.

DeepSeek V3.2 — Near-frontier performance on most benchmarks.

Llama 4 — Broadest commercial ecosystem and licensing familiarity.

Why This Matters

The landscape above isn’t your typical squishy 10-year VC projection. This is a live report from what is happening right now in the field.

These companies exist today and are scaling at unheard of growth rates. Surge AI crossed a $3B run rate without raising a single dollar from VCs. Baseten grew its valuation 3x in only five months. Fleet went from launch to $100M ARR in a matter of months. Mercor is circling a $20B valuation as we speak. And that’s just the tip of the iceberg. The infra to post-train, run, and own custom AI models is readily available to any enterprise that is ready to use it. That simply wasn’t the case even 18 months ago.

What’s left is deciding whether you’re ready to jump. Companies like Cursor, Harvey, Rogo and Ramp are all in and now operate models that are better and cheaper than off-the-shelf alternatives. Importantly, these companies have proven that you can build these competitive advantages while still being happy frontier AI lab customers. Their Claude bills are just significantly lower than they would be otherwise.

The infrastructure is ready. The only question is: are you ready to close your Claude terminal and enter the arena?

Next time (The State of AI Adoption in the Enterprise - Q2 Review): Back by popular demand! A follow up to our Q1 2026 Report: Do 95% of enterprise AI pilots really fail? See below for my Q1 Report on Enterprise AI adoption.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Views expressed are my own and should be considered as such, and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NYSE: NEE).

Scaling the Enterprise · Owning Your Intelligence, Part III, Vol. 2

Last week, we made the case that renting frontier intelligence is no longer the default move. That’s because open-source models have reached frontier quality at the same time frontier token prices have grown astronomically. The math no longer works.

Agentic workloads have made frontier API economics untenable. Open source is excellent. And the Mythos shutdown has highlighted the risk of relying on private companies for intelligence.

This week, we cover what enterprises can do about it: migrate to cheaper open models and regain control of your intelligence, your workflows, and your gross margin.

Post-Training Open Models is So Back

After a brief hiatus the past few years where model improvements from the AI labs made building your own models pointless and wasteful, the tide has shifted.

In short, the idea of post-training open source models is coming back. Previously, this strategy had been rejected on the basis that general models alone would become so good they would make such niche and customised approaches pointless.

But recently, things have changed. And many companies have proved it can work.

Take Baseten for example. The AI inference platform raised a massive $1.5 billion Series F in June 2026, achieving a post-money valuation of $13 billion. Altimeter's Apoorv Agrawal, the partner who led the firm's investment, laid out his thesis in a post titled Why we are doubling down on Baseten.

His thesis is remarkably simple:

• He believes inference will be the largest market

• He believes post-trained open source models deliver the best combination of capability, cost and control

• Baseten provides the entire model supply chain to harness open source. Train, deploy, and serve models all on the same platform

Of particular interest to this post is the bolded claim above about how that demand actually gets served: "Post-trained open source models deliver the best combination of capability, cost and control." Baseten's own numbers back that up — the company says its fastest-growing customers now direct 30-50% of total model spend toward custom, post-trained models built on their own data, rather than frontier APIs alone.

Agrawal points to Harvey as the proof case: a legal foundation model series built on Baseten's platform, inspired by Cursor's Composer, designed to serve frontier intelligence at an affordable price while building toward law firms owning their own models outright. This is the very playbook we’ll describe for you below.

The Companies Already Doing It

The best way to understand how owning intelligence works is to look at the companies already doing it — and what their approaches have in common.

1. Cursor + Composer

Cursor started with Kimi K2.5 — Moonshot AI’s open-source one-trillion-parameter mixture-of-experts model, with approximately 32 billion active parameters per inference pass. That’s the foundation. Everything else is Cursor’s.

The defining choice: 85% of total compute went to post-training and reinforcement learning, not pretraining. They ran thousands of agent rollouts in parallel, scoring each against private eval rubrics and updating the model toward what worked. Training data wasn’t benchmarks — it was actual Cursor work: agentic coding, multi-step debugging, refactoring across large codebases.

The result: Composer 2.5 matches Claude Opus 4.7 on coding benchmarks at one-tenth the cost. No competitor can replicate that by calling an API.

2. Harvey + Fireworks AI, Then Baseten Research

Harvey ran two experiments with two infrastructure partners. Both worked.

First, with Fireworks AI, Harvey fine-tuned Kimi K2.5 as the primary “worker” model and made Claude Opus 4.7 a callable advisor for only the hardest sub-tasks. Measured across 100 tasks on Harvey’s Legal Agent Benchmark: $84 versus $954 for Opus alone. Eleven times cheaper, not eleven percent.

Harvey then went deeper with Baseten Research — a post-training pipeline built around their Legal Agent Benchmark (LAB): 1,200+ tasks across 24 legal practice areas, expert-written rubrics, all-pass grading. Partial completion counts as failure.

Before training, they built a compaction harness — periodic context compression into structured memos so the agent can process long documents without losing prior findings. What emerged after training was telling: the post-trained 27B model adopted end-to-end document reading, not because Harvey programmed it, but because the reward signal made comprehensive reading the winning strategy.

Result: a 27B open-weight model inside the closed-source frontier band, at a fraction of frontier’s $50-per-task cost. The core lesson: the model and the agent harness must be developed together. The same training run without the harness produced minimal lift. With it: frontier-level performance.

3. Ramp

One of the first enterprise companies to run an own-vs-rent playbook at scale. Expense categorization, financial document analysis, and workflow automation are exactly the workloads where post-trained models outperform frontier API calls: real company data, clear evaluation criteria, predictable workflows. The raw material was already there. Ramp used it.

4. Genspark

Genspark’s Deep Research agent hit a ceiling on a closed-source model — prompt engineering was the only lever, and it couldn’t deliver the tool-calling behavior a mixture-of-agents architecture requires.

Working with Fireworks AI, Genspark RFT-trained Kimi K2 against their own reward function. Four weeks later: +12% quality, +33% tool-call volume, roughly half the cost. CTO Kay Zhu: “Fireworks enabled us to own our AI journey, and unlock better quality in just four weeks.” The post-trained model didn’t just match the frontier. It beat it.

5. Vercel

Vercel’s v0 runs as a composite model — retrieval, a reasoning LLM, and a custom auto-fixer that corrects errors mid-generation. That last piece started on Gemini Flash 2.0 and hit the same wall every closed-model does: prompting was the only lever.

Vercel RFT-trained the auto-fixer with Fireworks and paired it with speculative decoding. Result: 93%+ error-free generation rate, beating Claude Opus 4, Sonnet 4, Gemini 2.5 Pro, and GPT-4.1 on the same benchmark at roughly 40x the inference speed of GPT-4o-mini. CTO Malte Ubl: “The SOTA in this space changes every day, so you don’t want to tie yourself to a single model.”

The wrinkle Vercel adds: you don’t need to own the whole pipeline. Owning one narrow, high-volume sub-task was enough to beat frontier across the board.

Subscribe now

6. Plus More!

There are many other examples I could share, but you get the point. Moving on…

Step-by-Step: How to Build Your Own Intelligence 101

Friends, below you’ll find a step-by-step guide on how modern enterprises can build their own intelligence from the ground up. The future is here.

Step 1: Audit your data assets
Before you touch a GPU, answer three questions. How many real user interactions have been logged from your production AI workflows? Do you have agent eval traces — the complete chain of reasoning steps, tool calls, and outputs your agents produce in production? What are the highest-volume, highest-cost AI tasks in your product?

If you have more than a million real user interactions and a year of eval traces, you have the raw material. Most scaled AI companies do. Almost none have inventoried it.

Step 2: Pick the right open-source foundation
The tier-one open models in 2026: Kimi K2.5 (Cursor’s choice, strong on coding), Qwen 3.5 (exceptional reasoning, 1M token context), DeepSeek V3.2 (near-frontier on benchmarks), and Llama 4 (broad commercial ecosystem).

Match model size to use case. For cost-efficient production serving, 7–27B parameter models handle most domain-specific workloads. For complex long-horizon tasks — legal analysis, financial modeling, medical diagnosis — 70B+ is the right starting point. Check licensing before committing; terms vary widely.

Step 3: Build your private eval rubrics before you train anything
This is the step most teams skip, and it’s the most important one. You cannot optimize toward a target you can’t measure. Harvey’s LAB is the gold standard: expert-written rubrics defining what good output looks like for each task type, with all-pass grading that treats partial completion as failure.

Build the equivalent for your domain. Aim for at least 200 held-out evaluation tasks before training begins. The rubric writing is time-consuming. Do it anyway — it’s the lever that makes everything else work.

Step 4: Design the agent harness before training

Your model will train on the same context windows it sees in production. If the production system uses a compaction strategy — periodic context compression, tool use sequences, memory structures — train with that structure in the loop from the beginning.

Harvey’s research made this point definitively: the 27B model showed minimal lift without the harness. The same training run with the harness produced frontier-band performance. Harness design and model training are not separate workstreams. They are one.

Step 5: Post-train iteratively
Three techniques in order of complexity and payoff.

• Supervised fine-tuning (SFT) is the fastest path. Take your best agent traces, filter to rubric-passing examples, and train a base model on them. Immediate domain adaptation, minimal infrastructure investment. Start here.

• Iterative SFT is what Harvey used with Baseten Research. Run the base model through your eval tasks. For rollouts that fall short, use a teacher model to produce a corrected version and train on it. Iterate. Critical constraint: teacher hints must be invisible at inference.

• RL with domain reward is Cursor’s approach and the most powerful. Run thousands of rollouts in parallel, score each against your private rubrics, and train toward what scored highest. More compute required — but it produces domain fit that SFT alone cannot replicate.

Step 6: Evaluate relentlessly against your private holdout
Never compare your custom model to frontier models on public benchmarks. Use your private holdout, your private rubrics. The metric that matters: all-pass rate on end-to-end tasks. A model that’s 80% correct on a legal contract review is not good enough — the missed 20% is the liability.

Run evals after every training run. The trajectory tells you whether the training signal is working before you spend further compute.

Step 7: Deploy on the right infrastructure
Custom models need GPU serving infrastructure, a continuous update pipeline, and traffic routing between your custom model and frontier models for workloads where frontier still wins.

Three production-ready options: Fireworks AI offers training and inference under one roof. Baseten is built for high-throughput agentic workloads — their research arm did Harvey’s post-training work. Modal is developer-first and maximally flexible, for teams that want deep production control.

Step 8: Build continuous learning
Post-training is not a project with a launch date. It’s an engineering discipline. Your model should improve every quarter as production data accumulates, base models improve, and eval rubrics sharpen.

Treat it as a product: production data → eval filtering → retraining trigger → new model version → deployment → repeat. The companies that build this loop compound their advantage every quarter. The ones that treat post-training as a one-time project will find themselves starting over when the next base model drops.

The Ecosystem Making This Possible [Preview]

Thankfully, enterprises don’t need to build intelligence alone. There exists a growing ecosystem of startups providing the infrastructure that enables enterprises to post-train and serve their own custom models without building a frontier research lab.

(More on this next week when we cover a deep dive into the landscape)

Below are the companies making it possible for enterprises to own their intelligence without building a frontier research lab. The picks and shovels layer.

Introducing: The Owning Your Intelligence Market Map

The market map above is a preview of next week’s post, where I’ll dive deep into each of these categories. But for now, here’s what you need to know.

Sectors to Know

1. Inference & Serving Infrastructure

2. RL Training Environments (”AI Gyms”)

3. Open-Source Foundation Models

4. RL & Post-Training Specialists

5. Real-Time Knowledge (Retrieval & Search)

6. Domain Model Builders

Each sector is early. All six are growing faster than most investors anticipated a year ago. The VC dollars are flowing in: Baseten’s $13B Series F, Fireworks’ latest round potentially valuing the company at $15B, Applied Compute’s $1.3B valuation.

These valuations reflect a market that has concluded this category is real, is compounding fast, and is one with tremendous upside potential.

The Bottom Line

Owning intelligence is not a research initiative. It’s an engineering discipline, and the infrastructure to execute it now exists.

The era of renting the best frontier intelligence for every task is over. The companies that own their intelligence won’t just have lower costs. They’ll have better accuracy, better latency, and a clear path to competitive advantage.

Remember: Your intelligence. Your terms. Your moat.

• Next time (Part III, Volume III): The Market Map: Owning Your Intelligence [Deep Dive]

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Views expressed are my own and should be considered as such, and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NYSE: NEE).

Last week, Anthropic launched its two most powerful models: Fable 5 and Mythos 5.

Three days later, both were gone.

A US government export control directive, issued on June 12, ordered Anthropic to shut off access immediately. With little warning, the company was forced to shut down its two most powerful models for all users — globally, immediately, with no transition period. Anthropic didn’t choose to do this — they had no choice.

It was a moment few people saw coming, including Anthropic. They went into damage control immediately.

This is what renting intelligence actually looks like. If your product runs on Anthropic, that means it runs on someone else’s infrastructure, operates on someone else’s terms, and is even subject to someone else’s government directives. Any company operating at the application layer now faces the unsettling possibility that everything it has proudly built could disappear overnight.

Sadly, thousands of companies are still building this way.

We are entering an era of token scarcity and rapidly shrinking gains from each new frontier model release. At the same time, the smartest companies are starting to increasingly build and host their own models to power their AI tools, rather than renting general-purpose intelligence from the AI labs. In other words, we’ve reached…

The End of Renting Intelligence

For the last few years the default for AI apps was simple: call the best general-purpose model through an API, steer it with detailed prompts, give it tools, and improve behavior through evals. But times have changed.

Subscribe now

What happens when a fine-tuned open model can match frontier model quality at a fraction of the cost? Because that’s increasingly our reality today. And what happened this week with Mythos shows why that’s never been more important.

The Math That Used to Work

From 2022 to 2024, renting frontier intelligence was the right call. If you were building an AI company and didn’t build on GPT-4 or Claude, you were making a mistake. The best models were dramatically better. The open-source alternatives weren’t close. And even if you wanted to train your own model, the infrastructure to do this cost effectively didn’t exist yet.

The math worked. That’s why a recent Felicis report showed that most enterprise teams still use closed frontier models exclusively. Even at rate-card pricing, frontier API costs were manageable for most workloads. Until the tokenmaxxing era.

Fast-forward to 2026, and things have changed fast. Three things changed between late 2024 & today — and together, they may have broken the math on renting for good.

Open source caught up.

In December 2024, DeepSeek released V3 — a 671-billion-parameter open-source model matching GPT-4o and Claude 3.5 Sonnet on most benchmarks, trained for “roughly” $5.6 million versus GPT-4’s estimated $50–100 million. Weeks later, R1 matched OpenAI’s o1 on AIME and MATH-500 at roughly 30x lower inference cost. The market understood immediately: Nvidia’s stock fell 17% the following week, the largest single-day market cap loss in US stock market history.

By mid-2026, the open-source ecosystem — Kimi K2.5, Qwen 3.5, DeepSeek V3.2, Llama 4 — matches or exceeds frontier performance on coding, legal reasoning, and general intelligence. That’s why you’re seeing companies like Harvey, Rogo, and Cursor start to invest heavily into fine-tuning their own models for certain tasks.

Frontier progress has slowed to incremental gains.

For three years, each new frontier model release felt like a step change. That era is over. Parameter growth in frontier models has slowed to roughly 5% annually — a stark contrast to the more than 100x growth between 2019 and 2021. The benchmarks that once separated frontier models from everything else are increasingly noisy.

The new reality: marginal gains at extreme and growing cost. If your strategy depends on always being on the latest frontier model, you’re running on a treadmill at 10 MPH. The gap between frontier performance and a well-tuned open-source model on your specific workloads has never been narrower.

Agentic workloads broke the token economics.

Single-shot API calls cost pennies. Agentic workflows — the kind every serious AI product is now built on — consume 10x to 1,000x more tokens per task as the agent reasons, plans, calls tools, retrieves context, and self-corrects. At rate-card frontier pricing, that’s not a margin problem. It’s an existential one, especially when employees were being rewarded only a few months ago for maxing out token spend.

All of the companies that began the year defaulting to frontier API calls for every agentic workload are running an old playbook against 2026’s cost structure. It’s no wonder why companies like Microsoft and Uber have clamped down on AI spend.

The Risk Nobody Priced: What happens if your landlord kicks you to the curb?

Getting an eviction notice is never easy. That’s exactly what happened to Anthropic.

Most AI application companies ran some version of the following risk analysis: What if our frontier API provider raises prices? What if a model gets deprecated? These are normal risks that all companies typically forecast.

But Mythos exposed a category of risk that almost nobody had in their analysis: what if the government makes the decision for you?

The Fireworks AI team put it cleanly in their post-Mythos note:

“Renting works great right up until the day it doesn’t. The landlord can raise the rent. They can change the rules. And every once in a while, for reasons that have nothing to do with you, they can tell you it’s time to leave.” - Li Qiao, Fireworks AI

For a company like Anthropic, with a revenue run rate of $47 billion, that’s valued at nearly $1 trillion ahead of its IPO, the downstream impacts on the application layer are alarming. For every company building a flagship product on a model they don’t own, suddenly, that product is one government directive away from going dark.

Subscribe now

The Companies That Are Two Steps Ahead

While most companies were still debating the question, a handful answered it by proactively adjusting to today’s new reality.

Cursor + Composer

Cursor built Composer — its proprietary coding model — on Kimi K2.5, an open-source one-trillion-parameter mixture-of-experts model. Cursor’s team spent 85% of total compute not on pretraining, but on their own post-training and RL pipeline — running thousands of agent rollouts in parallel, scoring each against private eval rubrics, and optimizing for the exact workflows their users actually run.

The training data wasn’t benchmarks or synthetic problems. It was real agentic coding tasks: multi-step debugging, refactoring across large codebases, tool-calling sequences that match what Cursor’s users do every day.

The result: Composer 2.5 matches Claude Opus 4.7 on coding benchmarks at one-tenth the cost. Not eleven percent cheaper. One-tenth. That’s not a margin improvement. That’s the difference between a business model and a business.

Harvey + Post-Training

Harvey took a different path to the same destination. Working with Baseten Research, they post-trained an open-weight model against their Legal Agent Benchmark — a proprietary evaluation suite covering more than 1,200 tasks across 24 legal practice areas. The tasks mirror real partner-level legal work, and performance is measured with all-pass grading: partial completion is treated as failure.

The research surfaced something revealing. The post-trained open model adopted the same strategy as frontier models — reading documents end-to-end, not skimming with pattern matching — not because Harvey programmed that behavior, but because optimizing against the right rubrics made comprehensive reading the best approach.

The result: an open-weight model landing inside the closed-source frontier band. Frontier models on Harvey’s benchmark run to roughly $50 per task and 20+ minutes of latency. Harvey’s post-trained model delivers comparable results at a fraction of the cost.

Both companies followed the same logic: real product usage data, domain-specific eval rubrics, and long-horizon workflows where general-purpose frontier intelligence is significant overkill. The raw material was already there. Most enterprise companies have it. Almost none are using it.

What Owning Intelligence Actually Buys You

The case for owning intelligence isn’t primarily about cost — though the cost argument stands on its own. It’s about control.

Cheaper. An open-weight model post-trained on your domain can reach frontier-equivalent performance at a fraction of API cost. Cursor’s one-tenth-of-Opus economics. Harvey’s fraction of $50-per-task frontier pricing. These aren’t incremental improvements. They’re structural.

Faster. When you control the model, you control the inference stack. You can optimize latency for your specific workflows, tune serving parameters for your traffic patterns, and cut context windows without waiting for an API provider’s roadmap.

Your IP. The model trained on your proprietary data, your eval rubrics, and your domain workflows is uniquely yours. It cannot be replicated by a competitor who calls the same API you do. For the first time in the AI application era, the application layer has a defensible, compounding moat.

Sovereign. No government directive shuts down your custom model. No lab policy change alters what your model will and won’t do. No pricing reprice forces an emergency renegotiation. When your intelligence is yours, someone else’s regulatory situation is not your operational crisis.

The Bottom Line

By 2027, owning your intelligence won’t be contrarian. It’ll be status quo.

The companies that figured this out earliest are sitting on a massive advantage that is likely to compound quarterly. Their models improve as more user data accumulates. Eval rubrics sharpen. Inference costs fall as they optimize for their specific workloads. The gap between them and competitors still renting — still running on someone else’s terms — widens every month.

Here is the smart money story across this series so far:

• Part II: hyperscalers are buying watts at the infrastructure layer.

• Part III: smart application-layer companies are building their own models.

The companies doing neither — renting frontier intelligence, with no infrastructure layer moat, and no proprietary model — will be exposed.

• Next time (Part III, Volume 2): How to Own Your Intelligence Like a Pro — step-by-step guide and market map of the ecosystem building the infrastructure to make it happen.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Views expressed are my own and should be considered as such, and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NYSE: NEE).

Key Findings

• Anthropic and OpenAI together are valued at roughly 2x the combined market cap of the 10 largest US investor-owned utilities. The market is pricing intelligence as the scarce resource. But is it?

• The US has not executed a coordinated, national-scale grid buildout in roughly 50 years. We’re now asking that same grid to absorb the largest industrial load growth in a generation.

• JP Morgan estimates $1 trillion of US grid investment is needed over the next decade. But the constraint isn’t dollars — it’s time. The average new transmission line takes 10+ years to build.

• Chip efficiency has been the bottleneck on AI economics for the last three years. My POV is that grid capacity will be the bottleneck for the next 10.

• Data-center electricity demand is skyrocketing, with AI the biggest driver. Access to firm power will be the biggest competitive advantage in the AI race.

The AI Bottleneck isn’t Chips. It’s Watts

2026 saw OpenAI announce a $122 billion fundraise at an $852 billion valuation. Anthropic followed with a $65 billion raise at a $965 billion valuation. Between the two, they’re now worth roughly twice the combined market cap of the ten largest US investor-owned utilities.

The companies selling intelligence are valued at a multiple of the companies selling the electricity required to produce that intelligence. The market believes the battle for AI supremacy is a contest between Nvidia, Anthropic, and other technology giants. But the AI race is quietly becoming a race for electricity, not chips.

Last week I wrote about how CEOs are pushing back on AI spend and why the application layer is heading into a cost-accounting era.

That post focused on the demand side:

Because right now, enterprises are doing some deep soul searching to determine whether token consumption actually translates to business outcomes.

Subscribe now

Today’s post focuses on the supply side: AI cost conversations in board rooms are quickly becoming energy conversations. And right now — the math ain’t mathing.

I: US Power Grid: The Largest Machine on Earth

Universally available, reliable, and affordable electricity is taken for granted.

But delivering power to 350 million people in the United States is no easy feat. The American power grid is the largest machine on Earth — roughly 7 million miles of transmission and distribution lines, 240,000 miles of high-voltage backbone, and more than 180 million power poles. Concerningly, in the early 2010s we added 1,700 miles of new high-voltage transmission per year. Today, we’re adding 350. So what gives?

A grid that nobody fully owns

Quick: who runs the American power grid?

Trick question. Nobody does.

There is no National Grid Authority.

The US power grid has evolved from Thomas Edison’s localized 1882 Pearl Street Station into a massive, interconnected network. Today, it functions as one of the world's most complex engineering feats, a massive, complex industry that works in harmony to deliver power to consumers.

That’s right, the Eastern Interconnection, the Western Interconnection, and ERCOT, which covers most of Texas, are physically separate machines. Each of them is about the size of a continent, connected by only a handful of small DC ties. When a 2021 storm took down Texas, ERCOT could pull only ~6% of its peak demand from its neighbors. The result was the largest deliberate electricity shutoff in US history.

Inside the three grids, things get more fragmented. Seven regional operators — PJM, MISO, ERCOT, CAISO, NYISO, SPP, ISO-NE — coordinate roughly two-thirds of the country. The rest is run by individual utilities making their own decisions about generation, transmission, and who gets connected when. Large parts of the Southeast, Northwest, and Southwest have no formal wholesale power market at all.

Layer on roughly 3,000 utilities, 50 state public utility commissions, and one federal regulator (FERC) with limited authority over retail markets, and what you have is less a system than a federation of systems that mostly agree to keep the lights on.

It works, most of the time, because everyone behaves. But here’s the issue: the grid was not designed for anyone to add 50+ gigawatts of new industrial load in five years.

A grid older than most of the people running it

Most of the high-voltage backbone you see on a map today was built between 1950 and 1975. That backbone has been carrying power for fifty to seventy years. The average large power transformer in the US is roughly forty years old. Many of the substations being asked to serve modern hyperscalers predate the personal computer.

This is not a knock on the equipment — utility-grade hardware is built to last, and most of it does. But “keeping the lights on” and “expanding production” are two very different stories. And we need to significantly ramp capacity. So how do we do it?

The mechanisms for actually upgrading the system are slow. Permitting new transmission lines can take over a decade. Connecting new generation to the grid currently averages five-plus years and growing. There is no national grid authority that can compress the timeline; every project negotiates separately with state regulators, local landowners, and the utility on the other end of the line.

This is the part that people in the AI world tend to underestimate. You can spin up a Kubernetes cluster in an afternoon. You cannot spin up a 765-kV transmission line.

A grid getting greener and dirtier at the same time

The most interesting thing about the US energy mix right now is that it is moving in two directions at once.

In March 2026, renewables generated more US electricity than natural gas for the first time. Combined with nuclear, clean sources supplied more than half of all US power. Solar, wind, and batteries are projected to account for 93% of new capacity added to the grid this year, per EIA. The transition is real and accelerating.

At the same time, AI demand is keeping coal alive. Nine coal plants scheduled to retire last year had their operating lives extended, five of them under emergency orders from the Department of Energy. The White House has made its opinion on coal clear. Total coal capacity retired in 2025 was the lowest in fifteen years.

So how can both things be true? It’s because new megawatts are mostly green, but old megawatts are running longer than anyone planned. The grid we have today is more renewable, more fossil, and older — all at once.

This is the system we are now asking to underwrite the next decade of AI demand growth. Reinforcements are needed. Quickly.

Subscribe now

II: The Challenge Ahead

To meet surging electricity demand, significant grid infrastructure will need to be built and upgraded over the next decade.

Trillions of dollars are expected to flow into AI infrastructure over the next few years. JPMorgan estimates $1 trillion of US grid investment alone is needed over the next decade — $5.8 trillion globally. Those are big numbers. But the real question is, how long will it take to get that capital deployed?

The average new transmission line in the United States takes more than ten years to permit and build. Per recent New York Times analysis of FERC data, the interconnection queue currently holds more than 2,600 gigawatts of generation and storage waiting to connect — roughly twice the size of the entire installed US generation fleet. Average wait time in the queue has grown from less than two years a decade ago to more than five years today.

Meanwhile, the AI buildout is being announced on a different timeline. Every frontier model and AI chip release feels like a sprint. Announcements for new data center projects happen almost as frequently - with projects getting increasingly massive in scale. Project Stargate alone targets nine-plus gigawatts across seven US sites.

AI Data Centers Pushing the Grid to the Limit

The scale of AI data centers and their commensurate power needs are growing exponentially. To make the point more clear: a single gigawatt-scale campus needs more new firm power than most US states have added in the LAST DECADE.

A 1 GW AI data center consumes the equivalent power of roughly one million US households and requires $10B+ in upfront investment.

To meet growing AI demand, the semiconductor industry is constantly churning out increasingly powerful chips, which in turn requires a new generation of power-hungry data centers that are set to overwhelm our aging grid infrastructure. There’s a real risk that we run out of power before we can actually serve all of this demand.

The reason is simple: these AI data centers are using more power than ever.

By 2035, Deloitte estimates that power demand from AI data centers in the United States could grow more than 30x, reaching 123 gigawatts, up from 4 gigawatts in 2024.

The real constraint on AI is no longer talent or chips. It's power. The best models in the world are worth nothing if you can't plug them in.

III: While Everyone Fights Over Tokens, the Smart Money Is Buying Watts

There are two AI races happening right now. Only one of them is making headlines.

The Race for Compute

This is the visible race that everyone loves to cover: GPU allocation, model benchmarks, frontier capability, who gets API capacity first. It plays out in quarterly product launches and investor decks. It is loud, fast, and largely commoditized — every player is fighting over the same scarce inputs and the same downstream customers, and the gap between first and tenth place narrows every quarter.

The Race for Electrons

The silent race is the one a handful of players figured out 18-24 months ago. They stopped buying inference and started buying electrons.

The largest AI buyers are treating electricity like a top strategic priority. They are getting more sophisticated with offtake agreements, M&A, land purchases, and in some cases even choosing to vertically integrate with power producers outright. Effectively what they are doing is converting a variable cost — electricity at whatever the spot market clears — into a fixed cost at known prices.

A new model is emerging: data centers supplying or acquiring their own power. Tech companies are increasingly building or contracting for dedicated energy resources to secure faster access to electricity and reduce pressure on the grid.

The Smart Move: “Bring Your Own Power”

In a stable demand environment, this wouldn’t matter. In a power market where power costs are spiking faster than anyone expected, this has the chance to become the single most important strategic move in the industry.

Four things follow from securing your own power:

1. You control your own destiny. A company that has signed long-term PPAs for its compute is no longer a passenger on someone else’s grid. It does not need to stress about waiting in line for interconnection. It can focus on its business.

2. You control your own margins. Power is the single largest variable cost in AI inference. Locking it in is the equivalent of what airlines did with jet fuel hedges or what Amazon did with logistics: take the input most exposed to inflation, fix the price for a decade, and let everyone else pay spot.

3. You secure long-term capacity. This is the one nobody talks about. New gigawatt-scale generation projects take five to ten years to bring online. The capacity for 2030 is being contracted as we speak.

4. You move before the window closes. Every quarter that passes, the menu of available generation shrinks and the prices climb on what remains. With each passing day, counterparties also get much pickier about who they’ll sign with.

Now look at the AI application layer companies. They are buying inference capacity from hyperscalers at market-rate. They have little leverage. Their cost of goods sold is fully exposed to whatever the underlying power market does over the next decade.

No app layer AI company breaks out electricity costs as a line item. But if they did, it would tell an interesting story about who truly has a competitive advantage, and who has a serious margin problem.

In my view, the AI companies that survive the cost-accounting era will be the ones that understand that firm-power can be a strong moat.

IV: A Wake-Up Call

For nearly two decades, US electricity consumption was flat. That era is over. U.S. electricity demand is experiencing its first sustained, nationwide surge in years.

How the AI Race Might Bring Necessary Changes

The AI buildout may end up being the catalyst that finally funds the grid investment America has been deferring for fifty years. The time to act is now. The New York Times recently framed data centers as “a wake-up call” for a system that is “too old” with electricity supply that is “too small.”

That’s not a crazy take. It’s also exactly why the firm-power moat matters — the companies that locked in capacity before the wake-up call are the ones who benefit most from it. The reckoning is what makes their position valuable.

There are significant risks of not acting now.

Power emergencies are now routine; the federal government declared one during a recent heatwave. Supply chain bottlenecks, permitting delays, rising costs, cybsecurity threats and hardware shortages are among other risks to weigh.

Final Thoughts

The AI industry has spent three years optimizing for chip efficiency, model architecture, and inference cost per token. Those optimizations have produced a 90%+ decline in the price of intelligence and an even larger increase in its consumption.

The next 3-5 years will be different. The bottleneck is moving from silicon to substations, from inference to interconnection, from what Nvidia ships to what local utilities can actually deliver. The companies that recognized this early and locked in capacity before the wake-up call are the ones who will benefit most.

Scaling the Enterprise · A Series on AI’s Cost-Accounting era

• Next week (Part III): The cost of renting intelligence: How enterprises can use open-source to build custom intelligence instead of relying on frontier models.

• Coming soon (Part IV): If power is the bottleneck, where does the venture capital go? Mapping the investable themes at AI × Energy.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Views expressed are my own and should be considered as such, and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NYSE: NEE)

Key Findings

• The AI narrative is turning. Corporate pushback against AI tools is primarily driven by an ROI reality check

• Anthropic is the new AWS — minus the lock-in. Bills are arriving that look like cloud infrastructure, but customers can leave on a Tuesday.

• “Tokenmaxxing” has become a real corporate phenomenon, and the early data on it isn’t pretty: 78% of IT leaders report unexpected AI charges, and 80–85% of enterprises miss their AI infrastructure forecasts by 25%+.

• Microsoft and Uber’s recent moves matter more than the headlines suggest — neither was about AI quality, and both signal that the next 18 months will likely look choppier than expected for OpenAI, Anthropic, and Enterprise AI teams.

• The cost-accounting era favors three things: switching-cost discipline, consumption-aligned pricing, and real cost visibility.

Introduction

Early in my career as an investor, one of the signals I looked for in breakout companies was a specific kind of love-hate relationship:

What product do CFOs hate paying so much for, but absolutely can’t live without?

AWS came first. Bills skyrocketed year over year. Finance teams stared at invoices in disbelief and paid them anyway. They had to. By the time you noticed how big the bill was, your data was in S3, your auth was tied to IAM, your apps were stitched together with Lambda and RDS, and leaving wasn’t a procurement decision — it was a re-platforming project. AWS earned its margins because switching costs were enormous and the cloud enabled new and improved capabilities.

Then Snowflake. Before COVID, Snowflake ripped through enterprise IT budgets like wildfire. CFOs were furious. Then they signed the renewal. Once your queries, dashboards, governance, and data science teams lived in Snowflake, you weren’t going anywhere. Same lesson, different mechanism.

And now Anthropic. Bills are arriving that look more like AWS than SaaS. The mechanics are simple. Claude charges based on tokens processed—every word input and output costs money. Agentic AI tools can consume up to 1000x more tokens than basic chat queries, especially when employees deploy them for multi-step workflows or complex integrations.

Maybe that’s why Anthropic is currently valued at $965 billion, making it the most valuable private AI company in Silicon Valley.

Part I: The Sticker Shock Is Real

The End of Tokenmaxxing?

Corporate leaders are finally starting to question whether soaring AI spending is delivering meaningful returns.

Tokenmaxxing — the practice of maximizing employee AI usage to game adoption metrics — has surprisingly become a real corporate phenomenon overnight.

Uber blew through its annual AI budget in four months. Microsoft pulled Claude Code licenses across multiple business units. Fortune 50 companies began slashing token usage as a P&L action item. Amazon and Meta pulled internal AI leaderboards after employees gamed them with fake tasks. One anonymous large enterprise even allegedly burned $500 million on Claude in a single month.

Let’s Look at the Data

The pushback you’re hearing about is real. The early data on tokenmaxxing is starting to paint a bleak picture.

Enterprise AI spending hit a $1.2 million average per organization in 2026, up 108% year-over-year. 78% of IT leaders reported unexpected AI charges they never budgeted for. Roughly 80 to 85% of enterprises missed their AI infrastructure forecasts by more than 25%. Many companies aren’t seeing measurable ROI from their AI projects at all.

It’s possible we’ve simply reached the natural conclusion for the Tokenmaxxing trend. Take Meta’s dashboard, for example. Created by an employee, it tracked the top 250 token users among Meta’s 85,000+ staff. It awarded playful badges such as "Token Legend" and "Cache Wizard" to gamify usage. Quite ridiculous.

Are we seeing a "healthy swing" away from AI overuse and tokenmaxxing, or something more concerning for long-term AI demand?

Subscribe now

Part II: How Insatiable Demand Is Breaking AI Platform Economics

The answer to the question above depends on something the AI industry has been wrong about for two years:

The assumption that inference costs — the cost of running an AI model every time a user sends a prompt — would keep collapsing indefinitely.

Models would get smarter every quarter. Open-source competition would commoditize intelligence. Hardware improvements would lower serving costs. And eventually, AI companies would inherit the same envied economics that made SaaS one of the great business models in history.

That belief shaped an entire generation of startups. Founders launched “unlimited AI” plans before understanding long-term usage behavior. Investors rewarded explosive adoption over contribution margins. Enterprises rolled out copilots across entire organizations, assuming economics would naturally improve with scale.

In the early wave, the assumption looked correct. Costs per token kept falling. Every new release made the previous generation cheaper. AI adoption exploded faster than almost any software category in history.

But underneath all that excitement, another trend was growing even faster.

Usage.

Per-token costs fell — but token consumption per task is up an estimated 10 to 100x since late 2023, as reasoning models and agentic workflows have taken over. This is an obvious, but important point. It seems that many companies did not realize that agentic AI can consume up to 1000x more tokens than standard AI.

The math broke in the simplest possible way: per-token costs fell, but total tokens per task rose faster.

Part III: Why Recent Moves by Microsoft & Uber Matter More Than You’d Expect

Recent experiences at Microsoft and Uber have made the new AI economics impossible to ignore. Both involve sophisticated enterprises with deep AI commitments — and neither was about AI quality.

Microsoft

Microsoft is reportedly removing most internal Claude Code licenses across major engineering groups, redirecting developers toward GitHub Copilot CLI instead. Yikes.

Microsoft was not rejecting Claude’s quality. It clearly still believes frontier AI matters. Microsoft still has deep ties with Anthropic — the Azure partnership remains intact, and Anthropic models are still available in Azure AI Foundry.

Microsoft was rejecting the cost of unlimited third-party AI usage. A global organization with over 200,000 people will always need to prioritize platform control, infrastructure optimization, and procurement discipline.

Uber

Reports suggest Uber burned through its entire 2026 AI coding tools budget in roughly four months after internal adoption exploded. Monthly AI tooling costs reportedly reached $500 to $2,000 per engineer, while nearly 95% of engineers were actively using AI tools every month.

That is strong adoption.

But the important part wasn’t the adoption. It was the ROI. Once Uber’s leadership started digging into whether the rising token spend was actually translating into proportional business outcomes, they didn’t like the answer they found. And so they pulled the plug.

The Takeaway

Microsoft and Uber land on the same point from opposite directions. Microsoft has the platform leverage to swap one AI tool for another and chose to. Uber has the engineering culture to adopt at 95% saturation and still pumped the brakes on cost.

Neither company asked whether AI platforms are effective. Both asked whether the tools are worth what they’re paying for them.

Part IV: The Cost-Accounting Era: What to Watch

Three trends worth following. Each one is a question I’d now ask of any AI company that I’m evaluating for an investment.

1. What is the customer’s real switching cost?

If the answer involves a router, a gateway, or a “model-agnostic” architecture, the answer is: not much. Durable AI businesses are building lock-in above the model layer — in the workflow, the data, the agent state, the proprietary tooling. The application layer is where customer retention lives. The model layer is increasingly rebar.

2. Does pricing match consumption, or hide it?

Flat-rate AI pricing was the 2024 playbook. It is now a margin-destruction machine. The companies signaling discipline are moving to outcome-based, usage-based, or hybrid pricing. The ones still on flat-rate subscriptions are buying revenue with margin. Watch for who shifts first — and who waits too long.

3. Who has cost visibility, and who is flying blind?

CFOs aren’t pushing back because AI is expensive. They’re pushing back because it’s unpredictable. The first companies to give their customers — and themselves — clean dashboards, predictable bills, and consumption governance will look like adults in a room full of toddlers.

The Bottom Line

AWS was a generational business because its customers couldn't leave. Snowflake was a generational business because its customers wouldn't leave. The AI startups raising at AWS-like valuations need to prove that their customers won't want to leave — a fundamentally different argument, and a much harder one to win.

The good news: the cost-accounting era is also when great businesses get built. The operators and investors who lived through the 2001 dot-com reckoning will tell you the same thing. The companies that survived a serious unit-economics conversation became the companies that ran the next decade. The ones that didn’t, didn’t.

The AWS moment for AI isn’t a disaster. It’s a filter.

Next week: if the bottleneck on AI economics is cost, the bottleneck on cost is energy. Why the AI bottleneck isn’t chips — it’s watts.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Views expressed are my own and should be considered as such, and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NYSE: NEE).

Key Findings:

• The C-suite is climbing back down to IC roles. A wave of public-company CTOs and founders — plus Andrej Karpathy — have left executive seats to join Anthropic as individual contributors under a single flat title: Member of Technical Staff (MTS).

• Middle management exists to move information — aggregating context from below, relaying strategy from above, keeping everyone aligned. AI now does that routing work, which is why the middle layer is eroding fast.

• Flattening cuts both ways. It removes managers, but it also amplifies high-leverage individuals: one great IC can now do what a coordinated team used to. AI doesn’t just automate labor — it amplifies agency.

• The best people want proximity to the work. Impact in the AI era is concentrating in the people closest to the technical frontier, not the ones managing the biggest teams.

• The hidden cost is the talent pipeline. Cheap grunt work was how juniors earned their reps. Automate it away and you risk severing the apprenticeship that produces tomorrow’s seniors.

Something Odd is Happening in Tech

C-Suite execs of billion-dollar companies are quitting to take IC roles at Anthropic.

In May 2026, Andrej Karpathy — OpenAI co-founder, former AI Director at Tesla, one of the most recognizable engineers alive — joined Anthropic. Not to run a team. Not to hold a title. He joined the pre-training team as an IC so he could “get back to R&D.”

And surprisingly, he’s not even an outlier.

Over the previous year, a growing cluster of executives walked away from the C-suite to take IC roles at Anthropic under a single, deliberately flat designation: Member of Technical Staff. The CTOs of Workday, Instagram, Box, You.com, Super.com, and Adept AI all made the same move. Peter Bailis left the CTO seat at Workday — an $8 billion company with 18,000 employees — after less than a year to do RL engineering at a lab that now builds the kind of HR software Workday sells.

These are not people who ran out of options. They had the equity, the titles, and the institutional authority that most of corporate America spends a career climbing toward. They traded all of it for something the org chart can’t price: proximity to the work, without layers of management between them and the frontier.

Subscribe now

How did we get here?

A story in five (short) parts.

I. Middle Management’s Disappearing Act

Before we declare anything dead, it helps to remember why it was born.

Middle management isn’t a conspiracy to inflate org charts. It’s a solution to a real problem: information doesn’t move through large groups of people on its own. Once a company grows past the size where one person can see everything, coordination becomes the bottleneck — who’s working on what, what’s blocked, what the priorities are, what the front line is seeing that the C-suite isn’t.

Managers were the routing layer. They aggregated context from below, relayed strategy from above, and kept everyone pointed in the same direction. In theory.

However, AI has changed the game completely.

When Meta laid off roughly 8,000 people earlier this month — about 10% of the company — the internal memo didn’t frame it as cost-cutting. Zuck said people appreciate “the ability to take greater ownership and execute their vision with less bureaucracy and management to navigate.” The CEO of one of the largest companies on earth described middle management as something to be navigated around.

Companies have been flattening since the 1990s — reengineering, enterprise software, cloud collaboration tools — and each wave widened spans of control. None eliminated the function, because something still had to do the coordinating, and that something was a human being. Until now.

II. AI Is Compressing Coordination Costs

The principle is simple. The fewer coordination bottlenecks you have, the fewer managerial layers you need.

What you get instead is the trio every restructuring memo now reaches for: flatter orgs, wider spans of control, more autonomous workers.

The numbers are getting absurd in places. Across Meta, the average span of control has climbed to roughly 14 and is still rising, with some managers carrying 90 direct reports. Inside Meta’s Superintelligence Labs — the most AI-forward corner of the company — spans have reportedly stretched to 50 to 1.

A 50-to-1 ratio is double the 25-to-1 ratio that is usually seen as the “outer limit” of the so-called span‑of‑control scale. What could go wrong?

III. The Return of the High-Leverage IC

Suddenly, being an IC is the hottest job in tech. Founders back coding with AI: Mark Zuckerberg, Garry Tan, and from LlamaIndex.

When AI absorbs coordination overhead, it doesn’t just remove managers. It changes what a single talented person can do.

One great individual contributor can now analyze faster, execute faster, synthesize faster, and communicate faster than an entire small team could a few years ago. The research that took a week takes an afternoon. The deck that needed three people needs one. The report that waited on a chain of reviewers ships in a morning.

AI doesn’t just automate labor. It amplifies agency. Early-stage startups have always punched above their weight for the same reason — smaller teams, higher output, less bureaucracy.

Turn out, the pyramid was in many ways an artifact of expensive coordination. Make coordination cheap, and any high-agency individual — regardless of whether it’s a junior sales rep or CEO — can become the most valuable unit in the company again.

IV. So Why Are the CTOs Running for the Exits?

For decades, the conventional wisdom in tech was that an IC’s impact could only scale so far — to do more, you had to manage more. Recent evidence seems to indicate that’s no longer the case. Three reasons why:

1. Leverage beats span. The move isn’t seen as a demotion, but rather a bet on impact. In the AI era, the people closest to model training and research agendas hold more leverage than the ones managing large enterprise teams.

2. The flat structure is the point. Anthropic’s “Member of Technical Staff” designation deliberately avoids heavy management layers — and at top-tier compensation (base salaries roughly $300,000 to $405,000, with substantial equity), the roles are financially competitive with C-suite packages. You give up the title, not the upside.

3. Everyone wants to be at the frontier. With figures like Karpathy joining pre-training to use AI to accelerate AI’s own development, the most ambitious technologists want to be on the front lines of the work itself — not two layers removed, approving someone else’s roadmap.

   

V: What Risks Lie Ahead if Everyone is an IC?

Here’s the problem with optimizing away the middle of the pyramid: Companies may accidentally optimize away the very people who teach other people how to grow.

Entry-level work traditionally was an exchange: the junior did the grunt work — the research, the first drafts, the modeling, the documentation — and in return got reps, feedback, and mentorship. But what happens if there’s no one left to teach?

Wharton researchers call the thing we’re losing “cognitive apprenticeship” — people learn complex skills through observation, guided practice, and the gradual assumption of responsibility. Take venture capital - it’s widely known for its apprenticeship model of learning.

The apprenticeship model in VC has survived for decades because it’s critical for a firm’s survival: every generation of GPs has to produce the next one, or the firm doesn’t outlive its founders.

But what happens when there’s no one left to train the next generation of GPs? Is the next crop of Forbes Midas List investors supposed to learn the craft from AI agents?

Flattening an organization is easy. Rebuilding a severed talent pipeline is not. Just ask Disney, or GE — two companies that spent the last decade learning what happens when succession breaks.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Views expressed are my own and should be considered as such, and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NYSE: NEE).

Since then, Anthropic surveyed 81,000 Claude users about their fears around AI. Somewhat surprisingly, work anxiety tracks observed AI usage almost line-for-line. For folks that are short on time, the TLDR is below.

Key findings:

• A survey of 81,000 Claude users by Anthropic showed that people who work in roles that are more exposed to AI have more concerns about AI-driven job displacement. In other words, ignorance is bliss.

• Those in the highest- and lowest-paid occupations report the largest productivity gains, most commonly from increases in scope (doing new tasks). Concerns are also higher among early-career respondents (for good reason).

• My 2¢: Anxiety is the signal you’re paying attention. Keep paying attention and learning how to use AI tools, and there will always be a job for you. That job might just look different than the one you’re doing today.

Part I Recap: Picking up where we left off

A few months ago, I wrote about the gap between theotrical and observed AI exposure - the paper that quantified which jobs AI could automate vs. which jobs Claude was actually showing up in. That gap, I argued at the time, meant that much of the market panic about AI was being overblown - at least short-term. Long-term, however, I continue to worry about our future children’s job prospects.

Subscribe now

Part I was helpful in setting important context, but it was also missing an important piece of the puzzle. Anthropic’s new paper, What 81,000 people told us about the economics of AI, adds the perception layer: what workers think is happening to them. Plot it on the same axes and the picture sharpens dramatically.

Time to dive into Part II.

Part II: The New Data

The survey asked people about their visions and fears around advances in AI.

Chart #1: Workers’ fears track the usage data eerily well

Web developers and computer programmers anchor the top-right corner — high exposure, high fear. CEOs and school teachers sit in the bottom-left. The trend line is clean: every 10-percentage-point increase in observed exposure produces a 1.3-percentage-point increase in perceived job threat. Workers in the top quartile of exposure mention the worry three times as often as those in the bottom quartile.

This matters because it’s the first piece of evidence that AI usage data and AI sentiment data are converging. People aren’t panicking based on Twitter doomscrolling — they’re calibrating to what’s happening based on their own jobs.

Chart #2: Unpacking the U-curve

The most striking finding is that results are U-shaped: workers reporting the largest productivity gains from AI are also the most worried about being replaced by it.

(I love Claude, and yet, why does it feel like Claude will replace me?)

The people who report being slowed down by AI (left side — mostly fine artists and writers who find it stifling) are scared. Fine. But look at the right side: the people who report the largest boost are equally scared. The fastest movers and the most-stifled creatives both cluster at ~4% perceived threat, well above everyone in the middle.

A recent MIT Sloan piece on workflow redesign gives this a name: coordination cost. AI doesn’t have to be better than at every task to replace your role in the workflow. It just has to be good enough that removing the human checkpoint produces a faster overall system. Every handoff requires review, validation, adjustment — that friction adds up. The system optimizes by eliminating handoffs, not by speeding them up.

Power users see this clearly because they’re the ones constantly deciding whether to validate AI output or just let it run. They’ve internalized the math. Task-level mastery lacks a true moat. Owning entire workflows end-to-end is the right one.

Chart #3: The early-career squeeze

In Part I, I flagged the most uncomfortable discovery of a 14% drop in the job-finding rate for young workers entering high-exposure occupations. The unemployment numbers weren’t moving yet, but the entry-level pipeline was already narrowing.

Two months later, that’s now showing up as tremendous anxiety. Early-career respondents are roughly twice as worried about displacement as senior workers (8.2% vs. 4.0%). Only 60% of early-career workers say they personally benefit from AI, compared to 80% of senior professionals.

Chart #4: The optimist’s counterweight: scope, not speed

When users describe how AI helps them, the most common answer isn’t “I can work faster”. It’s actually “I can now do things I couldn’t do before.” 48% of users who described a productivity gain pointed to scope — new capabilities, new workflows, and tasks previously out of reach. 40% pointed to speed. Quality and cost were afterthoughts.

The optimistic frame is real. The market for tools that turn one person into ten is bigger than the market for tools that make ten people into one.

(One caveat: about 70% of survey respondents who named a recipient said the productivity surplus accrued to themselves, not their employers — but this survey only went to people with personal Claude accounts. Enterprise users, the ones whose employers paid for the seat, weren’t included. So take the 70% with a grain of salt.)

So what does this all mean? My takeaways

#1: The terrified are the prepared.

The right side of the U-curve is not where the casualties are — it’s where the survivors are.

The people reporting the largest speedups are also the people building the most accurate mental model of how AI changes their job. They’ve already noticed that the system optimizes by removing handoffs. They’re anxious because they understand the math — and the same understanding that’s making them anxious is what’s going to let them adapt. They’ll be the ones moving up the workflow ladder while their less-engaged peers are still optimizing tasks that don’t need optimizing anymore.

The ones who should actually be worried are the ones in the middle of the chart — high exposure, low perceived threat, low engagement. Casual users, modest gains, not really paying attention. Comfort with the tool is a leading indicator. Discomfort with what the tool implies is a lagging one — and the lag is the warning.

#2: The corporate ladder is being slowly disassembled rung by rung.

If the early-career anxiety holds — and the underlying hiring data suggests it will — it’s going to fundamentally reshape where ambitious young people go. The deal the corporate ladder offered for fifty years was: take an entry-level job, grind for a decade, climb. That deal assumed entry-level jobs existed. If those jobs no longer exist…what’s the rational move for a 22-year-old with a college degree? Only a few years ago, places like Meta and Microsoft offered enviable job security and perks. But times have changed.

Here’s what I think happens next:

1. Big Tech and the F500 lose their pull. The pitch — “join us, learn from senior people, climb” — only works if the rungs of the ladder still exist. If you’re a software engineer at Meta right now, you’re living in constant. Be sure to check on your friends working at M7 companies. The stress is real.

2. Graduate school becomes harder to justify. A two-year MBA or law degree is a bet that the credential premium will exist when you graduate. That bet looks worse when the work the credential gates is the work AI is best at automating. (I say this as someone who got an MBA from Harvard and loved it. Times have changed)

3. Startups, accelerators, and entrepreneurship win the talent war. YC and a16z Speedrun become the new prestigious early-career path — the only places where a 22-year-old can credibly bet on owning a workflow instead of being a node in someone else’s. YC’s next batch is going to be insanely competitive.

4. AI-native service firms eat the legacy ones. Trade organizations, agencies, boutique law firms, accounting/tax shops, consulting practices — anywhere a small team can punch above its weight using AI tools — wins. The 50-person AI-native law firm beats the 500-person legacy firm on price and turnaround.

#3: The Industrial Revolution, in reverse.

I would not be surprised to see a meaningful reverse migration from white-collar work back into blue-collar and manufacturing.

The data supports this: the most exposed occupations required the most education. The least exposed are physical, embodied, hands-on. The Industrial Revolution moved labor from fields and workshops into factories and offices. The current shift might run the same arc backwards.

Layer on the political moment: tariffs are pushing reshoring, the CHIPS Act is spinning up domestic semiconductor manufacturing, and electrical/HVAC/plumbing trades are facing a generational labor shortage with starting wages that look more attractive every quarter. Now imagine a generation of would-be knowledge workers looking at the corporate rat race and choosing the trades instead.

It sounds insane. So did “the cooks and the dishwashers will be the only ones with stable jobs” two years ago, and now I have a Part I post that literally made that joke. (Grounds Maintenance Technicians, here I’ve still got my eyes on you)

Am I betting the farm on this one? Probably not, but only because it means that I could be out of a job sooner than later.

The Bottom Line: Fear Not?

Strip away the doom narrative and the data is actually telling an optimistic story.

• First, most AI productivity gains are accruing to individuals more than to employers — which is a welcome change from how every prior wave of workplace technology played out. For now, the leverage is in human hands.

• Second, despite the media outlets love for reporting on layoffs and social media doomers obsession with telling people the world is ending, the AI layoffs are really more myth than reality. Many layoffs are due to declining performance, restructuring, and economic uncertainty from tariffs and middle east turmoil.

So if you’re sitting in the high-exposure, high-anxiety quadrant of these charts, here’s what I recommend you do based on the data:

1. Use AI more, not less. The U-curve says engagement is correlated with awareness, and awareness is correlated with adaptation. Casual users are the ones not paying attention. Become a power user — and become indespendible at work.

2. Move up the workflow. Task-level skill is depreciating. The unit of value has shifted to workflows you own end-to-end — not the discrete steps inside it. Ask yourself which workflow you can credibly run alone with AI, and master it.

3. Bet on yourself sooner. The traditional 5 or 10-year corporate apprenticeship is increasingly looking like a bad risk-adjusted bet. If you were going to start something in 5 years, the data says start it in two. Startups, AI-native services firms, and small high-leverage teams are where the surplus is being captured.

4. Pay attention to the trades. We love plumbers here! The reverse-migration thesis might be wrong, but the underlying truth — that physical, embodied, hands-on work is the most defensible part of the labor market right now — isn't far off.

The full paper is available from Anthropic’s research page. Part I of this series is here. Stay tuned for Part III -- coming in a few months. Happy Monday friends!

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Views expressed are my own and should be considered as such, and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

The enemy of my enemy is my friend, yet again.

In shocking news, Cursor and SpaceX have announced a groundbreaking partnership which could eventually lead to a $60 billion acquisition.

You see, both Cursor and xAI have a shared enemy — Anthropic and OpenAI — two companies which are increasingly moving full-stack beyond the model layer.

The Cursor Perspective

There was a time when Cursor and Anthropic were bffs and Cursor was happy to be Anthropic’s biggest customer. Then Anthropic launched a direct competitor and potential Cursor-killer in Claude Code after first getting a first-hand look at what made Cursor successful. Needless to say, they’ve cancelled their Friendsgiving plans.

The rise of Claude Code has come at Cursor’s expense. While Cursor remains a popular IDE, Claude’s models and the CLI tool, Claude Code, have set new standards for handling complex, multi-file tasks and large context, pushing Cursor to evolve.

In short, this deal shows that Cursor is actively looking to diversify away from OpenAI and Anthropic, who have been after Cursor’s lunch money for some time.

The xAI Perspective

There was also a time when Sam Altman and Elon Musk were good friends. The two former OpenAI co-founders and AI visionaries have a long history with each other. However, in 2018 Musk left the OpenAI board citing phoilosophical differences. Since that time, let’s just say things have not exactly been warm and fuzzy.

• Musk is suing OpenAI, Altman, co-founder Greg Brockman and the company’s biggest investor, Microsoft, which owns a 27% stake in the company.

• OpenAI has called Musk’s lawsuit “baseless,” a “sham” and a “harassment campaign.”

• Musk and Altman are also both actively vying for preferred treatment from the current Trump Administration, as seen in projects like The Stargate Project.

Subscribe now

xAI + Cursor: Deal Breakdown & Strategic Implications

Overview

xAI (now under SpaceX) is reportedly partnering closely with—and potentially acquiring—AI coding startup Cursor for up to $60B. The deal appears structured as a strategic partnership with an acquisition call option, combining compute, data, and talent to accelerate xAI’s position in coding AI.

Why the Deal Makes Sense

For Cursor

Cursor’s position has shifted from early leader to increasingly pressured player:

• Supplier risk: Heavy reliance on Anthropic (Claude) and OpenAI, both of whom are now direct competitors in coding tools.

• Vertical integration threat: Model providers are moving into IDE-like experiences, compressing Cursor’s differentiation.

• Strategic hedge: Partnering with xAI gives Cursor:

  • A third model provider (diversification)

  • Access to large-scale compute (GPUs)

  • Potential liquidity via acquisition

Bottom line: Cursor reduces risk while gaining infrastructure and a potential exit.

For xAI (SpaceX)

xAI is behind in coding—arguably the most commercially valuable AI use case today:

• Musk has acknowledged Grok is not competitive in coding

• xAI has faced talent churn and is undergoing a rebuild

• Recent hires from Cursor signal urgency in closing the gap

Cursor provides:

• High-quality coding data + user traces (critical for training)

• A proven product layer with real developer usage

• Talent and domain expertise in code generation workflows

Bottom line: Instead of building from scratch, xAI is effectively buying acceleration in the most important AI battleground.

Likely Deal Structure (Best Guess)

The partnership appears to function like a staged acquisition:

1. Data + Compute Exchange

   • Cursor provides usage data / traces

   • xAI provides GPU capacity + model access

2. Model Improvement Loop

   • xAI uses Cursor data to improve Grok + coding models

   • Cursor potentially integrates xAI models into its product

3. Option-Based Acquisition

   • If successful → xAI acquires Cursor (~$60B)

   • If not → Cursor receives ~$10B (data + partnership value)

Takeaway: This is effectively a low-risk call option for xAI and a high-upside partnership for Cursor.

Deal Context: Signals We Should Have Noticed

The more I think about this, the more I think we should’ve seen this move coming, for three main reasons:

1. xAI recently hired two key Cursor engineers (Andrew Milich & Jason Ginsberg)

2. Musk publicly stated Grok needed to be rebuilt from the ground up

3. All 11 co-founders who launched xAI with Elon Musk in 2023 have departed the company (or were pushed out?) as of late March 2026. Reinforcements were needed.

This is not even to mention the high-stakes AI coding war brewing between the Model-native tools (OpenAI, Anthropic) and the Application-layer tools (Cursor, Cognition, GitHub Copilot).

Market Implications

1. Acceleration of the Coding AI Arms Race

• Coding is emerging as the highest-value AI use case

• This deal intensifies competition between:

  • OpenAI (GPT + Codex direction)

  • Anthropic (Claude + strong coding usage)

  • xAI (now aggressively catching up)

2. Pressure on Anthropic & OpenAI

• Cursor is a major downstream consumer of Claude tokens

• If usage shifts to xAI:

  • Anthropic could lose a key distribution channel

  • xAI gains immediate demand + feedback loops

This is less about absolute growth slowing—and more about share shifting.

3. Vertical vs Horizontal Battle

• OpenAI/Anthropic: moving down the stack (into apps)

• Cursor: sits at the application layer

• xAI: appears to be choosing a hybrid strategy:

  • Stay model-focused

  • Partner (not compete) at the app layer

4. Data Advantage Becomes Decisive

• Cursor’s real value isn’t just product—it’s:

  • Developer workflows

  • Code edit traces

  • Feedback loops

This reinforces a key trend: The winners in AI coding will be those with the best data and compute resources, not just models.

5. SpaceX IPO Narrative Risk

• Raises questions about:

  • Whether xAI was overvalued at $250B

  • Why it needs external help to compete

• At the same time:

  • Shows aggressive action to fix weaknesses

  • Supports Musk’s broader “AI + space + robotics” narrative

Closing Thoughts

This deal signals that the AI coding market is entering its most competitive phase yet. xAI is no longer trying to catch up organically—it’s buying its way into the race, while Cursor sees the writing on the wall and is seeking to an opportunity to stay relevant in a race that’s increasingly dominated by model providers.

Key Takeaways

• Cursor is trading independence for distribution, compute, and optionality.

• The deal could re-route a meaningful portion of coding AI demand toward xAI.

• This puts an even bigger spot light on the upcoming SpaceX IPO. A lot of VC funds will be watching this one anxiously.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed are my own and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

TLDR: **Enterprises are better at adopting AI than you think**

• 📰 The Headline: MIT says 95% of enterprise AI pilots fail. The stat ran like wildfire. The truth is more interesting.

• 👑 The AI Elite: 92% of the C-suite are cultivating a new class of AI super-users — while 60% plan layoffs for non-adopters.

• 💰 The Money: Average enterprise AI spend hit ~$7M in 2025 — and is projected to jump 65% to $11.6M in 2026.

• 🏆 The Agent Milestone: 80%+ of Fortune 500 now run AI agents in production.

• 🤖 The Paradox: 97% of executives say they’re benefiting from AI. Only 29% see significant organizational ROI.

Part I: About That MIT Report…

You’ve seen the headline: “95% of enterprise AI pilots fail.”

LinkedIn. Board decks. Twitter posts. Earnings calls. Forbes articles. Conference panels. It went everywhere.

The stat comes from MIT’s NANDA initiative — The GenAI Divide: State of AI in Business 2025 — based on 150 executive interviews, 350 employee surveys, and 300 public AI deployments. The finding: only 5% of AI pilots achieved rapid revenue acceleration. The vast majority delivered “little to no measurable impact on P&L.”

This data was real. However, the media narrative was not. Here’s where it went sideways.

What the stat actually measures

The 95% figure measures one thing: whether an AI pilot produced rapid P&L impact within six months. Not productivity. Not cost savings. Not efficiency gains. And it mostly measured pilots in sales and marketing — the lowest-ROI area in the study.

Measured that way, most projects will “fail.” A new hire doesn’t move the P&L in six months either... they often take six months or more to ramp up!

The study’s most important finding got buried: vendor-led deployments succeed 67% of the time. Internal builds succeed one-third of the time. This was always a story about strategy, not technology. This is a better takeaway for enterprises to focus on.

Do 95% of pilots really fail? The counter-evidence:

The data from Q1 2026 paints a more optimistic picture:

• 72% of enterprises have at least one AI workload in production — up from 55% in 2024 and 20% in 2020.

• Organizations deploying AI across core operations report 20–40% productivity gains in year one

• Firms moving AI to production average 1.7x ROI; top performers see 10–18x

• AI super-users report 5x productivity gains and save nine hours per week

• 66% of organizations are already reporting real efficiency gains (Deloitte, 2026)

The MIT Report headline should’ve been: “AI works great once you have a plan to demonstrate ROI and start embedding it into actual workflows.” But that doesn’t fit nicely in a tweet. So here we are.

Subscribe now

Part II: Where AI Is Actually Working

The enterprises seeing returns aren’t chasing flashy deployments. They’re doing something less glamorous: embedding AI into high-specificity, high-integration workflows that rarely make headlines. The highest-ROI deployments share three traits: domain specificity, deep workflow integration, and buy vs. build.

2A: The Top Use Cases

✅ Where It’s Working: The Winners’ Playbook

1. Software Development is the clearest success story. AI coding tools have moved from experimental to “how does anyone write code without this?” Engineering teams are cutting PR review time by a third and boosting code throughput by 30-100%. Anthropic has emerged as the clearleader here.

2. Back-Office Automation produces the highest returns in the MIT study despite getting the least board attention. Case studies show $2–10M in annual savings from replacing outsourced document review and support. Cost savings of 26–31% across supply chain, finance, and operations. The less glamorous and more manual the function, the higher the return.

3. Healthcare & Clinical Documentation — AI is now demonstrably improving accuracy and efficiency in clinical documentation. Microsoft’s Dragon Copilot documented 17 million patient encounters in Q1, up nearly 5x year over year. The Permanente Medical Group (TPMG) recently published analysis finding that Gen-AI scribes saved physicians an estimated 15,791 hours of documentation time — equal to ~1,800 eight-hour workdays — while also improving patient-physician interactions and satisfaction. Real value — yet nobody’s writing X posts about it.

   

4. Customer Support — 90%+ of enterprises are now testing third-party AI rather than building internally. One public fintech abandoned their internal build mid-development to buy instead. The data is clear: specialized, integrated tools win. Robinhood is one notable exception, successfully launching Robinhood Cortex.

5. Financial Services - Purpose-built AI tools for finance has been another quiet success story. A few cool examples:

   1. Rogo: AI for Financial Services

   2. Concourse: AI for Corporate Finance Teams

6. Fraud Detection — AI-powered transaction monitoring has cut false positives by up to 200% at leading financial institutions. It’s no longer a pilot. It’s mission-critical infrastructure. See below for an example:

   1. Variance: AI Agents for Fraud Detection

2B: The Rise of the AI Elite

The most underreported story in enterprise AI right now comes from Writer’s 2026 Enterprise AI Adoption Survey (1,200 C-suite executives, 1,200 employees).

The enterprise workforce is splitting in two. Fast.

92% of C-suite execs admit they’re actively cultivating “AI elite” employees, while 60% plan layoffs for non-adopters.

The outcomes gap is stark: AI super-users save nine hours per week — 4.5x more than laggards. They’re 3x more likely to have received both a promotion and a raise in the past year. They’re 5x more productive. The C-suite has noticed.

Who are these super-users?

• Skew younger: 43% Gen Z vs. 25% of Baby Boomers

• Cluster in marketing, HR, sales, and customer support

• Save nearly a full workday every week vs. colleagues who haven’t adopted

77% of executives say employees who refuse to become AI-proficient won’t be considered for promotions. 90% say the rise of AI super-users will require completely rethinking performance evaluation.

Your relationship with AI in 2026 isn’t just a productivity preference. It’s a career decision. Choose wisely.

Part III: What’s Actually Failing — And Why

Now for the honest section. Because AI is also genuinely failing in predictable, avoidable ways — and it’s worth exploring exploring what’s going wrong.

Reason #1: The pilot graveyard

59% of companies invest over $1M annually in AI. Only 29% see significant ROI from generative AI. Only 23% from AI agents. The math is not working.

42% of companies abandoned most AI initiatives last year — up from 17% the year before. The average org scrapped 46% of POCs before production. Not a technology problem. An organizational problem wearing a technology costume.

The top barriers from enterprise CIOs: technical complexity and integration (26%), security and privacy (26%), uncertain ROI (24%). But underneath all three is one root cause: organizations are trying to solve a transformation problem with a highly limited “technology’ budget.

Reason #2: The build-vs-buy trap

Gartner research shows organizations buying existing AI solutions can deploy 60% faster than those building from scratch. And as we covered earlier, our friends at MIT found that vendor-led AI projects had a 67% average success rate (!!).

In our sample, external partnerships with learning-capable, customized tools reached deployment ~67% of the time, compared to ~33% for internally built tools. ~ MIT NANDA

Is the Buy vs. Build debate over for good?

Almost everywhere MIT researchers went, enterprises were building their own tools. And almost everywhere, it was the wrong call. Vendor-led deployments succeed at 2x the rate of internal builds. A specialized vendor has solved the integration and governance problems a hundred times. Your internal team is solving them for the first time, while fighting seventeen other priorities.

Reason #3: Employees are actively pushing back

Here’s the data point that should be in every CIO’s deck right now. 29% of employees admit to sabotaging their company’s AI strategy — jumping to 44% among Gen Z. How? Entering proprietary data into public tools, using non-approved software, intentionally generating low-quality outputs, refusing AI training, and tampering with performance metrics to make AI look bad.

76% of executives say employee sabotage poses a serious threat. And yet 75% of those same executives admit their AI strategy is “more for show” than actual guidance.

The sabotage is rational. When executives impose mandates without genuine strategy, resistance is the logical response.

Subscribe now

Part IV: The Leaderboard — Who’s Winning Enterprise AI Right Now?

The enterprise AI battleground is heating up.

A16z’s Q1 2026 CIO survey of 100 Global 2000 companies is the best public dataset we have. Here’s what it shows.

OpenAI is the enterprise leader today, but the tide is shifting…

78% of Global 2000 CIOs use OpenAI in production. It dominated the first wave — chatbots, knowledge management, customer support — and those beachheads are sticky. Switching costs are real — but not insurmountable.

• But momentum is shifting. Anthropic and Google have made meaningful share gains with Anthropic’s rise particularly striking, even excluding AI coding startups.

• Wallet share tells a similar story. OAI still commands a majority at ~56%, but Anthropic and Gemini are steadily gaining at OpenAI’s expense. The survey also suggests that respondents expect that shift to continue into 2026.

• Multi-Model is New Default: The chart below tells the full story. OpenAI’s “only” share has been declining since mid-2024. The total adoption line is holding because customers are also now using both Open AI and Anthropic. The multi-model approach is now the default enterprise.

Anthropic: The story of Q1

Since May 2025, Anthropic posted the largest share gain of any frontier lab — up 25% in enterprise penetration. 44% of enterprises use Claude in production, rising to 63% including testing. Gains are concentrated in software development and data analysis. Critically, 75% of Anthropic customers are already running the latest models — vs. OpenAI, where many enterprises are still on earlier models that “work well enough.”

Use case by vendor

The model race is several parallel races:

• OpenAI dominates chatbots, knowledge management, and customer support — the early horizontal use cases

• Anthropic leads software development and data analysis, where the capability gap is widest

• Google Gemini is competitive across most use cases — except coding, where its share lags

81% of enterprises now run 3+ model families, up from 68% a year ago. Nobody is betting on a single horse.

And the actual winner is… Microsoft

While everyone debates Claude vs. GPT-5, the actual enterprise AI winner is a productivity suite from 1989. 90%+ of Fortune 500 companies use Microsoft 365 Copilot in daily workflows. GitHub Copilot has 26 million users. 65% of enterprises prefer incumbent solutions — citing trust, integration, and procurement simplicity.

The incumbent has a 30-year head start and a line item in every enterprise budget. Don’t underestimate that. But the more interesting question is - will adoption stick after a few renewal cycles?

Part V: Q1 2026 Scorecard

One of the persistent problems with enterprise AI coverage is that the data is scattered across a dozen surveys, each with their own methodology and agenda. Starting this quarter, I'm going to be consolidating all of these signals into a single, consistent scorecard — updated every quarter, same format, clean comparisons. The goal is to present a clean view of how Enterprise AI Adoption changes over time.

Here's what Q1 2026 looks like.

Same format, every quarter. Bookmark it.

Final Thoughts: What to Watch in Q2

The 95% stat became a media darling because it confirmed what skeptics wanted to believe. The actual story is more interesting. AI is working — in back-office automation, software development, fraud detection, and clinical workflows. It’s failing where someone bolted a chatbot onto a workflow and called it a strategy. And it’s failing because organizations haven’t figured out how to turn individual productivity wins into enterprise-wide transformation.

Three things I’m watching in Q2:

1. Can anything stop Anthropic’s current momentum? The a16z adoption chart is clear. If Anthropic’s trajectory holds through Q2, we’re looking at a genuine leadership change by year-end. Switching costs are high — but not infinite.

2. Build vs. buy debate trending towards “buy”. Vendor-led deployments outperform internal builds 2:1. Enterprises still dedicating engineering resources to LLM wrappers are heading for a painful Q3 reckoning. The emerging category of companies combining model access, governance, and workflow integration in one platform will be the biggest M&A targets of 2026.

3. Revenue accountability becomes the new mandate. If it continues that only 29% of executives see meaningful ROI from AI initiatives…Boards and activist investors will start asking hard questions. The 29% camp share four traits: AI tied to revenue outcomes, governance built before scaling, business teams owning AI workflows, and the whole thing treated as Org redesign.

AI spend is no longer an innovation budget. It's the cost of staying alive. The companies still treating it like an experiment aren't being cautious — they're falling behind. Q2 will separate the winners from the losers. More soon.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed are my own and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

Story’s over, right?

Nope. Since then, DeepDelver published Part II of a five-part series. A whistleblower materialized with a data dump. And the Delve saga officially graduated from “embarrassing startup scandal” to something considerably more sinister.

🏃 The Generational Run: A Recap

For anyone just joining us, here is the complete Delve mixtape:

Delve has been on a generational run.

1. Got into YC

2. Onboarded Lovable, Cluely, and other unsuspecting companies

3. Allegedly faked hundreds of SOC 2 & ISO 27001 audits

4. Spammed a “wall of love” with logos they claimed to have “helped close”

5. Got caught

6. Made employees delete all the old posts

7. Put out an explanation — with replies turned off

8. Allegedly stole a fellow YC company’s open-source IP (Sim.ai / SimStudio)

9. Rebadged it as a standalone product called “Pathways”

10. Sold it to Brex, Anthropic, Gusto, and Notion — one deal was $50k+

11. Got caught again

Want the full pre-scandal arc — the billboards, the Forbes 30u30, the Garry Tan tweet? Read Part I. Then come back.

Subscribe now

📬 Part II: The Whistleblower Enters the Chat

After DeepDelver’s first article went viral, CEO Karun Kaushik issued a denial on X and LinkedIn — reportedly with YC backing. Standard playbook: automation platform, independent auditors, industry-standard templates. Nothing to see here.

DeepDelver’s response was to go bigger. Part II isn’t a single article — it’s five consecutive days of drops, fueled by a Delve employee who came forward with a cache of internal data, screenshots, and videos. He brought the receipts.

DeepDelver

Delve - Fake Compliance as a Service - Part II - Day 1 of 5

Reporters: aicpnay@proton.me…

Read more

4 months ago · 84 likes · 6 comments · DeepDelver

The most damning revelations:

• Fake reports. A screenshot shows internal engineer Sazzad Islam writing on August 9, 2025 that “v0 of Delve AI was live” — after the Series A. Internal staff were actively participating in something called “Project Audit Automation.” Delve claimed they don’t generate reports. The internal receipts say otherwise.

• Services company, not AI platform. One internal note reads: “To supplement, we’ve asked vCISOs to support them off-platform, and in effect become exactly what we didn’t want to be: a services company. This presents a large problem for us.” Services. Not AI. Services.

• Auditor swap. Delve quietly migrated clients to a new auditor — Ezzy & Associates — while telling them they wouldn’t need to restart their SOC 2 Type 2 observation periods. The founder of Ezzy? A Certified Fraud Examiner. You genuinely cannot make this up.

If you’re having a bad week, just remember — Delve is having a worse one.

🕵️ Hold Up: How Is DeepDelver Getting This?

Let’s slow down, because the sourcing here deserves serious scrutiny.

Think about what DeepDelver has produced across two articles: internal Linear tickets, Notion documents, AI call-notetaker logs from private sales calls, pitch decks, employee DMs, full database access, and now an internal whistleblower dump. This is not what an angry ex-customer does. A former customer doesn’t have timestamped Slack conversations between engineers captured during an active sales call.

My read: DeepDelver is almost certainly a disgruntled ex-employee — or a competitor who planted a spy inside the company.

The latter has Silicon Valley precedent. In 2025, Rippling alleged — and eventually proved via bank records — that Deel had planted a mole inside the company to exfiltrate trade secrets. The full story is here and is worth reading.

None of this discredits what DeepDelver is publishing. The allegations are specific, evidence-backed, and increasingly corroborated by third parties. But the source matters — for understanding motivation, and for any downstream legal proceedings.

Subscribe now

🧑‍💻Day 2: The Sim.ai Swindle

Sim.ai (SimStudio) is a fellow YC company. In April 2025, CEO Emir Karabeg got on a sales call with Karun Kaushik to purchase compliance services — a $15,000 SOC 2 package. Karun promised personal attention, great service, and threw in Arcteryx jackets and a box of donuts.

DeepDelver

Delve - Fake Compliance as a Service - Part II - Day 2 of 5

I had to take this ‘episode’ into a completely different direction at the very last minute, so I’m sorry for dropping it this late…

Read more

4 months ago · 61 likes · 11 comments · DeepDelver

Here’s what was happening on Karun’s end of that call.

From the timestamps: Karun posted about SimStudio internally while still on the call. Delve’s Linear shows the first mention of Sim Studio under the Pathways project appeared in April 2025. A Notion page in the whistleblower dump — titled “Sim Studio Port Plan” — lists exactly which folders to copy from SimStudio’s codebase.

He was selling them compliance services and stealing their product. Quite the multi-tasker.

When Emir later tried to sell Delve a licensing deal, Karun said the ROI “wasn’t there” and went silent — while Delve was closing Pathways deals with Brex, Anthropic, Gusto, and Notion using SimStudio’s code.

DeepDelver calls this “stealing IP” which is a bit dramatic, since open source tools are freely available to use. But only if proper credit is given, hence the problem.

When DeepDelver reached out to Sim.ai directly, the team confirmed: no white-label agreement was in place, and they had no idea Delve was selling Pathways as a standalone enterprise product until DeepDelver’s first article published.

🤔 Can We Still Trust the Trust Industry? A Much Needed Wake-Up Call

The recent allegations surrounding Delve should make every SaaS founder, CTO, and security leader stop and ask a difficult question:

Are we building trust — or just engaging in Compliance Theater?

Recent reporting from popular tech outlets have served as a major wakeup call to the industry. Whenever a company gets caught with blantantly fake evidence of meetings, reports and tests that never occurred - it makes you wonder how many other compliance startups are doing the same. Delve is not the only bad apple.

This is the moment many of us in security and compliance have been expecting. What comes next is anyone’s guess.x

🔥The Bigger Picture: 7 Incidents. 10 Days.

Let’s zoom out: March was NOT a good month for governance and security.

LiteLLM. Axios. Railway. OpenAI Codex. Mercor. Claude Code. And right there in the middle: Delve. 494 fake compliance reports.

Delve didn’t create this problem — they just found the most cynical possible way to profit from it. The AI wave is moving faster than security infrastructure can keep up, and a generation of startups has learned that looking compliant is often good enough to close a deal, land a customer, and raise a Series A.

The shortcuts compound. A fake SOC 2 cascades into a LiteLLM breach. A rubber-stamped ISO 27001 becomes a HIPAA liability. The trust page listing controls that were never implemented becomes evidence in a lawsuit. Compliance isn’t a box-checking exercise — it’s the load-bearing wall of enterprise software. And right now, too many teams are building on a foundation of drywall and donuts.

⌚More to Come?

What’s crazy is that DeepDelver series isn’t done. Parts 3, 4, and 5 are coming. Part 4 or 5 reportedly contains what the author called “quite the nuclear bomb.”

We’ll be watching.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Views expressed are my own and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

Delve had all of it. Then one anonymous Substack blew it all up.

In a well-researched article that nonetheless reads like scorned ex-girlfriend, Substack user DeepDelver alleges that Delve “fabricated evidence of board meetings, tests, and processes that never happened,” then forced customers to “choose between adopting fake evidence or performing mostly manual work with little real automation or AI.”

It got me wondering…how on earth did we get here?

Here’s the full timeline — buckle up.

🗓️ The Delve Timeline

April 3, 2024 — YC Winter 2024 Graduation

Delve graduates from Y Combinator’s Winter 2024 batch. The pitch: an AI-powered compliance platform that eliminates the brutal manual grind of SOC 2, HIPAA, GDPR — the whole alphabet soup. Smart wedge. Real pain point. Investors love it.

January 2025 — $3M Seed Round

Delve closes a $3M seed from General Catalyst, FundersClub, and Soma Capital. The company reports 100 customers. Growing fast. Humble beginnings.

July 22, 2025 — The $300M Series A

This is where it gets interesting. Delve — barely 2 years old — closes a $32M Series A led by Insight Partners at a $300M valuation. A roughly 10x jump from the seed in under six months. The growth rate is mindboggling for many, including myself.

The TechCrunch headline said it all: “21-year-old MIT dropouts raise $32M at $300M valuation.” The company was careful to signal two things loudly: the premium valuation and the claim that it was already profitable. Over 500 customers. Fast-growing AI unicorns like Lovable, Bland, and Wispr Flow on the roster. The founders weren’t just building — they were arriving.

Subscribe now

September 13, 2025 — Garry Tan Tweets!

YC president Garry Tan amplifies co-founder Karun Kaushik posting about returning to MIT to give the same YC talk that inspired him to apply. Garry’s caption:

“Delve is now a top YC startup.”

175,000 views. The train continues on.

October–November 2025 — The Billboard Era

Delve runs what they’d later describe as one of the largest out-of-home campaigns of all time — billboards, bus wraps, and transit takeovers across SF, NYC, and Austin. Bus stops in SoMa. Panels on the 101. A startup playing a big company’s game.

(Delve even published a detailed blog post about how to run an OOH campaign. Weird flex)

December 2025 — Forbes 30u30

Co-founders Karun Kaushik and Selin Kocalar are named to Forbes’ 30 Under 30 list for 2026 in the AI category. The prestige flywheel is spinning at full speed. By every external signal, Delve is an absolute rocketship. 🚀

March 18, 2026 — The Substack Drops

An anonymous tip leads New York Times reporter Erin Griffith to post a link a Substack analysis about Delve’s misleading tactics.

In the article, a pseudonymous investigator operating under the name DeepDelver publishes an in-depth, months-long investigation on Delve. It goes viral instantly.

The allegations are serious: that Delve was generating fraudulent or misleading SOC 2 reports — using copy-paste templates, low-credibility auditors, and pre-filled evidence that customers didn’t actually customize. And apparently, many of the customers Delve proudly claims are real either?

DeepDelver

Delve - Fake Compliance as a Service - Part I

The version that arrived in your mailbox is truncated. Visit the full article here to view the rest…

Read more

4 months ago · 484 likes · 45 comments · DeepDelver

March 20, 2026 — Delve Issues a “Response to Misleading Claims”

The company publishes a weary blog post pushing back on five specific allegations. The way Delve tells it: the company is just an automation platform — it doesn’t issue reports, independent auditors do. The templates are industry-standard. The pre-filled evidence is just a starting point. Nothing to see here.

March 20, 2026 (Same Day) — Lovable Drops the💣

Later that afternoon, Lovable — one of Delve’s marquee, name-checked customers — posts this on X:

“We’re aware of recent reporting about Delve’s compliance practices. Lovable is not a Delve customer. We proactively moved to Vanta in late 2025, before any of this came to light. Our SOC 2 Type II was independently audited by Prescient Assurance.”

The same Lovable that was cited in Delve’s TechCrunch article for its Series A as a flagship customer. Still listed on Delve’s website (now gone). This is the kind of thing that tips the tide of war. When your own marquee customer publicly disavows you — on the same day you’re doing damage control — its tough to recover.

🧠 The Real Takeaway: The Downside of Silicon Valley’s Fake-It-Till-You-Make-It Culture

Delve isn’t an anomaly. It’s the inevitable output of a system with broken incentives.

Silicon Valley’s hustle culture has always had a complicated relationship with the truth. The canonical advice — “move fast, break things” works right up until it doesn’t. And increasingly, it doesn’t.

Truth is, in the quest for eternal Sand Hill Road Glory, Delve is not the first to bend the rules, and they won’t be the last:

Exhibit A: 11x

Remember these guys? The a16z and Benchmark-backed AI SDR startup that raised a $50M Series B at a $350M valuation in 2025. Their claim to fame? Selling one-year contracts with a 90-day “break clause” opt-out — and then counting the full annual contract value for every customer, including the ones who bailed. VC friends, this is why you do pipeline diligence calls..

^^I even wrote about it 365 days ago! Looks like covering tech scandals might be my new annual March 26th tradition…

Exhibit B: Frank (feat: Charlie Janice)

For those who need a reminder — this is the founder who defrauded JPMorgan out of $175 million by fabricating a database of 4 M student customers for her college financial aid startup, Frank, which JPMorgan acquired in 2021. The customers were fake. The money was real. Janice was eventually convicted of fraud. The DOJ's parting line at sentencing: "Today's sentencing should send a message to other entrepreneurs who may be tempted to cross the line into fraud.”

Sadly, the message wasn’t received by everyone.

Where We Go from Here

The thread connecting Delve, 11x, and Charlie Janavs isn't incompetence — it's incentive design.

The problem isn’t that some founders lie. The problem is that lying works — right up until one tweet blows the whole thing up.

The compliance angle is especially dangerous. SOC 2 isn't a vibe. It's a legal backstop. Companies trusting Delve weren't just buying SaaS — they were buying trust. If the compliance isn't real, neither is the enterprise product it claims to protect.

The Delve billboard on the 101 said “compliance is done in days, not months.”

Turns out it wasn’t done at all.

Subscribe now

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Views expressed are my own and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

In a fascinating report released yesterday called Labor Market Impacts of AI, there’s one chart that should make every knowledge worker a little uncomfortable. It maps where AI could theoretically disrupt jobs versus where it’s actually disrupting them right now. The gap between those two lines? That’s either a massive opportunity for founders — or a depressing countdown clock for the rest of us.

How They Measure It: “Observed Exposure”

This is where the paper separates itself from the usual “AI is coming for your job” slop. Anthropic built a new metric called Observed Exposure. Most prior research just asks “could AI theoretically do this task?” Anthropic is asking something sharper: of those tasks LLMs could speed up, which ones are actually seeing automated usage in professional settings right now?

TLDR: Fundamentally, this is a white collar disruption story. This is coming for the highly educated people who thought their education was a moat. Sadly, it’s not.

Note: Grounds Maintenance and Agriculture, here I come. There’s something romantic about the idea of driving landmower and getting zero Slack notifications.

Breaking Down the Chart

The gap between the Blue (Theoretical AI Coverage) and the Red (Observed AI Coverage) is the real story.

• Theoretical AI Coverage (Blue) = what LLMs could theoretically do.

• Observed AI Coverage (Red) what’s actually happening on Anthropic’s platform.

For example - take Computer & Math: 96% theoretical (Blue)… only 32% observed (Red). That means there’s a HUGE amount of disruption runway still ahead.

Now look at the bottom. Grounds Maintenance. Agriculture. Food & Serving. Barely a blip @ 2%! Might be time for a career switch - Food Catering here we come!

Subscribe now

So, who’s Most Exposed?

The top three most exposed occupations:

1. Software Programmers — 75% coverage (no surprise here)

2. Customer Service Reps— 70% (that CS rep you love? Probably an AI Agent)

3. Data Entry Analysts — 67% coverage. Time to pack it up.

On the other end, 30% of workers have zero coverage. Their tasks appeared too infrequently in AI usage data to register. This group includes Cooks, Motorcycle Mechanics, Lifeguards, Bartenders, Dishwashers, and Dressing Room Attendants.

(Note to self: Learn how to cook)

Reality Check: This is not the first-time that Anthropic has caused some market panic. Or the second. Or the third.

The Good News (For Now)

Here’s the thing that keeps this from being a full panic piece: the authors are not finding significant unemployment effects yet.

But here’s the catch: they found suggestive evidence that hiring of younger workers (ages 22-25) has slowed in exposed occupations — about a 14% drop in the job-finding rate for young workers entering high-exposure jobs. The entry-level pipeline is narrowing before the layoffs even show up. That’s a troubling trend.

That’s the pattern to watch. AI might not fire you. But it might mean the company doesn’t hire your replacement. Or your child can’t get that first job out of school.

Bottom Line

Every founder should be staring at the categories where blue dwarfs red and asking: what’s the product that closes that gap?

Subscribe now

Every worker should be looking at where they sit on this chart and thinking honestly about their future career trajectory.

And every policymaker should be struggling with the fact that the most exposed workers are the ones who did everything “right” — got educated, got office jobs, climbed the ladder. It’s a depressing fact for people like me —> VCs might be cooked (unless you ask Marc Andreessen 😂).

The full paper is available from Anthropic’s research page, and the underlying data is on Hugging Face. Happy Friday friends.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Views expressed are my own and should be considered as such, and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

Happy 2026, everyone. January is always a magical time: new beginnings, ambitious OKRs — and sales reps “circling back” about software you’re still not going to buy. February always feels like a return to the grind: the start of ambitious new projects, the announcement of big funding rounds, and the thrill of back-to-back meetings.

(Note: Yes, this post is late - I’ve been heads down with deals. Sorry not sorry!)

Last year, I kicked off Scaling the Enterprise with a set of predictions on agents, data centers, and the products-vs-features reckoning in AI. Most of them aged well. This year, we move on to arguing whether foundation models will eat the application layer.

Here are my Top 10 predictions for Enterprise AI in 2026, written from the POV of someone who spends most of his time thinking: Will an enterprise actually adopt this?

1. AI Governance Becomes a Competitive Advantage

Enterprise AI sprawl is real. Between copilots, agents, foundation models, internal fine-tunes, and third-party AI SaaS, most large enterprises now operate dozens of AI systems simultaneously. But how do you account for AI risk?

Winning organizations will build internal governance councils, cross-functional oversight, clear deployment playbooks, and defined risk thresholds. The enterprises that scale AI fastest will be the most focused and disciplined.

In 2026, I predict the AI Trust Layer becoming mandatory. A few companies I’m tracking in this space include Arize, Trustwise, Geordie AI, Reco AI, and Fiddler AI.

2. AI Might Take Your Job. But You’ll be Fine.

🔥 My spicy take is that AI’s impact on job destruction will be overblown. Yes, AI layoffs are coming. But society overall will benefit massively, and many new jobs will emerge for those willing to learn new, higher value skills.

Take yesterday’s announcement by Block about a 40% layoff of its workforce: while CEO Jack Dorsey was quick to blame AI for the cuts, the reality is that Block had become bloated and inefficient in recent years. The cuts were overdue (Wall Street responded with a 20% stock jump).

A report from McKinsey found that one-third of the jobs we have today didn’t even exist 25 years ago. So even as AI replaces workers in some occupations that are manual and repetitive, new job functions will be created.

To put this into practice, Dave Kellog makes an excellent analogy to elevator operators, a profession that peaked in the 1950s. Does anybody really miss these jobs?

3. Physical AI is next up for the AI Hype Train

The next wave of AI value won’t live inside SaaS dashboards in web apps. It will extend into robotics, logistics automation, manufacturing, grid intelligence, and industrial inspection. VC investment in robotics soared in 2025 — over $22B deployed, a 69% year-over-year increase — and I’d expect funding to stay strong as companies hit key milestones and investor appetite shows no sign of slowing.

This isn’t just about hardware. World models, led by efforts from Dr. Fei-Fei Li and Yann LeCun, aim to encode the physics and spatial properties of the real world. Many believe this will be foundational to both embodied AI and, eventually, AGI.

A few companies I’m tracking in this space include: Skild AI, Physical Intelligence, RLWRLD and Basetwo.

Subscribe now

4. The IPO Window Cracks Open

After a multi-year drought, some of the best-known names in Silicon Valley are gearing up for what could be some of the most anticipated IPOs of all-time. My prediction is at least two names on the list below make their public debut:

• Databricks: ~$150B–$220B estimated value @ IPO

• OpenAI: ~$830B–$1T+ projected

• Anthropic: ~$350B–$500B range

• Stripe: ~$180B–$250B

• SpaceX: $1T -$1.2 T

None of these are guaranteed — although we know OpenAI and Anthropic have no choice but to tap into public markets soon. Companies with such staggering capital needs can’t stay private forever.

5. AI Bubble Concerns Prove Overblown

Valuations will take their medicine. Capital will get more selective. But real usage and adoption stay firmly intact. This is digestion, not collapse.

For the top AI companies, the fundraising party continues. Looking at the top 10 most valuable private companies, it’s remarkable that 3 of the top 5 are pure-play AI firms. For this elite group of AI companies, I don’t see the music stopping anytime soon.

…however for the other 95% of the market, expect serious cracks appear.

6. Cybersecurity Becomes AI’s Most Defensible Category

In 2026, security will remain one of the most durable and fast-growing categories of IT spend for CIOs in 2026.

AI is both the threat and the solution. It’s expanding the attack surface through model poisoning, prompt injection, and more sophisticated social engineering — but it’s also the primary tool for detecting and responding to those same attacks.

A few categories I’m interested in:

• Email + Next-Gen Phishing: LLMs have made spearphishing more scalable and effect than ever. Most legacy tools weren’t built for this new reality.

• Security-Focused Frontier Lab: Can a founder build a best-in-class focused entirely on red-teaming models and setting the evidentiary standard the industry needs as LLM built apps become the norm?

• Agentic Security: Agents in the Enterprise are getting real credentials and taking real actions. But security is still lacking.

7. Power Becomes the Real Constraint

One of the most underappreciated bottlenecks in the AI buildout isn’t compute — it’s power. Data center energy demand is set to double by 2030, and there’s already an estimated 245 gigawatts of U.S. capacity in development or planning. In 2026, both the physical limits of power delivery and rising political pushback over electricity costs will become major points of contention.

Satya Nadella put it plainly: the biggest issue now isn’t a compute glut — it’s the ability to get builds done fast enough, close enough to power sources. Big tech is already moving aggressively to secure long-term energy, striking deals with decommissioned nuclear plants and entering electricity trading markets.

8. The Application Layer does not become Foundation Model Roadkill quite so easily

Another controversial one! I know it’s trendy to say Clade (which helped me with this) is goint to soon be capable of doing any task and every Cursor-like tool will soon be extinct. But I think AI Apps will prove more resilience than you might think.

Many categories benefit from being *multi-model* and this will force users to continue levreaging multiple providers. Plus, there are many categories where AI startups have built proprietary databases of workflow data from years of execution and are 5x better than the AI labs - Open Evidence (Health), Trullion (Accounting) Harvey (Legal), and Rogo (Finance). So this guy below is flat wrong.

The labs are ambitious and formidable but this is the same as saying Amazon will every online purchase or every search will run through Google.

9. Nvidia Competitors are coming…fast

In 2026, I predict we’ll see many more AI companies manufacturing custom AI chips at scale. We’ve all heard the rumors out of China, where an entire industry has emerged with the sole purpose of displacing the West’s global dominance. New entrants like MetaX and Moore Threads have joined tech giants Huawei, Baidu and Alibaba in the race to produce Nvidia alternatives.

In the US market, mong the first movers will be the large AI labs, for whom purpose-built chips will become increasingly critical part of the stack landscape on which to innovate and compete.

2025 saw VC funding for AI/ML semiconductor companies nearly double, jumping from $4.8B to $8.4B according to PitchBook. Companies to watch in this category include Ricursive Intelligence, Etched, Axelera AI, and MatX.

10. Enterprise AI Budgets Grow — But Consolidate Around Fewer Winners

Companies will increase their AI budgets significantly, but spend those dollars on far fewer vendors.

A TechCrunch survey of 24 enterprise VCs found overwhelming agreement. As Andrew Ferguson of Databricks Ventures put it, enterprises are currently testing multiple tools for single use cases — and in 2026, they’ll cut the experimentation budget and rationalize overlapping tools.

Rob Biederman of Asymmetric Capital was more direct: budgets will increase for a narrow set of AI products with clear results and decline sharply for everything else.

For startups, the implication is clear: if your product isn’t mission-critical by the end of 2026, you’re at risk.

Rapid Fire: A few other predictions I’m making

Here’s my 2026 enterprise AI checklist for separating real adoption from theater:

• In 2026, we’ll see the broader shift from seat-based licenses to performance-based economics. AI Agents start getting paid for specific outcomes.

• The market finally sours on “Cursor-for-X” copycats, which continue to be undifferentiated solutions in search of a problem (thanks YC!)

• Espionage from Chinese’s AI Labs and State-Sponsored enterprises grows 10x, creating a massive industry for Sovereign AI cybersecurity tools

  ---

Subscribe now

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

Follow me for more #EnterpriseAdoptionTips

In Part I of this series, we talked about how Perplexity launched Comet, its AI-native browser, in a direct attack on entrenched incumbents like Google Chrome. That move was bold enough. If you would like a recap of Part I, you can check that out below.

But this month, things have only continued to heat up. Perplexity has reportedly offered a staggering $34.5 billion all‑cash bid to acquire Google Chrome—a figure that obliterates its own ~$18 billion valuation. If you thought the browser wars were already over, think again—this movie is just getting good.

Yes, you read that right. A company valued privately at around $18 billion, with less than 2% of browser market share, just announced it wants to acquire Chrome — the crown jewel of Google’s distribution machine. 🚨This is not a drill 🚨

“Interesting to see Perplexity with 1.9% market share at $18B offering $34B for Chrome. While Antitrust pressure may force Google to sell Chrome, it is central, so selling is unlikely. But Perplexity intent is more of PR positioning of Perplexity as contender in AI Search.” — @JabesUSR

If Part I felt like a surprise strike, Part II is the public declaration of the Browser Wars.

Subscribe now

Can David Really Beat Goliath?

If history tells us anything, it’s that you can never count out a small scrappy startup. But Goliath rarely goes down without a fight.

Historical Parallels Part I: Netscape vs Microsoft

If Perplexity’s gambit feels familiar, it should. The history of the internet is chock-full of similar PR stunts and gimmicks. Take Netscape vs. Internet Explorer, for example.

The Netscape web browser was once dominant but lost to Internet Explorer (IE) in the first browser war (literally). In 1998, AOL famously bought Netscape for billions at the peak of dot-com bubble. While this appears to be a great outcome on paper, the deal was massively disappointing to Netscape, which saw its market share fall from over 90% in the mid-1990s to less than 1% by 2006. As a last-ditch effort to fend of Microsoft’s looming IE dominance, Netscape open-sourced its browser code in January 1998. It wasn’t done. Only a few months later in May 1998, the United States Department of Justice filed an antitrust case against Microsoft. While Netscape was technically not a plaintiff in the case, its executives were subpoenaed, and it contributed material to the case. The central argument? Internet Explorer’s $0 price.

Microsoft set a zero price for its browser for the purpose of depriving Netscape of revenue and protecting its operating system monopoly.

Unfortunately, by this time, it was too late. The damage was done, and Netscape’s share of the browser market never recovered.

Historical Parallels Part II: When Challengers Emerged Victorious

While it’s rare for a small tech company to literally "buy" a larger, more established one, there are certainly past examples where smaller, more agile players have successfully challenged and disrupted larger incumbents. While sometimes this leads to the bigger company acquiring the smaller one, often the smaller company's innovation forces a major shift in the larger company's strategy.

A few notable examples from the past:

• Nubank takes on Brazil’s incumbent banks: Nubank, a digital bank startup, identified a gap in the Brazilian banking sector due to high fees and poor service from incumbents. They leveraged smartphone adoption to create a digital-first banking experience, attracting over 90 million customers and reaching a $40 billion market cap, effectively outmaneuvering traditional banks.

• Netflix gets the last laugh against Blockbuster: The Perplexity–Chrome saga also channels one of the most iconic David vs. Goliath battles in modern business history: Netflix vs. Blockbuster. At one time, Blockbuster was the undisputed giant — a $6 billion behemoth with thousands of retail stores and near-monopoly distribution of home entertainment. Netflix? A scrappy mail-order DVD startup that looked like a niche curiosity. When Reed Hastings offered to sell Netflix to Blockbuster for $50 million in 2000, Blockbuster literally laughed them out of the room. Fast forward a decade: Blockbuster filed for bankruptcy. Netflix not only survived — it reinvented how the world consumes entertainment.

  

What’s Really Going On? Perplexity’s Chess Move

Perplexity isn’t alone in reshaping browsers with AI. Edge has Copilot, Opera has Aria, Safari is adding in Apple Intelligence (lol), Brave has Leo, and Firefox is experimenting with OpenAI and Gemini (writes Brian from his Chat-GPT enabled Firefox browser). Everyone is racing to slap “AI” on their browser before its too late.

Still, Perplexity stands out because it’s skipping incremental product announcements and going straight for the jugular. It’s not just saying, “We built a cool AI-browser.” It’s saying, “We might buy the browser you use every day.”

Here are a few other narratives to watch:

• Antitrust Angle: With DOJ cases ongoing, Chrome could be forced to spin out. As many have noted, even seasoned DOJ lawyers are parsing this bid for its antitrust implications. Perplexity’s positioning, then, is smart theater: “If you’re going to break it up, break it with us taking care of it.”

• AI is Eating Traditional Search: Generative AI’s impact on search is undeniable and multifaceted. As consumers shift discovery and commerce to AI-native platforms, the opportunity ahead is greater than ranking in AI search rankings. The agentic internet is reshaping all aspects of consumer behavior, and Big Tech incumbents like Google will need to rapidly innovate to navigate this change.

• Chrome: The Crown Jewel: Chrome isn't just a browser—it’s a distribution engine for Google’s ecosystem. Many argue Chrome functions more as leverage to lock in ad, search, and browser control—not a neutral platform. A forced Chrome divestment would crack open a critical lever in Google’s competitive advantage. Google would likely fight tooth and nail rather than hand it over.

• An Elaborate Recruiting Scheme: Nothing excites top engineers like being on the “crazy but possible” team. This type of aggressive moonshot tactic will surely help Perplexity recruit top AI engineers.

• Brand Play: Attention is currency. A Chrome headline guarantees global awareness, far beyond what a video about a new AI feature would generate. By throwing out a $34.5B number, Perplexity signals to VCs: “Price Gone Up”

Conclusion: Your Move, Google.

Chrome once looked like Silicon Valley’s impenetrable stronghold. Now it has a cocky, well capitalized rookie challenging it for its starting spot with the entire internet watching. Welcome to the new frontlines of the browser wars.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

TLDR Version: Perplexity and OpenAI are launching their own AI-driven browsers to challenge Google. Perplexity has stolen homecourt advantage with the release of Comet.

• Agentic AI comes to the Browser: Comet can actively group tabs, assist with tasks like form-filling, and even operate in a truly agentic fashion by executing tasks like purchases, social media posts, or sending meeting invites on your behalf.

• Browser Wars Begin: All eyes on Google, as Chrome’s dominance (~68% market share) faces an existential threat. With OpenAI planning to launch its own browser soon, the race to own the interface layer of the internet is heating up.

• Browsers become more than Browsers: For over a decade, the browsers have seen only incremental innovation. We're finally entering a new era where browsers will quickly become full-on productivity tools and even human-like AI teammates.

Part I: Perplexity Enters the Chat

Perplexity just dropped its bombshell: Comet, a full-fledged AI browser built on Chromium and deeply integrated with Perplexity’s agentic AI. Initially available to premium “Perplexity Max” subscribers (~$200/month), Comet launches with AI on every tab—summarizing pages, managing tabs, even booking meetings or purchases via its assistant sidebar. In short, very cool stuff - unless you’re a Perplexity Pro subscriber - then you’ll have to wait a little longer to get your hands on Comet.

According to Perplexity, Comet was designed to:

• Personalize by learning your habits and aiding in research;

• Be Powerful, providing support across languages and domains;

• Enhance Productivity by automating tasks and summarizing inboxes or docs

This move ignites what may be the biggest browser shake-up since Chrome was first released by Google - all the way back in 2008! Let’s unpack what it means for consumers, Google, Microsoft, OpenAI… and predict what’s next.

Subscribe now

Overview of Comet from Perplexity

Perplexity's Comet is an AI-first browser that aims to change how people use the web. It enhances traditional search engines by integrating a context-aware AI assistant into the browsing experience, acting as a browser with integrated AI capabilities.

Below is an overview of some of the headline features:

• Agentic Assistant: The Comet Assistant can intelligently observe your page content and context, offering to summarize articles, answer questions, and even compare products you’re browsing. It can also act on your behalf, booking meetings, sending emails, purchasing items, and setting calendar reminders.

• Seamless Context & Tab Management: Say goodbye to tab overload. Comet understands browsing context, intelligently grouping or collapsing tabs, surfacing relevant past content, and closing duplicates to maintain a clutter-free workspace. Are the days of having hundreds of Chrome tabs open gone?

• Voice Support: Comet supports both voice and text interaction, allowing you to “think aloud” to initiate complex tasks—such as comparing shipping rates or cross-referencing news—directly within the browsing context

In essence, Comet isn’t just for search, it’s a full-on digital command center, unifying discovery, analysis, and execution into a single smart interface. Comet is the first mass-market consumer browser designed for an AI-native world. That’s a huge deal.

Key Use Cases for Comet

Perplexity’s Comet is far more than a simple Chromium* fork. Powerful features such a contextual memory, task automation, intelligent shopping support make it a truly unique browser experience. Fully equipped with a deep bag of AI tricks, it’s hard to argue that Comet won’t immediately jump to the front of the AI Browser race.

1. Deep research & summarization: Whether you're reading a dense academic paper or diving into a 15-tab rabbit hole about YC startups, Comet can use AI to instantly digests articles, pages, and tabs.

2. Memory Across Sessions: Unlike traditional browsers, Comet doesn’t forget. It can remember your workflows, preferred sources, and past searches and suggest relevant materials based on your interests. It even has a “flashback mode” to revisit how you arrived at past conclusions.

3. Task automation: From emails to booking, Comet streamlines repetitive workflows and suggest ways to automate tediousz parts of your daily grind.

4. Enhanced multitasking: Keeps context across tabs and remember it. No more losing samples of your flow.

5. Inbox and Calendar: For those working across Gmail or Outlook in-browser, Comet can surface important emails, summarize email threads, suggest calendar slots based on preferences, and auto-draft responses using a personal touch.

6. Intelligent Shopping: While browsing e-commerce or vendor catalogs, Comet can compare products across sites automatically, flag shady or low-reputation sellers, and surface contextual reviews, specs, and discount codes.

*Note: Chromium, for those unaware, is an open-source web browser project, primarily developed by Google, that serves as the foundation for many popular browsers, including Google Chrome, Microsoft Edge, and Brave.

The Death Beam has found a new target: Perplexity HQ

Part II: How Will the Market Respond?

While Perplexity has clearly made the first major move in the Browser Wars, it’s clear this thing is just beginning. Perplexity is aiming to take control of the most valuable real estate in the consumer tech world, the Internet Browser, which has long served as Ground Zero for the attention of mass-market internet consumers.

For this reason, the market response is sure to be fast and furious.

As we speak, the death beam of unlimited AI resources at the disposal of Google and OpenAI is setting their sights on a new target — Perplexity HQ. Now the question becomes: what’s likely to happen next?

Stack Ranking the Browser Wars Favorites

Below, I rank the challengers and incumbents by those most and least likely to emerge as the winner in the AI Browser Wars. Kalshi Market Odds Wen?

The Favorite: Google Chrome (+200 Betting Odds)

• Google’s Stakes: Chrome’s current market share is historically dominant. Losing browser ownership has the potential to threaten Google’s entire business model. So trust me when I say this: Sundar and team will treat Perplexity as an existential threat. They will not go down without a fight, and it would be unwise to bet against them.

With an astonishing 68% of global browser use, Google Chrome currently dominates the global web browser market as of July 2025. Its competitors are essentially fighting for scraps: Safari ~16%, Edge ~5%, Firefox ~2% . But that’s not what makes Google the betting favorite. The chart below tells the full story.

In short, Comet is challenging a technology giant with a $2 trillion+ market cap that has been a dominant internet force for decades. Just like Meta used its immense resources as a weapon to steal talent from OpenAI and other research labs, expect Google to aggressively use its resources to deepen its technology moat, retain customers, and fend off threats from Perplexity and OpenAI.

The Consumer AI Darling: OpenAI (+250 Betting Odds)

• OpenAI’s Stakes: As the current leader in the mass-market consumer AI race, Sam Altman and team stand the second-most to lose with Perplexity’s release. Expect the company to come out firing.

  

According to numerous reports, OpenAI plans to launch its own AI browser in the coming weeks, likely built on Chromium with ChatGPT and a whole slew of advanced features fully baked. Don’t be surprised if the company successfully leverages second-mover advantage and one-ups Comet offering with better features across the board.

For those paying close attention, OpenAI has already been preparing its users for this moment. ChatGPT users might have noticed prompts to use the ChatGPT Chrome extension and use it as the default search option in Chrome. This is all to reinforce the consumer behavior of using ChatGPT for everything - including search.

The Challenger: Perplexity (+350 Betting Odds)

• Perplexity’s Stakes: For Perplexity, it’s clear the company has made a bet-the-company type of move towards owning the internet browser of the future. This is the type of moonshot-like move that recent history rewarded favorably: favorable PR, strong consumer adoption, and piles of venture funding. While this move doesn’t make Perplexity the betting “favorite”, given the immense size of the prize, it doesn’t have to become #1 for this launch to be viewed as successful.

The Field: Microsoft, Brave, Safari & Others (+1000 Odds)

• The Field: Expect Microsoft Edge to roll-out deeper AI integrations via Copilot. Smaller alternatives (Opera, Brave) are likely emphasize to privacy-first AI and niche appeal. Will another AI player emerge in the Browser Wars? Wouldn’t shock me.

  

Final Thoughts

The debut of Comet marks a pivotal moment: the browser is finally evolving from a passive display tool to an AI-enabled productivity suite. Perplexity’s premium $200/month price tag all but confirms this shift is here to stay. With OpenAI's browser launch imminent and Google's response likely coming soon thereafter, the AI Browser Wars are officially underway.

This is more than just another product release. This will truly be a battle for who controls how we navigate digital life by some of the biggest names in Silicon Valley. Stay tuned for explosive competition, privacy debates, controversial regulation, it’s all coming soon to a browser near you.

Stay tuned for Browser Wars Part II next week.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

Is MCP the Future—or Just Another Flash in the Pan?

If you've spent any time in tech Twitter, GitHub, or a VC partner meeting this year, you've probably heard three letters over and over: MCP. The Model Context Protocol has exploded onto the scene as 2025's most hotly debated technology. Proponents hail it as the “USB-C for AI agents”—a universal plug-and-play interface that could finally unlock truly useful autonomous systems. Skeptics say it’s just another overhyped protocol destined for the same graveyard as LangChain plugins and OpenStack. So which is it? A fundamental breakthrough for the agentic era—or a shiny standard that’ll be forgotten by Q4? In this deep dive, we cut through the noise and take stock of what’s actually happening across startups, open source, and Big Tech.

Subscribe now

Reminder: In Part I of my MCP Mini-Series, which you can check out below, we discussed several key topics, including:

1. What MCP is

2. What its origins are

3. Why it’s useful for developers

4. Why MCP has huge implications for enterprise AI adoption

In this post, we turn our attention from theory to practice. What can MCP do for you?

Introduction: So, what can MCP actually do for you?

In just a few short quarters since its introduction by Anthropic, MCP’s ecosystem has exploded, fueled by a vibrant open-source community, support from tech giants, and a large and maturing base of venture-backed startups building around the standard. Seemingly every week in 2025 has brought new startups emerging from stealth, product announcements from AI leaders, and companies announcing MCP support.

And yet, there remain many skeptics among us. What’s the point? Can’t we just use existing LLMs and orchestration tools?

Source: Tumblr, imgflip.com

MCP Use Cases: Current and Future

Incorporating MCP into LLMs and AI agents provides massive advantages over traditional methods. Consider how AI tools were built before — custom functions for database access, file manipulation, version control — with detailed documentation explaining when to use each one. Doesn’t sound too bad, right? However, when trying to leverage tools across different frameworks, builders often need to spend hours rewriting essentially identical code for different projects. Total headache.

🚀Early Use Cases: Where MCP Is Already at Work

While MCP is still in its early stages, several compelling use cases are emerging:

1. Context-Aware Assistants (Enterprise Copilots): Companies are building internal copilots that can “plug into” an MCP layer and instantly understand key workflows, org charts, knowledge bases, and CRM data. Instead of spending months integrating with APIs or retraining models, these copilots simply ingest the MCP context & go.

Image Source: Modelmetry

2. Multi-Vendor AI Orchestration: Enterprises are testing multiple LLM providers simultaneously—OpenAI, Anthropic, Mistral, and others. MCP allows them to abstract the context layer so they can swap out the underlying model without having to re-engineer everything.

3. Security and Governance Layers: MCP is being used to standardize how context is passed into models, making it easier to apply consistent policies around data sensitivity, access permissions, and audit logging.

4. Coding Assistants: Vibe coding tools such as Cursor, Replit, and Windsurf have become extremely popular use cases for MCP. In Cursor, for example, simply open the Cursor editor, click the settings icon (top-right), select "MCP" from the left menu, then click "+ Add new global MCP server" to get started.

🔮 Future Use Cases on the Horizon

As the ecosystem matures, MCP will expand beyond copilots:

1. Composable AI Workflows: Chaining together several narrow models, each grounded in a different slice of enterprise context—HR, finance, customer ops—via MCP, enabling orchestration of complex workflows across domains with minimal glue code.

2. Autonomous Agents with Real Context Memory: MCP could become the memory layer for autonomous agents, helping them retain and reason over context across long-running sessions or even across tools.

3. Plug-and-Play Enterprise Apps: Developers could ship context-aware apps that work out of the box inside any enterprise as long as there’s an MCP-compatible context layer.

🧱 What Else Is Needed for MCP to Thrive?

History is littered with examples of protocols like OpenStack that launched to great excitement and ran out of steam. Additionally complicating matters is that MCP “complements” such as Google A2A and Cisco ACP could theoretically expand to become full-fledged alternatives in the future.

To move from early pilots to widespread adoption, several developments are necessary:

• Better Tooling: Developer-friendly SDKs, wrappers, and devtools that simplify defining, managing, and updating model context.

• MCP-Native Model Support: Broad buy-in across major model APIs, including native support for MCP schemas, would be a significant advancement.

• Cross-Vendor Standards: Similar to how JSON and REST.API enabled interoperability across the internet, MCP needs a community-driven spec that's open, flexible, and widely adopted.

• Human-in-the-Loop UX: Enterprises desire visibility into what context is being used and how it’s influencing outputs. Dashboards and explainability tools are expected to emerge later this year to address this concern.

Examining the MCP Ecosystem: Startups and Tech Industry Adoption

In many ways, the MCP landscape has matured far sooner than expected. Below is a summary of the companies that are working hardest to accelerate MCP’s path to becoming a ubiquitous layer in the AI stack:

🚀Venture-Backed Startups Building on MCP

The number of interesting startups that are choosing to actively build products and infrastructure around the MCP ecosystem is growing every day. See below for a sample of VC-backed startups solving different challenges with MCP:

1. Cursor / Anysphere (SF): This company needs no introduction. AI-powered code editor that added MCP to access internal APIs and docs. Raised ~$1.1 B from investors such as Thrive Capital and a16z.

2. Dust (Paris): An “AI OS” startup helping enterprises build agent workflows. Uses MCP to connect agents with business tools; backed by Sequoia Capital and XYZ.

3. StackOne (London): Secure AI platform supercharging SaaS and AI developers to build real-time enterprise integrations. Supports MCP for complex enterprise use cases. Raised $20M+ GV, Workday, and angels from OpenAI and DeepMind.

4. Aracor AI (Miami): AI legal assistant for contract review and compliance. Raised $4.5M pre-seed; MCP allows it to tap into legal data securely.

5. Nace AI (Palo Alto): AI agents for risk and compliance. Raised $5M from General Catalyst; uses MCP to access audit logs, finance tools, and compliance systems.

Note: List is obviously non-exhaustive.

Honorable mentions include Replit, the popular online IDE, which recently announced MCP support…Sourcegraph, a developer platform, which also incorporated MCP into its Cody code-assistant…mcp.run (Dylibso), an enterprise productivity platform that offers a secure registry of AI tools to manage MCP servers…and Composio, an integration platform for AI Agents & LLMs that offers managed MCP support.

What are some other MCP startups I missed? Feel free to comment your favorites below…

🥇 Support from Tech Incumbents and AI Leaders

It’s rare to see fierce AI competitors (OpenAI, Anthropic, Google, Microsoft, Amazon) align to a single standard, but that’s exactly what’s happened with MCP. Not to be outdone, many tech industry giants in other sectors have followed suit and also hopped on the MCP bandwagon train.

1. Zapier - Workflow Automation Meets the AI Era

An early leader in workflow automation has entered the AI era. Zapier-loyalists can now use the company’s MCP Server to connect AI Agents to over 8,000 apps. Zapier MCP is currently available in beta and free to use.

2. Shiprocket – AI-Integrated eCommerce Enablement

India-based Shiprocket, preparing for its IPO, announced what it calls the “country's first AI-integrated MCP server”. The initiative aims to drive autonomous digital commerce by enabling real-time decision-making and personalized consumer experiences through MCP.

3. Block (Square) – AI-Powered Payment Integrations

Block integrated MCP internally to connect their payment systems and databases with in-house AI assistants. Now, with a dedicated Square MCP server, sellers using the company’s products can also enlist the help of an AI agent to build their business on Square’s entire suite of API resources and endpoints.

4. Cloudflare – Edge Computing and AI Integration

While primarily known for its web infrastructure services, Cloudflare has adopted MCP to enable AI models to interact with edge computing resources, ensuring low-latency and secure data processing.

5. Stripe - Add Payments to Agentic Workflows

You can use the Stripe MCP server if you use code editors that use AI, such as Cursor or Windsurf, or general-purpose tools such as Claude Desktop. By connecting AI Agents to Stripe via MCP, business can also address common payment-related queries from customers using AI instead of human agents.

📊 Current State of MCP: By the Numbers

Looking at the raw numbers, it’s clear that MCP’s explosive growth is driven primarily by the creation of new servers and development tools — the “supply side” of the ecosystem. These are the building blocks that developers make available for AI agents to use: search tools, API connectors, data access layers, and more.

To better understand MCP's trajectory since its inception in November 2024, let's examine some key metrics:

1. GitHub Star Growth for MCP Repositories

The number of active MCP servers has expanded rapidly, showcasing the protocol's widespread implementation across various platforms and services.

Source: star-history.com

Summary: The latest data pulled from GitHub shows the MCP servers repository has witnessed exponential growth in GitHub stars, reflecting the developer community's increasing interest and adoption. The steep growth curve beginning in early 2025.

2. Top MCP Server Classifications

The number of available MCP servers is growing rapidly. The official MCP typescript SDK is currently averaging around 2.8M weekly downloads from npm. Interestingly, data from Madrona shows that actual usage is concentrated. Of Smithery’s 2,500+ MCP servers, only 8 have surpassed 50,000 installs, a classic power law distribution.

Source: Madrona Venture Group

Analysis: The data suggests that supply of MCP offerings is still outpacing demand, with only a handful of MCP servers achieving meaningful adoption out of thousands. Additionally, it’s interesting to note that the “web search” category has the highest number of MCP servers, while “file search” and “code/dev” servers have high installs per server - over 3k each. This suggests that early users are likely to be technical founders looking for practical use cases, as opposed to mass-market ones. Despite the hype, it’s a clear reminder that we’re still early.

Final Thoughts & Takeaways

What’s Next for MCP

It’s too early to tell whether MCP will be the winning standard of the “AI Era”. The AI landscape is moving too fast, with new groundbreaking new AI products seemingly getting released every day. But MCP’s strong early support shows a clear demand for extending agents’ functionality with a well-understood protocol. However, it’s clear from the data that supply clearly outpaces demand, with issues such as security and governance likely holding back broader MCP adoption.

What this Space Might Look like in 2026

While it’s too early to tell what MCP will look in the future - we can have a little fun by making some educated guesses. By mid-2026, here are a few predictions for MCP:

• MCP integration into several open-source frameworks (LangChain, LlamaIndex, Semantic Kernel, etc.)

• Deployment of the first “MCP-native” enterprise copilots in production environments

• Formation of an MCP Global Standard Body— the global version of Anthropic’s MCP Steering Committee

• Emergence of startups addressing tooling gaps: context governance, debugging, data security, schema evolution, etc.

• Large companies updating their Vendor Security Assessments with the following question: “Are you MCP-compliant?”

Will these predictions come true? Only time will tell. I’ll plan to re-examine my predictions this time next year - excited to watch this space in the meantime.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

OpenAI has been busy lately

This morning, OpenAI announced the launch of Stargate UAE, the first international deployment of Stargate, OpenAI’s infrastructure platform, in a unique arrangement that could also include up to $20 billion in investments from the UAE-based investment vehicle G42.

Yesterday, OpenAI made headlines for paying $6.5 billion to buy Jony Ive’s AI device startup, io, in an all-stock deal. It’s by far OpenAI’s largest acquisition to date and comes just weeks after the company agreed to buy AI-assisted coding tool Windsurf for $3 billion. The steep price tag paid for Jony Ive’s startup has raised almost as many questions as the curious picture below that accompanied the announcement…

But let’s get back to Windsurf. The acquisition of Windsurf (aka Codeium), an AI-native coding platform, reflects OpenAI’s urgent need to make a big move. While OpenAI has largely served as the pacesetter in the AI Arms race in recent years, it now finds itself trying to defend its slipping market position while continuing to expand into new frontiers. In the market for AI coding tools, arguably the hottest and arguably most fiercely contested battleground, OpenAI has undoubtedly fallen behind its peers Claude and Gemini - and also was likely caught off guard by the popularity of Agentic coding startups such as Cursor, Replit and Windsurf.

The acquisition was Silicon Valley’s worst-kept secret for weeks before becoming official. And now, for the small price of $3 billion, OpenAI has its prize.

Introduction: What is Windsurf?

Windsurf, known as Codeium until April 2025, is an AI code assistant for developers.

At its core, Windsurf provides AI-powered code completion, refactoring, and search capabilities, enabling developers to write, review, and navigate code more efficiently. These tools ensure that development teams can work faster and smarter while maintaining control and precision in their workflows. Windsurf also integrates deeply into developer workflows while prioritizing security and context awareness.

Model choice is also part of Windsurf’s appeal. Windsurf lets you swap between GPT-4o, Claude Sonnet and Meta’s Qwen models with just a click. But it will be interesting to see whether this model-agnostic approach survives inside OpenAI’s walled garden, or if it will eventually give way to a one-vendor approach.

The Evolution of Windsurf

Now that we know the basics about the company, let’s take a look at how Windsurf first got its start, and what lessons we can learn from its journey….

Windsurf: A Startup Known by Many Names

The company, initially known as Exafunction in 2021, rebranded to Codeium in 2022, focusing on AI-powered coding extensions. Following the launch of their Windsurf Editor in November 2024, the company decided to rebrand again in April 2025. For those following at home, the timeline looks something like this:

• Exafunction: 2021-2022

• Codeium: 2022-2025

• Windsurf: April 2025 to Present

Origin Story: Exafunction

Windsurf was founded in 2021 by Varun Mohan and fellow MIT grad, Douglas Chen. Prior to Windsurf, Chen was at Meta, where he helped to build software tools for VR headsets like the Oculus Quest. Mohan was a tech lead at Nuro, the autonomous delivery startup, responsible for managing the autonomy infrastructure team.

The startup began as a radically different company called Exafunction, focused on GPU optimization and virtualization for AI workloads. Exafunction’s original goal was to reduce AI development costs by abstracting away hardware, allowing customers to run workloads in Kubernetes clusters. See below for a clip from an early TechCrunch article on Exafunction - the company is almost unrecognizable.

But in 2022, Mohan and Chen sensed a bigger opportunity in generative coding and decided to rebrand — and pivot!

From Kubernetes to Vibe Coding

By 2022, the company saw a clearer vision and rebranded to Codeium, integrating vertically into the application layer to bring AI-powered coding extensions to every IDE possible. Codeium launched its beta coding product in October 2022 and hasn’t looked back since.

By the time it raised its $150 million Series C in August 2024, the company supported around 70 programming languages and integrates with a number of popular development environments, including Microsoft Visual Studio and JetBrains.

Entering the Windsurf Era

In November 2024, the company launched the Windsurf Editor, which it referred to at the time as the “first agentic IDE”. After the successful launch of Windsurf Editor, the company felt strongly about its new identity, and thus made its final name change official in April 2025. This rebranding reflects the company's evolution and its broader vision beyond just assisting with code writing, aiming to combine human ingenuity with AI for powerful developer experiences.

Windsurf is Proof that Pivoting Can be Good for Business

A pivot is a fundamental change in the direction or strategy of a company. This could involve changing the target market, introducing a new product or service. Many founders balk at the idea of pivoting their company, believing that is represents a failure. But in reality, pivots aren’t inherently bad, especially if the founders are acting rationally given the market feedback they’ve received on their original idea.

In the words of Albert Einstein, “Insanity is doing the same thing over and over and expecting different results.” We’ve all met many founders who would rather be insane and lose their business than pivot to a new idea that might actually save it.
The Windsurf co-founders sensed a much bigger opportunity in generative coding and decided to bet the farm - and their boldness was rewarded.

Successful Pivots

When looking at today’s most successful companies, it’s easy to forget that many of them began with an initial idea that was radically different. Many successful companies are successful because they pivoted, not in spite of their pivots. A few famous examples:

• Slack began as a gaming company called Tiny Speck. After the company’s primary game, Glitch, failed to perform in the market, it was on the verge of shut down before they were able to pivot to the internal messaging tool we know today.

• Twitter began as a podcasting platform called Odeo. After Apple released iTunes and emerged as an existential threat to the podcasting market, Odeo realized the danger, and started holding internal hackathons in search of new products. The founders pivoted hard into Twitter, and were rewarded handsomely.

• Instagram began as a company called Burbn, a location-based check-in platform, before pivoting to be the photo-focused platform that it is today.

• Amazon began as a bookstore before pivoting to the diversified e-commerce platform we know today. In 2006, Amazon also launched a subsidiary, Amazon Web Services, that offers cloud computing services and is now the major growth driver for the company.

Moving Beyond the Stigma Of Pivoting

Pivots are not an indication of failure. Pivots are the refusal to fail - the willingness to do whatever it takes to make your company succeed. And for that reason, companies like Windsurf should be applauded for taking the bold actions necessary to generate a fantastic outcome for all involved.

For investors, this is another reason why it’s always best to focus on investing in high-quality teams - not ideas. Without execution, even outstanding ideas are worthless. Windsurf is another example of how investing in an A+ team can yield better returns down the road if that team is determined to find success - even if the idea changes.

So, Why did OpenAI do this Deal?

Before wrapping up this post, I’d like to consider one final question: Why did Open AI do this deal? Can’t they just build this themselves?

The answer is not as obvious as it seems on the surface.

Reason 1: OpenAI was late to the Vibe Coding Party

Previously, it was rumored that OpenAI tried to buy Cursor, the leading company in the AI-assisted IDE space, but it couldn’t agree on a price. So, it settled on buying Windsurf, the second biggest player. But what’s so special about these companies?

Both Cursor and Windsurf are known for helping to popularize the recent trend of Vibe Coding. Essentially, Vibe Coding is interacting with a code base through natural language prompts, so that the engineer's role is simply to converse with the tool, rather than needing to pump out millions of code. The term was coined in a tweet by Andrej Karpathy where he described the experience of AI-enabled programming.

Reason 2: Windsurf Has What OpenAI Needs

Still, you might be thinking, “Why would OpenAI buy Windsurf? There are lots of players in the market. And that’s a totally fair question. But Windsurf brings a few things to the table that made it a no-brainer.

1. It’s Already a Go-To Tool – Windsurf has become the go-to coding assistant for a ton of people. In fact, it’s the second most popular one out there, and lots of companies rely on it every day. When you’re already part of someone’s daily workflow, that’s a huge win—and hard for competitors to break into.

2. It Just Works with AI – While Windsurf’s interface might not be totally unique, it really gets the job done. It’s super fast, great at predicting what you need, and can handle big chunks of code without breaking a sweat. Plus, it supports a ton of languages and plugins, so users immediately see value—no hand-holding required.

3. Data, Data, Data – Here’s the secret sauce: Windsurf is learning from real, live code across tons of different use cases. That means their models are getting trained on high-quality, real-world data that most public tools don’t have access to. That kind of edge is hard to copy.

Only time will tell whether OpenAI’s Windsurf move will pay dividends. But the many motivations behind the deal are crystal clear.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

Introduction

Large language models (LLMs) have become incredibly powerful, but they often operate in isolation. One of the biggest challenges in developing AI applications is giving these models the context they need from external data sources (documents, databases, APIs, etc.) in a reliable and scalable way. Traditionally, each new integration between an AI assistant and a data source required a custom connector, creating an ever-growing web of bespoke integrations that’s difficult to maintain.

To address this, developers at Anthropic introduced the Model Context Protocol (MCP) in November 2024. MCP is a universal, open standard designed to bridge AI models with the places where your data and tools live, making it much easier to provide context to AI systems. Despite receiving little fanfare upon its initial release, MCP has recently exploded in popularity – already passing LangChain in GitHub Stars and on pace to soon overcome OpenAI and CrewAI. Major AI labs and open-source communities have both rallied behind MCP in a big way. Why? Let’s dive in below.

1. What is MCP?

Model Context Protocol (MCP) is an open protocol that standardizes how applications provide context to AI models (particularly LLMs). In other words, it’s a framework that defines a common language for connecting AI assistants to external data sources and services. Anthropic aptly describes MCP as “like a USB-C port for AI applications” – a universal connector that lets AI models plug into various tools and databases in a consistent way. Just as USB-C standardized how we connect devices, MCP standardizes how AI systems interface with different data sources and functionalities.

The protocol defines how the AI model can call external tools, fetch data, and interact with services. As a concrete example, below is how Resend, an AI-native email app for developers, uses its MCP server to orchestrate across multiple MCP clients.

Source: a16z enterprise

The purpose of MCP is to break down the silos between AI models and the vast information they may need. It enables developers to set up secure, two-way connections between AI-powered applications and the systems where data lives. For example, with MCP, an AI assistant could retrieve a document from your knowledge base, query a database, or call an external API – all through a single protocol.

2. Origins of MCP

Early Days: From Anthropic, with Love

The MCP protocol was conceived and built by two software engineers at Anthropic, David Soria Parra and Justin Spahr-Summers. If you haven’t seen it, David shared the origin story in a Latent Space podcast episode, and it’s a recommended listen.

When you look closer, you see that the “AI integration” problem is an MxN one. You have M applications (like IDEs) and N integrations.

Source: Latent Space

MCP was developed as a response to a growing problem in the AI field - the lack of a common standard for connectors. Each new pairing required customization, making it difficult to scale and maintain AI systems in real-world applications. This led to what Anthropic engineers call the “M×N problem” — the compounding headache of connecting M different AI models with N different tools or data sources. Seeing this pain point, Anthropic designed MCP to become a universal standard.

Source: Anthropic

Rapid Industry Adoption

The beauty of MCP is that it’s not just for Anthropic or OpenAI’s models. MCP is open and extensible, meaning anyone can adopt it. That’s part of its charm—it’s not a walled garden.

Early adopters like Block (formerly Square) and Apollo.io integrated MCP into their systems during its initial launch. While others in the market took longer to come around, MCP eventually began to pick up steam from all the major IDEs. Companies like Zed, Cursor, and Windsurf (formerly Codeium) announced MCP support in early 2025, with dozens of household names following shortly thereafter.

Source: X (@Sama)

OpenAI officially announced MCP integration with its Agents SDK in March 2025, with CEO Sam Altman tweeting his enthusiasm for MCP. This put the hype into overdrive, as rarely do the Red Sox (Anthropic) & Yankees (OpenAI) agree to play nicely.

3. Why Use MCP?

Why should enterprises use MCP? In short: providing context to AI models has been challenging and MCP offers a strong solution to those challenges.

Why should developers use MCP? See below for my two cents:

1. Less Custom Integrations: Before MCP, if you wanted an AI model to access a SharePoint site, customer database, or Slack, you’d need to implement three different plugins or connectors – each with its own custom code.

2. Context Quality and Relevance: The smartest AI is only as good as the data it has access to. Without easy access to live context, models will give generic or outdated answers because they’re not able to see the most recent data.

3. Interoperability: In an AI landscape that’s constantly evolving, it’s common to experiment with different LLM providers or tools. Without a standard protocol, this means re-integrating data sources and building new connectors each time.

4. Lower Maintenance Costs: Custom integrations don’t scale. Every time an API changes or you want to adopt a new AI model, the work must be redone. This becomes a nightmare to maintain over time.

5. Improved Cybersecurity and Privacy: Providing context often means giving an AI access to sensitive data. Because MCP is an open protocol, you can host MCP servers within your own infrastructure, keeping data within your environment.

4. What does MCP Mean for AI Adoption?

MCP may not be as flashy as the latest LLM, but it’s arguably just as important. As foundational models have gotten more intelligent, the ability for AI Agents to interact with external tools, data, and APIs inherently has become increasingly fragmented.

For developers, this means needing to implement agents with custom business logic for each system agents seek to connect with. That’s why the context integration that MCP provides is huge for developers.

So, what does this mean, in practice? Consider the example below of an AI coding assistant like Cursor or Windsurf. Both of these IDEs are increasingly rolling out agentic capabilities, which allow them to suggest edits to multiple files at once. However, in order to actually take advantage of these capabilities, a small army of bespoke connectors is required, a major buzzkill for AI developers.

Now with MCP, Cursor and Windsurf developers have access to a shared protocol that lets their IDEs access external tools on their behalf. To use a consumer metaphor, AI developers effectively have a universal remote control (MCP) to connect to all of their digital devices and services.

MCP offers a win-win: better performance and capabilities for AI models, and improved efficiency and safety for developers and organizations. By adopting MCP, enterprises can build AI solutions that are easier to maintain.

It’s hard to imagine vendors not scrambling to add MCP servers – already we’ve seen a flurry of announcements by the foundation model companies and startups across the AI tool stack. Whether MCP remains relevent long-term is anyone’s guess, but I’m excited to see how things shake out.

Until then, to MCP or Not (MCP), that is the question.

Note: Stay Tuned for Next Week’s Article: MCP Part II

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

However, the company’s recent press has eliminated any lingering doubts. As I foreshadowed in my first edition on 2025 Predictions, AI startups have finally started to draw increased scrutiny from investors - with customer churn a top issue.

A recent TechCrunch article shows the reality of the problem. Instead of offering a traditional “trial period”, 11x offered customers a break clause, typically at three months, that made it easy for customers to break the contract. Many customers took advantage, leading to some healthy churn.

Note: For what it’s worth, both 11x and a16z have come out and denied the allegations.

To be clear, there’s nothing wrong with this. It’s actually a clever way to potentially lock in long-term deals. But here’s where we have an issue:

The Issue: Churn Happens

Enterprise clients churn after short pilot periods. This is a well-known fact of enterprise sales. Pilots play an important role in the sales cycle by giving enterprises the opportunity to “try-before-you-buy”, and sometimes companies decide that products simply aren’t a fit. In these cases, it’s best for both sides to move on.

Source: Jason Lemkin (Twitter)

But, for some reason, many startups still often recognize these contracts as “ARR” upfront, despite them not being recurring or annual in nature. This can be extremely tempting, especially when trying to fundraise or hit aggressive sales targets, but is unwise for many reasons. This idea is well captured in the tweet above by Jason.

Infographic from 11x’s Series B Press Release (Source: 11x)

The Current State of ARR

While there’s a lot more I can say about the 11x situation, I’m going to resist the temptation to dunk on this particular company, and instead choose to focus on the broader theme at hand. Not all ARR is made the same.

It’s ironic that the tweet above is over five years old. This is an evergreen issue. But in recent years, I’ve noticed that nearly 50% of startups we come across use CARR as a primary metric to run their business (if not, the primary one), believing it to be a better reflection of how their companies are growing.

CARR is defined as contracted annual recurring revenue, whether in production or not yet in production.

For some, reporting CARR has become even more popular than using ARR. At times, the terms are even used interchangeably, along with Annualized Run Rate. This is a mistake.

CARR ≠ ARR

If you’re wondering the difference between “CARR” and “ARR”, you’re not alone.

Put simply, Contracted Annual Recurring Revenue (CARR) includes revenue that's booked but can’t yet be recognized. Often these are deals that customers have verbally “committed” to but aren’t fully signed and aren’t yet in production. Committed ARR is sometimes used synonymously with Contracted ARR.

Subscribe now

Why is this Distinction Important?

When startups look to raise their next round, they often are surprised to find that growth-stage investors ignore “CARR” calculations when analyzing revenue quality. That’s because verbal commitments can easily be cancelled, and future contract start dates easily be pushed back. Unfortunately, many startups find that once their CARR metrics are swamped with ARR, investor interest quickly fades.

The Lesson: Not All ARR Is Made the Same

There are three key lessons that emerge from the analysis above.

1. Know Your ARR Metrics. It’s hard to trust ARR figures from startups today. Always triple check how ARR is being calculated. Annualized run-rate does not mean ARR.

2. It’s not just 11x. Lots of early AI revenue is experimental in nature. Plan accordingly.

3. Don’t fall for the CARR trap. Even for startups in the circle of trust, CARR is being increasingly misclassified as ARR.

(1) Know Your ARR Metrics

ARR is no longer ARR. At this point there are so many different derivations of ARR metrics that’s it’s impossible to keep count. A few of my favorites are below:

1. Break Clause ARR (made famous by 11x) - Annual contracts that customers can “break” three months in. Still reported as ARR.

2. Annualized ARR - Customers pay on monthly contracts (or revenue is non-recurring), yet revenue is “annualized” assuming they will pay through year-end.

3. Contracted ARR - Startup has received a “verbal commitment” they believe is solid. So customer is not paying now, and the product is not yet deployed, but customer has promised to pay in the future.

4. Pilot ARR - Customer agrees to a $250 K, 3-month pilot. The startup reports this deal as $1M in ARR, believing it will convert to an annual deal. Similar to #1.

Takeaway: ARR should only be used if revenue is annual and recurring in nature.

(2) It’s Not Just 11x. Lots of ARR is Now Experimental

When it comes to AI startups, ARR is getting increasingly misclassified, and it’s hurting both startups and investors.

Why Investors Like SaaS

Historically, investors have preferred SaaS business models because ARR is inherently predictable - you know what revenue will be coming for the next 12 months. An added benefit of ARR is that, when calculated correctly, retention tends to be predictable. The key term here is when calculated correctly, because when ARR is misused, it has a huge downstream effect on churn and retention, which destroys the entire value proposition of SaaS.

With AI Startups, ARR Can’t Be Trusted

In recent times, AI startups have begun to claim a lot of experimental revenue (ERR) that lacks the same retention qualities of true ARR. The urge to book experimental revenue can be strong, as it allows startups to show explosive growth. However, mistaking ERR for ARR can lead to tremendous revenue volatility, poor investment decisions, and misaligned expectations that have significant downstream implications.

It’s important for founders and investors to be honest with themselves about the revenue quality of AI startups. If revenue is experimental in nature, it should be classified accordingly until it’s locked in.

(3) Don’t Fall for the CARR Trap

Even for startups in the circle of trust, we’re seeing CARR get misclassified as ARR more and more. Startups must resist the temptation to count pilot revenue and almost-but-not-quite-closed contracts as part of their ARR figure. The health of their startups (and the trust of VCs) will be significantly better off long-term because of it.

Where We Go from Here

In some ways, VCs have created a monster with their fanatical obsession with ARR. Startups know the answer key, and are gaming the system. Something needs to give.

Source: Michael Mignano (Twitter)

Investors need to more closely examine revenue quality during due diligence. What questions to ask specifically? The answer will be different for everyone. But a good place to start would be that whenever you read that a company has grown to $10M ARR in only a few months, be skeptical. Very skeptical.

Until next time.

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of this information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this website or posted to social media and other platforms are my own and are not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).

DeepSeek-R1 launched less than two months ago, and somehow that feels like an eternity ago in the world of AI, which has quickly turned into a global Arms Race.

Recently, my entire social media feed has been filled with posts, videos and comments about Manus AI, a seemingly picture-perfect Autonomous AI agent. The media was quick to declare we were having “another DeepSeek moment” before folks were able to even get their hands on it.

It cannot be disputed that 2025 has seen China rise as a formidable player in the AI game. Bolstered by a strong tech community, forward-leaning AI government policies, and a thriving open-source community, China has put Silicon Valley on notice. But the emergence of Manus, another small startup out of China, so soon after DeepSeek, begs the question: is the US starting to lose its grip over AI supremacy?

Source: theconversation.com

Manus AI: What We Know

According to its website, Manus is a general AI agent that turns your thoughts into actions. It excels at various tasks in work and life, getting everything done while you rest.

Said differently, without the jargon, it’s a general-purpose AI Agent capable of handling a wide variety of tasks in one’s personal and professional life. Below is a selection of use cases - some rudimentary, others exotic - that Manus claims to be able to perform today on its website.

Source: Manus.im

Sounds impressive, right? If we’re to believe what we’re being told, the AI Agent from Manus can handle everything from hedge fund analysis of Tesla stock to Professor-level course writing to booking international travel for executives.

But here’s the catch: It’s not clear Manus can actually do these things well.

Have we entered the era of fully Agentic AI? Not so fast.

In just a few hours of research, I’ve found not one, or two, or three, but over a half-dozen articles, videos, and Reddit threads where people talk about Manus being unable to back up many of the audacious claims it makes on its site. And then I put the platform to the test myself, and found the results largely to be a mixed bag.

The Secret Sauce

Manus claims to be the world's first general AI agent. First things first, there are already a wide-range of AI Agents currently in production at large enterprises. So while it could still emerge as “the best”, Manus AI is certainly not the “first”.

How does it work? The company uses multiple AI models (such as Claude 3.5 Sonnet and fine-tuned versions of Alibaba's open-source Qwen) to orchestrate a team of AI Agents that are able to act autonomously on a wide ange of tasks. While they are not the first to do this, the autonomous nature of the platform does make it different from AI Chatbots, including DeepSeek, which are designed for conversational interactions.

Personal Assistant vs. Force Multiplier: How AI Agents are Different

• AI Chatbots: Collaborative tools designed to assist and work alongside you in a conversational manner

• AI Agents: Autonomous tools designed to take on tasks independently while keeping you in the loop

Manus: Hype or Reality?

My personal belief is that Manus is mostly a marketing stunt to generate some hype. Although the promise of AI Agents is real, the company’s product is clearly not ready for primetime - yet.

Don’t get me wrong - at first glance, I was blown away by the YouTube videos from Manus - and those social media pundits and influencers. The company’s demos looked incredible, and it certainly seemed we were closer to AGI than ever. It’s absolutely deserving of a YouTube rabbit hole (dive in at your own risk below).

A few examples of what Manus claims it can do:

1. Campaign explanation maps that “bring historical events like the Battle of Lexington to Life” (yes, that’s a direct quote)

2. Financial statement review and verification

3. SEO optimization for your website

4. Realistic simulation a bacteria colony after reading the latest research on it

5. Create a video editor that can add effects to a video like wave, pixelate

6. Take a video of a lecture or a research paper and create detailed lecture notes with math and interactive simulations

7. 3D simulation of a futuristic city with traffic, buildings, weather, time

8. Create a candidate interview schedule for 40 candidates to maximize time efficiency based on availability

Conclusion: Manus is not DeepSeek 2.0

If you’re like me, you likely noticed the proliferation of articles recently that claimed things such as “Manus Changes Everything”, “fundamental breakthrough,” and “paradigm shift”. These themes largely aligned with the prevailing market narrative after Manus’ initial release. Color me skeptical.

Why do I believe this? Three main reasons:

1. Technical Problems - The company’s release video make it look like Manus has perfected general-purpose Agentic AI. But when you take these AI Agents for a spin, you’ll see they come with many glitches, often are unable to complete tasks, and are prone to delivering results of marginal value.

2. Little of Transparency — Manus claims to outperform OpenAI’s Deep Research agent, but I had a hard time finding hard proof. Until the company releases independent benchmarks, raw performance results, and evidence of real-world testing, it’s hard to take the company’s claims seriously.

3. Manus may just be a Claude Wrapper - As noted by many X users as well as Techcrunch, Manus leverages many popular LLMs under the hood, such as Claude. It’s unclear what features, if any, are truly proprietary and “gamechanging” vs. nicely repackaged ones from other popular AI companies.

Here’s the Truth: Manus isn’t China’s Second DeepSeek Moment. And that’s okay.

So, why did Manus blow up on the Internet?

A few factors contributed:

1. Exclusivity created by limited invites: The air of exclusivity saw folks, allegedly, selling invite codes for thousands of dollars on Chinese reseller app Xianyu. This created hype and immense anticipation across China and felt all over the world.

2. Most People Believe What they See: The 2016 US Presidential Elections proved that the folks on the internet love to reshare articles without actually watching or validating them. We just can’t help ourselves - we see a buzzy headline and we have to share it. We’ve all done it - including yours truly.

3. Cherry-picked demos: Companies like to only show their greatest breakthroughs and “best hits”, and Manus is no different. But if Manus is as amazing as it claims, why not release raw performance data, and discuss its limitations honestly?

Subscribe now

Disclaimer: The information contained in this article is not investment advice and should not be used as such. Investors should do their own due diligence before investing in any securities discussed in this article. While I strive for accuracy, I can’t guarantee the accuracy or reliability of all information. This article is based on my opinions and should be considered as such, not a point of fact. Views expressed in posts and other content linked on this site or posted to social media are my own and not the views of NextEra Energy Investments (NEI) or NextEra Energy (NEE: NYSE).