Uber reportedly burned through its entire annual AI budget in just four months.

Then it capped employees at $1,500 per month, per agentic coding tool.

Most people will read this as an “AI is too expensive” story.

I think that is wrong.

This is the first clear signal that the enterprise AI market has entered a new phase.

The problem is no longer adoption.

The problem is control.

Companies are not asking, “Will employees use AI?”

They are asking, “Can we scale AI without blowing up the budget?”

In this newsletter, I break down:

1. Why Uber’s AI budget problem is actually an adoption success story

2. Why AI does not behave like traditional SaaS

3. Why enterprises are entering the “AI cost shock” phase

4. The new categories being created around AI spend, control, and ROI

5. Why AI FinOps could become a massive market

6. Why agent observability is becoming critical

7. Why the next AI buyer will increasingly be the CFO

8. Where founders should look for the next big AI opportunities1. AI FinOps

Hey, welcome to AI Market Fit!

Every week you’ll get two things, fast:

1. What’s happening now (and where the money’s headed).

2. Who’s already building it (so you can invest, partner, or steal their playbook).

However, most of the content from this newsletter will be only accessible to the paid subscribers.

If you are serious about building or investing in AI, you can subscribe here:

If you want to sponsor The AI Opportunity or Product Market Fit email me to g@guillermoflor.com

Read more

Arvind Jain built two unicorns before turning 50. His second bet is that enterprises don't need more AI tools, they need one company that makes all those tools actually work.

In this issue

1. Why I’ve been spending months building agent templates and what I learned

2. What Arvind told me about how enterprises actually experience AI right now

3. The horizontal vs. vertical stack split and what it means for founders

4. The real moat in AI companies

5. Five things to carry into your next product or investment decision

A note on this conversation: Glean didn’t pay for this. They reached out because they follow the newsletter. I took the call because Arvind has built two unicorns and his company sits at a uniquely useful vantage point in the enterprise AI stack. Everything I’ve written here reflects what I actually took away which happens to validate things I was already thinking, and challenged me in a few places too. That’s what a good conversation does.

Hey, welcome to AI Market Fit!

Every week you’ll get two things, fast:

1. What’s happening now (and where the money’s headed).

2. Who’s already building it (so you can invest, partner, or steal their playbook).

However, most of the content from this newsletter will be only accessible to the paid subscribers.

If you are serious about building or investing in AI, you can subscribe here:

If you want to sponsor The AI Opportunity or Product Market Fit email me to g@guillermoflor.com

Previous articles you should read:

🔒 The rest of this issue is for paid subscribers.

Read more

Anthropic dropped the cleanest read we have on AI’s labor market footprint. Most outlets are running with the calm headline no unemployment spike yet. They’re anchoring on the wrong number.

The real story is buried in a new metric called observed exposure: how much of a job is actually being done on Claude in work contexts right now. Computer programmers are at 75%. Data entry keyers at 67%. Customer service reps wired directly into API traffic. And entry-level hiring into exposed jobs has already dropped 14%.

This week, I pulled Anthropic’s report together with research from Microsoft, the ILO, Stanford, and the OECD to map exactly where the openings are.

Inside this article I breakdown:

1. What “observed exposure” really measures and why it changes the game

2. Where is the opportunity for founders to build (according to Anthropic’s map)

3. The top 3 occupations getting peeled apart first

4. Three cross-cutting plays that span every vertical

Here’s what I’d give to anyone building out their GTM motion for the first time: GTM Atlas.

Curated by Attio, the AI CRM, it’s the ultimate resource for early-stage teams navigating growth from scratch. You’ll get:

• Systems thinking that scales with you

• A point of view on your GTM, not borrowed from a playbook

• Exclusive stories from industry leaders setting the standard

The first installation features entries from Elena Verna (Lovable), Kyle Norton (Owner.com), Roniesha Copeland (Vercel), James Pastan (Framer), and more.

Mapped by operators. Curated by Attio.

Explore Now!

Hey, welcome to AI Market Fit!

Every week you’ll get two things, fast:

1. What’s happening now (and where the money’s headed).

2. Who’s already building it (so you can invest, partner, or steal their playbook).

However, most of the content from this newsletter will be only accessible to the paid subscribers.

If you are serious about building or investing in AI, you can subscribe here:

If you want to sponsor The AI Opportunity or Product Market Fit email me to g@guillermoflor.com

Previous articles you should read:

🔒 The rest of this issue is for paid subscribers.

Read more

Eighteen months ago, enterprise AI agents were one-off pilots. Today the Big Four and the systems integrators have converged on the same five-move playbook, they’re running it at scale, and they’re packaging the result as a repeatable business.

• PwC has deployed 250+ agents internally.

• EY built 50,000 agents in nine months.

• IBM reports $4.5 billion in productivity gains.

• Accenture is shipping toward 100+ industry agent solutions by year-end.

This is not experimentation anymore. The motion is set, the buyers are committing budget, and the lanes that remain open are narrowing fast.

So I decided to dive deep into the strategy they're following to access the enterprise, and pull out the moves that actually repeat across all six firms.

Here’s what’s inside:

1. What every firm is now shipping, and what it closes off.

2. How internal deployment became the sales asset.

3. The five control points every stack needs.

4. How trust became a wedge.

5. Why workflow beats automation.

Hey, welcome to AI Market Fit!

Every week you’ll get two things, fast:

1. What’s happening now (and where the money’s headed).

2. Who’s already building it (so you can invest, partner, or steal their playbook).

However, most of the content from this newsletter will be only accessible to the paid subscribers.

If you are serious about building or investing in AI, you can subscribe here:

If you want to sponsor The AI Opportunity or Product Market Fit email me to g@guillermoflor.com

Previous articles you should read:

Read more

Two companies sold roughly $90 billion of one drug class last year. That class will hit $200 billion by 2030. Most of the value is still locked behind problems AI is uniquely good at solving: oral delivery, tolerability, manufacturing, and patient matching.

If you are building or investing in AI for biology, this is the most concentrated profit pool in healthcare and the cleanest case study for where AI moves the needle in drug development. Here is what is actually going on, what is already won, and what is still open.

In this issue

1. What peptides actually are (and how they differ from normal drugs)

2. Why peptides are exploding now: the five unlocks

3. The peptide family map: leaders, niches, and frontier

4. What AI can solve that traditional chemistry can’t

5. Is GLP-1 actually the biggest opportunity?

Check it out with the free trial (no strings attached), or jump straight into premium if you’re ready for the full scoop👇

With the premium access you’ll get also access weekly to:

1. What’s happening now (and where the money’s headed).

2. Who’s already building it (so you can invest, partner, or steal their playbook).

Subscribe now

👋 Just a quick note before you continue!

Building great AI systems and agents is quickly becoming one of the most important advantages a company can have. So we’ve been building non-stop and this will become a much bigger part of AIMF and PMF going forward.

From now on, as a PMF + AI subscriber, you’ll get access not just to the best resources on fundraising and growth (160+ Playbooks, 65+ Success Stories, 30+ Pitch Deck Collections, 25+ Investor Resources), but also:

• 🤖 3+ New AI agent Templates Every Week

• 📚 A Growing Library of 20+ AI Agents

Here, few examples:

• The McKinsey Slide Playbook for Claude

• Claude x Clay Playbook

• The Ultimate Guide to Building Skills for Claude

• Claude For Finance: The CompIete Guide

• The Claude Playbook for Reddit Prospecting

• Claude for AI SDR: The Complete Guide

• Linkedin X Claude Agent Playbook

• The Claude Code Founder Guide

Get a 7 day free trial

We’re already 70,000+ founders and operators inside!

Read more

For the last 3/4 years we’ve heard one time after another that AI is going to kill our jobs.

From lawyers to accountants, consultants, level entry jobs, and more, we’ve heard constantly AI will throw them out the window.

1. Dario Amodei, co-founder & CEO of Anthropic

Prediction: AI could wipe out half of entry-level white-collar jobs and push unemployment to 10-20% within 1-5 years.

2. Sam Altman, co-founder & CEO of OpenAI

“Jobs are definitely going to go away, full stop.”

3. Elon Musk, founder of xAI / Tesla / SpaceX

“In a benign scenario, probably none of us will have a job.”

4. Mustafa Suleyman, co-founder of DeepMind, co-founder of Inflection, CEO of Microsoft AI

“Most, if not all, professional tasks” will reach human-level performance.

Jamie Dimon, says people will work less and live longer, but why would people work less?

Now, all these predictions have led popular culture to come up with the concept: the Permanent Underclass.

The theory is simple:

“Some even believe that artificial general intelligence, or A.G.I., will create a permanent underclass. In the United States, the term “underclass” gained currency in the 1960s to describe the factory workers left behind by the postwar automation boom.

Today, it has become repopularized as a viral term for a theory that posits that people have a limited window of time to build wealth before A.I. and robotics are advanced enough to fully replace human labor.

At that point, we will get frozen in our current class positions: The rich will be able to deploy superintelligent machines to do their bidding, and everyone else will be rendered useless and unemployable, left to live off welfare scraps”

In a very summarized explanation, this is what the process of the permanent underclass:

1. AI replaces cognitive labor, not just manual labor.
   It can automate many white-collar tasks that used to provide stable middle-class jobs.

2. AI removes the first rung of the career ladder.
   If companies hire fewer juniors, young workers may struggle to gain experience and move up.

3. AI concentrates wealth among owners and operators.
   The biggest gains may go to people who own capital, companies, data, compute, or distribution.

4. Skills expire faster than people can adapt.
   AI may change so quickly that many workers cannot retrain fast enough to stay competitive.

5. Labor-market exclusion compounds over time.
   Losing access to stable work can lead to worse income, housing, health, networks, and opportunities.

6. Humans risk becoming non-competitive labor.
   In the extreme case, many people remain socially valuable but lose economic value in the market.

Now, this is all theoretical, and there are many reasons why this theory doesn’t make any sense.

As Jensen Huang has said multiple times, when a company becomes more efficient or productive, it doesn’t decide to keep itself small but to hire more people and grow.

Nevertheless, I’ve been playing the scenario in my mind for a while, thinking about how it could be possible to escape the permanent underclass, if such thing existed.

In this newsletter I run you through the systems/strategies I’ve come up to make the most out of this AI Opportunity and avoid getting behind.

For PAID SUBSCRIBERS, Unlock with your free trial or subscribe now!

Subscribe now

💭 Remember!

As a AI Opportunity + PMF subscriber you’ll also get access to all the other Claude resources we put together, such as:

• 1️⃣ The Claude Code Founder Guide

• 2️⃣ want to keep your job? read this.

• 3️⃣ claude just made outbound fully automatable.

• 4️⃣ 🚨 The Ultimate AI Fundraising Copilot (Built With Claude)

• 5️⃣ look what anthropic just released

• 6️⃣ The Ultimate Guide to Building Skills for Claude 🔥

• 7️⃣ The Prompting Skills Guide

• 8️⃣ The Garry Tan (YC) Playbook for Building with Claude

• 9️⃣ Claude. md Full guide

• 🔟 Understanding Claude Dispatch (Complete Guide)

• 1️⃣1️⃣ Claude For Finance: The CompIete Guide

• 1️⃣2️⃣ How I Rebuilt My SEO Strategy Inside Claude (The Playbook)

• 1️⃣3️⃣ The McKinsey Slide Playbook for Claude

We’re already 70,000+ founders and operators inside!

Subscribe now

Read more

Sequoia recently hosted their annual AI Summit: AI Ascent, probably one of the most important AI events in the year.

They brought together more than 150 leading founders and researchers in AI, including Demis Hassabis, Andrej Karpathy, Greg Brockman, Boris Cherny of Anthropic, Dmitri Dolgov of Waymo, Jim Fan of Nvidia, and many more.

I broke down what Andrej Karpathy shared last week.

Today, I want to share my breakdown of Sequoia’s analysis for the future of AI.

👋 Just a quick note before you continue!

Building great AI systems and agents is quickly becoming one of the most important advantages a company can have. So we’ve been building non-stop and this will become a much bigger part of AIMF and PMF going forward.

From now on, as a PMF + AI subscriber, you’ll get access not just to the best resources on fundraising and growth (160+ Playbooks, 65+ Success Stories, 30+ Pitch Deck Collections, 25+ Investor Resources), but also:

• 🤖 3+ New AI agent Templates Every Week

• 📚 A Growing Library of 20+ AI Agents

Here, few examples:

• The McKinsey Slide Playbook for Claude

• Claude x Clay Playbook

• The Ultimate Guide to Building Skills for Claude

• Claude For Finance: The CompIete Guide

• The Claude Playbook for Reddit Prospecting

• Claude for AI SDR: The Complete Guide

• Linkedin X Claude Agent Playbook

• The Claude Code Founder Guide

We’re already 70,000+ founders and operators inside.

Get 7 day free trial

PART ONE — PAT GRADY: CALIBRATION

1. The Waves of Technology

Modern technology has come in layered waves, each building on the one before.

• Silicon-based transistors gave Silicon Valley its name.

• Transistors were assembled into systems.

• Systems were connected by networks.

• Networks went public as the internet.

• The internet supported applications like social media in the cloud.

• The cloud reached our pockets through mobile devices.

Today all of those layers power something indistinguishable from magic: artificial intelligence.

All of these waves are additive.

This AI wave is different in 3 ways:

1. It’s the biggest one yet

During the first 15 years of cloud, software total addressable market grew from roughly $350B to $650B, with cloud capturing about $400B of that.

AI adds an entirely new layer beneath: Sequoia estimates a roughly $10 trillion services market is now potentially addressable by machines.

The exact figure is uncertain — it could be $5 trillion or $50 trillion — but the magnitude is clear. Legal services in the United States alone is a $400B market: one vertical, one country, equal in size to the entire global software industry.

2. Is the fastest yet

The companies that crossed $1B+ in revenue from the cloud, mobile, and AI tectonic shifts are arriving faster with each successive wave. At Sequoia's current trajectory mapping, more billion-dollar AI businesses are likely to emerge in a fraction of the time it took during previous platform transitions.

3. Why This Wave Is Different

There are 2 types of revolution:

• Communication: information distribution

• Computation: how information is processed

Grady identifies three discontinuous moments that define the current trajectory of AI. The first was November 2022, when ChatGPT revealed the power of pre-training to a mainstream audience.

The second came roughly two years later with o1, which surfaced a second scaling law around inference-time compute, reasoning models.

The third is happening now, with Claude Code and the Opus 4.5 / 4.7 generation, which have demonstrated the power of long-horizon agents.

The first two inflections feel like points on a continuum. Between the second and third there is a hard break, a discontinuous shift in what software can do.

Grady offers a deliberately commercial, rather than technical, definition of AGI: if a system can be dispatched to perform a job, recover from failure along the way, and persist until the job is complete, it functionally qualifies.

From Sequoia's perspective as investors, what matters is the practical capability, not the label. For the past several years, AI produced faster horses, applications that made knowledge workers 10–40% more productive without changing the shape of their work.

The current generation produces cars, applications that deliver 10–40x improvements and fundamentally restructure the work, the workflow, and the organization.

AI changes the way that we work, the nature of the work and the nature of the organization.

Why does this matter?

Echoing the famous question of Sequoia founder Don Valentine — so what? — Grady argues that what matters is that the race for the $10 trillion services opportunity is now underway.

Two strategies are visible.

1. Foundation model labs are advancing tech-out, expanding capability and pushing into applications from the model layer down.

2. Startups (largely Sequoia's portfolio shape) are advancing customer-back, building specialized applications on top of the labs and wrapping themselves around real customer workflows.

Driving a car is nothing like riding a horse, and building one is nothing like caring for one, the strategic playbook for this race looks fundamentally different from prior platform shifts.

2. The pillars to build on top of models: Moats, Affordance and Diffusion

1. Moats

Grady's most counterintuitive piece of advice: in a computation revolution, capabilities change faster than customers do.

The instinct of most technical founders is to chase the technology — to optimize endlessly against the latest model release.

But that's where the ground is least stable. Sequoia's recommendation is to wrap tightly around specific customers and their specific problems.

Product still matters; best product generally still wins. But customer intimacy, workflow integration, and earned trust are the moats that compound over time even as the underlying tech keeps shifting.

In a revolution in computation the things that you build might today might be irrelevant tomorrow.

The degree you wrap around your customers is going to be way more valuable.

2. Affordance

Affordance, borrowed from the design world, is the quality of an object that signals how it should be used without explanation.

A hammer has high affordance, its shape suggests its function instantly. Grady notes that raw foundation model access has low affordance: a terminal interface with Claude Code is enormously powerful, but the average enterprise employee has no idea how to make use of it.

The opportunity at the application layer is to take immense underlying capability and package it into surfaces so intuitive that the customer reaches the desired outcome with minimal effort.

3. Diffusion

Grady highlights a widening gap between the rate at which new AI capabilities are created in the labs and the rate at which those capabilities diffuse into real enterprises. Foundation models advance daily; the average Fortune 500 absorbs change in months or years.

Every day that the gap widens, the application-layer opportunity grows.

In Sequoia's view, closing that diffusion gap — turning frontier capability into deployed business value — is the core mission of the application layer.

PART TWO — SONYA HUANG: AGENTS, THE STORY OF 2026

1. The Year of Agents

Sonya Huang opens the second section with a clean thesis: if 2022–2024 was defined by chat interfaces and 2024–2025 by reasoning models, 2026 is defined by agents. The transition, in her framing, is from systems that respond to a prompt to systems that pursue a goal.

2. Flashback to 2022 — AutoGPT and BabyAGI

Huang reminds the audience that the idea of an autonomous AI agent is not new. In 2022, projects like AutoGPT and BabyAGI became overnight GitHub sensations by wrapping GPT-3 in a planning loop and pointing it at a goal.

The concept was right; the execution was not. The agents looped, hallucinated, and failed repeatedly.

The lesson she draws in retrospect is that agents were a known destination years before they were viable, the models simply weren’t yet capable of sustaining performance.

3. The Three Functional Components

Huang breaks agents into three functional capabilities.

1. Reasoning and planning provide the baseline intuition required to think through a task.

2. Action-taking provides the means to do something in the world — call APIs, run code, search the web, write files.

3. Iteration provides persistence across long time horizons. The combination is what produces agency: the practical ability to get things done.

4. The Sliding Scale of Agentness

Huang argues that agentness is a spectrum, not a binary, and uses coding as the canonical example. First came tab autocomplete in 2023 — an AI suggestion in the editor, useful but not transformative. Then came agentic development, where a human directs an agent or coordinates a small team of them. Now emerging are background and async agents— long-running workers that operate without continuous human supervision, often spawning sub-agents of their own. At the frontier sit dark factories: pipelines in which human review has been removed entirely. Huang notes that dark factories sound implausible until they appear in production, which is already happening in narrow domains such as cybersecurity. The same trajectory — assistant to managed intern to self-managing intern to trusted autonomous worker — is repeating across every domain in which agents are being applied.

5. Services Is the New Software

Huang flags this as the most important takeaway of her section, attributing the framing to Pat Grady and to fellow Sequoia partner Julian, whose article on the topic is referenced.

The largest commercial implication of capable agents is that services, not software, becomes the addressable opportunity. In medicine, an agent can analyze a genome, recommend interventions, prescribe medications, and identify clinical trials. In law, an agent can negotiate contracts, perform litigation work, and settle disputes. In the sciences, agents are solving Erdős-class math problems and proposing novel superconductors. In consumer life, personal agents are taking over inboxes, calendars, financial accounts, and tax filings. The economic activity once captured by professional service firms is now being absorbed into the application layer.

PART THREE — CONSTANTINE BUHLER: WHAT'S NEXT

1. Two Kinds of Work

Constantine Buhler picks up where Grady left off, offering a second bifurcation to complement the computation vs. communication framing from the opening. There are two fundamental kinds of work. Physical work moves matter through space — the package on the Pony Express, the satellite on a Falcon 9. Cognitive work moves ideas through minds — Pythagoras formulating his theorem, DeepMind solving protein folding. The two are profoundly different in substance, but Buhler argues their revolutions follow a strikingly similar arc.

2. The Industrial Revolution as Precedent

Buhler walks through the industrial precedent. For millennia, virtually all physical work serving humans was performed by muscle — human or animal. The transition began around 1700 with water and wind, accelerated through steam and combustion, and reached its modern form with electric motors. By 2026, more than 99% of physical work done on behalf of humans is performed by machines: the aircraft that move people, the manufacturing systems that produce goods, the logistics networks that deliver them. Muscle has been almost entirely displaced by machine.

3. The Cognitive Revolution

Buhler’s central claim is that a parallel transition is now underway in cognition. For most of history, nearly all thinking on Earth was performed by biological brains, with a small assist from animals and mechanical aids like the astrolabe or the clock. Electronic computation began to shift this balance in the twentieth century — at any moment today, trillions of calculations are running in service of human needs. The neural network, in his view, is the next wave. The trajectory suggests that within a generation, 99.9% of cognition on Earth will be performed by machines. The cognitive revolution will resemble the industrial revolution in shape, but will be substantially larger in scope and substantially faster in tempo

4. What the future looks like? 4 Stories

1. Aluminum and the Washington Monument

Buhler illustrates the coming devaluation of cognitive skills with the story of aluminum. In the mid-1800s, the Washington National Monument — then the tallest building in the world — was capped with 100 ounces of aluminum. Aluminum was, at the time, the most precious metal known; it was displayed at Tiffany’s in Manhattan before being installed. Within a few decades, the invention of electrolysis allowed aluminum to be separated cheaply from ore, and the metal became disposable — candy wrappers, sandwich foil, packaging discarded after a single use. Buhler’s analogy: aluminum is intelligence, electrolysis is artificial intelligence. Capabilities that once required decades to acquire — PhD-level expertise in narrow domains — are becoming instantly invocable and effectively disposable.

2. The Age of Alien Design

Buhler’s second story addresses what AI-led design will look like. The built environment today is overwhelmingly optimized for human cognition because humans have been doing the cognition. When machines take over the design process, the outputs begin to look strange. In 2006, NASA used an evolutionary algorithm — a precursor to modern reinforcement learning — to optimize an antenna for a satellite mission. The conventionally designed antenna was a clean, symmetric, geometric form. The machine-designed antenna was a contorted, asymmetric shape that no human engineer would have proposed — and it dramatically outperformed the conventional design. Buhler’s forecast is that as AI takes over the design of chips, vehicles, buildings, and software, the outputs will increasingly look alien by human standards. The aesthetic and structural intuitions that have governed human design for centuries may not apply.

3. Emerging Sciences

Buhler’s third story is about the scientific revolution he expects to follow the engineering one. The mature engineering disciplines often emerge long after the practical technology they describe. For nearly a century, steam engines were perfected through pure tinkering — engineers like Newcomen and Watt iterating empirically without a unifying theory. Then, in the early 1800s, Sadi Carnot formalized thermodynamics, giving the field a rigorous foundation that ultimately enabled the next several generations of progress. AI, Buhler argues, is currently in its tinkering phase. The systems work, but the science of why they work is incomplete. His forecast: within the next two decades, a new fundamental science of intelligence — at the level of thermodynamics in scope and rigor — will emerge. It will be taught in high schools. It may provide the conceptual tools needed to genuinely master AI, and possibly to understand consciousness itself.

4. The Art of Unreason

Buhler’s fourth story addresses how human creative activity tends to respond to mechanization. For tens of thousands of years, visual art trended steadily toward realism. From cave paintings to Egyptian hieroglyphs, Greek pottery to Renaissance masterworks, the dominant pursuit was capturing the world as the eye perceived it. The arrival of the daguerreotype and early photography abruptly made that pursuit obsolete: a camera could capture a scene in seconds with more fidelity than a master could achieve in months. Many predicted the end of painting. Instead, painting reinvented itself by asking a different question — not how the eye sees the world, but how the heart and the soul see it.

Impressionism, Expressionism, Cubism, and Neo-Expressionism followed. Buhler’s expectation is that a similar reorientation will occur as machines take over the literal execution of cognitive work.

Human creative and interpretive activity will shift toward the dimensions machines cannot replicate

Man Is the Measure of All Things

Buhler closes with a line from the Greek philosopher Protagoras, written twenty-five centuries ago: “man is the measure of all things.” His reading: nothing in a vacuum has value in itself. Not aluminum, not art, not intelligence.

Value is conferred by human experience and human meaning. AI can perform the work. AI will perform the work. But only human relationships and human attention supply the reason any of it matters.

A decade from now, in Buhler’s view, the substance of daily work will look unrecognizable compared to today.

What endures is the human side, the relationships formed during this transition, among builders, collaborators, and the communities that gather to make sense of moments like this one.

Three forces now define this cycle: a technology floor that moves weekly, an addressable opportunity roughly 10x larger than prior platform shifts, and a playbook that breaks from cloud and mobile precedent.

Hope this was valuable,

Cheers,

Guillermo

A few weeks ago, I wrote about Sequoia’s big claim:

The next $1 trillion company won’t sell software. It will sell the work.

This week, OpenAI made a move that confirms exactly that.

They launched the OpenAI Deployment Company: a forward-deployed engineering unit that goes inside enterprises and builds bespoke AI systems from scratch. And they backed it up immediately by acquiring Tomoro, an AI consulting and engineering firm founded in 2023 that already counts Fidelity, Virgin Atlantic, and DPD among its clients.

This is not a side product. This is a full structural bet on the services model.

We’ve been this coming along for a couple of years now, but the AI Consulting opportunity is still there.

So today I want to go further than the thesis and give you something more useful:

a concrete playbook, sector by sector, for how to deploy AI the way these companies actually do it.

In this deep dive, you’ll get a concrete deployment playbook for each of the sectors Sequoia flagged as the biggest opportunities in AI services:

1. Insurance brokerage: how to replace the broker entirely, starting with standard commercial lines

2. Claims adjusting: how to own the other side of the insurance policy

3. IT managed services: how to sell “your IT runs” as an outcome, not a tool

4. Payroll and compliance: how to automate the most repetitive back-office function in every company

5. Accounting and audit: how to close the books without a CPA shortage slowing you down

6. Healthcare revenue cycle: how to own medical coding and billing without touching clinical decisions

7. Tax advisory: how to use multi-jurisdiction complexity as a moat, not a barrier

8. Legal transactional work: how to become the firm that drafts the contracts, not the software lawyers use

9. KYC/AML: how to own financial compliance workflows before the incumbents wake up

10. Mortgage origination: how to replace the loan processor, not build software for them

11. Paralegal and LPO: how to go after the outsourced legal work that never makes headlines

12. Supply chain and procurement: how to find money clients didn’t know they were losing

13. Recruitment and staffing: where the autopilot wedge actually exists, and where it doesn’t yet

14. Management consulting: how to disaggregate the intelligence from the judgement and own the former

Each section covers the entry point, the expansion path, the pricing model, and who is already building it. By the end you’ll have a clear picture of where to start and how to build from there.

👋 Just a quick note before you continue!

Building great AI systems and agents is quickly becoming one of the most important advantages a company can have. So we’ve been building non-stop and this will become a much bigger part of AIMF and PMF going forward.

From now on, as a PMF + AI subscriber, you’ll get access not just to the best resources on fundraising and growth (160+ Playbooks, 65+ Success Stories, 30+ Pitch Deck Collections, 25+ Investor Resources), but also:

• 🤖 3+ New AI agent Templates Every Week

• 📚 A Growing Library of 20+ AI Agents

Here, few examples:

• The McKinsey Slide Playbook for Claude

• Claude x Clay Playbook

• The Ultimate Guide to Building Skills for Claude

• Claude For Finance: The CompIete Guide

• The Claude Playbook for Reddit Prospecting

• Claude for AI SDR: The Complete Guide

• Linkedin X Claude Agent Playbook

• The Claude Code Founder Guide

We’re already 70,000+ founders and operators inside.

Get 7 day free trial

Read more

Last year, Sequoia AI Ascent felt like the clearest map of the AI market.

The big ideas were obvious but important:

• AI was not just another software cycle.

• The application layer could capture massive value.

• Agents were becoming the next platform shift.

• And the best founders had to move from building tools to building systems that do real work.

One year later, Sequoia opened AI Ascent 2026 with Andrej Karpathy.

That choice matters.

Karpathy is not just another AI commentator. He helped cofound OpenAI, led important AI work at Tesla, helped make concepts like Software 2.0 mainstream, and last year gave the world the phrase “vibe coding.”

But this talk was not just about vibe coding.

It was about something much bigger:

Software is changing again.

Karpathy’s core point is that we are moving into Software 3.0.

• Software 1.0 was code written by humans.

• Software 2.0 was neural networks trained with data.

• Software 3.0 is software programmed through prompts, context, agents, tools, memory, and verification.

That sounds abstract, but the implications are very practical.

This is the full breakdown of Karpathy’s opening conversation at Sequoia AI Ascent 2026.

We’ll cover:

1. What Software 3.0 actually means

2. Why Karpathy says he has never felt more behind as a programmer

3. Why some AI apps are temporary wrappers around model limitations

4. Why verifiability explains where AI will automate first

5. The difference between vibe coding and agentic engineering

6. Why the best founders should build for agent-native workflows

7. And why understanding, taste, and judgment become more valuable as intelligence gets cheaper

8. Vibe coding raised the floor. Agentic engineering raises the ceiling

9. Hiring also has to change

10. Human taste, judgment, and understanding become more valuable

11. The agent-native world is coming

12. The final lesson: intelligence gets cheap, but understanding stays scarce

👋 Just a quick note before you continue!

Building great AI systems and agents is quickly becoming one of the most important advantages a company can have. So we’ve been building non-stop and this will become a much bigger part of AIMF and PMF going forward.

From now on, as a PMF + AI subscriber, you’ll get access not just to the best resources on fundraising and growth (160+ Playbooks, 65+ Success Stories, 30+ Pitch Deck Collections, 25+ Investor Resources), but also:

• 🤖 3+ New AI agent Templates Every Week

• 📚 A Growing Library of 20+ AI Agents

Here, few examples:

• The McKinsey Slide Playbook for Claude

• Claude x Clay Playbook

• The Ultimate Guide to Building Skills for Claude

• Claude For Finance: The CompIete Guide

• The Claude Playbook for Reddit Prospecting

• Claude for AI SDR: The Complete Guide

• Linkedin X Claude Agent Playbook

• The Claude Code Founder Guide

We’re already 70,000+ founders and operators inside.

Get 7 day free trial

1. Karpathy felt behind because the agentic shift became real

Open with the surprising line:

Andrej Karpathy, one of the people who helped build modern AI, said he has never felt more behind as a programmer.

Not because he forgot how to code.

Because the tools crossed a threshold.

His point: around late 2025, coding agents stopped being “helpful autocomplete” and started producing large chunks of useful code that often worked without correction.

This matters because it changed the emotional experience of programming:

• before: the model helped with snippets

• now: the model can push entire workflows forward

• before: you corrected the model constantly

• now: you increasingly supervise it

• before: AI felt like a tool

• now: it feels closer to a junior team

2. Software 3.0 means prompts are becoming programs

Karpathy’s framework:

Software 1.0: humans write explicit code
Software 2.0: humans train neural networks with data
Software 3.0: humans program models through context

The important part is not “prompt engineering.”

The important part is that the context window becomes the new programming surface.

You are no longer only writing deterministic instructions for a computer. You are giving context to an intelligent interpreter that can read, reason, call tools, inspect environments, debug errors, and adapt.

This changes what building means.

In Software 1.0, you write a shell script.

In Software 3.0, you give the agent a piece of text and say:

“Install this. Inspect my environment. Fix what breaks. Make it work.”

3. The OpenClaw installer shows why agent-native software will look different

Karpathy’s example: installing OpenClaw.

In the old world, you would write a complex bash script that tries to handle every environment, platform, dependency, and edge case.

In the new world, the installation is just a block of instructions for an agent.

The agent reads the instructions, looks at your computer, understands what is missing, runs commands, sees errors, and adapts.

That is much more powerful than a rigid script.

The bigger point:
Most software today is still built for humans clicking buttons and reading docs.

But agents need something else:

• clear instructions

• structured context

• machine-readable documentation

• direct permissions

• APIs and workflows designed for agent execution

• fewer human-only menus and settings pages

Founder opportunity:
The next wave of infrastructure will be agent-native. Not user-native.

4. MenuGen shows the biggest trap: building apps that should not exist

This is the best story in the talk.

Karpathy built MenuGen: an app that lets you take a picture of a restaurant menu and generate images of the dishes.

The traditional app flow was:

• upload photo

• OCR the menu

• extract the dishes

• generate dish images

• rebuild the menu

• display everything in a UI

Then he saw the Software 3.0 version:

Take a photo of the menu. Give it to Gemini. Ask it to overlay images of the food directly onto the menu.

No app.
No OCR pipeline.
No UI.
No complex workflow.

Just model input and model output.

Key insight:
A lot of AI apps are temporary wrappers around model limitations.

As models get better, entire product categories collapse into a prompt.

This is the most important founder lesson in the talk:

Don’t just ask what AI can help you build faster. Ask what AI makes unnecessary.

5. The next obvious companies will come from things that were previously impossible

Karpathy says the exciting part is not just speeding up existing workflows.

It is creating things that could not exist before.

Example: an LLM-generated knowledge base.

Before AI, there was no normal software program that could take a messy pile of documents, understand them, restructure them, rewrite them, and turn them into a useful wiki.

Now that becomes possible.

This is the category to watch:

information transformation.

Not just search.
Not just chat.
Not just summarization.

But systems that can take unstructured information and recompile it into new forms:

• company wikis

• investor memos

• research maps

• product specs

• operating manuals

• market maps

• internal training systems

• strategic briefs

• personal knowledge bases

6. Verifiability explains where AI will automate first

Karpathy’s framework on verifiability is one of the most important parts of the talk.

Traditional software automates what you can specify.

AI automates what you can verify.

That is why AI is so strong in domains like:

• code

• math

• tests

• security

• data workflows

• certain research tasks

• structured business processes

These domains are easier because the model can be trained with clear feedback.

The model can try something, get a reward, and improve.

But this also explains why AI feels so weird.

It can refactor a huge codebase and still fail at a simple common-sense question.

That is what Karpathy calls jagged intelligence.

Key insight:
AI capability is not evenly distributed. It spikes in places where labs have data, rewards, and verification loops.

7. The founder playbook: find valuable verifiable environments

The big founder question is:

Where can you create a valuable verification loop that the foundation labs are not focused on yet?

The labs are already going hard after coding, math, reasoning, and general agents.

But there are many high-value business domains where outputs can be verified and models can improve:

• finance operations

• tax workflows

• compliance checks

• insurance claims

• procurement

• contract review

• medical admin

• customer support QA

• cybersecurity

• internal reporting

• sales operations

• logistics

• accounting

• enterprise data cleaning

The opportunity is not “build another wrapper.”

The opportunity is:

1. Find a valuable workflow

2. Break it into tasks

3. Define what good output means

4. Create verification loops

5. Collect edge cases

6. Fine-tune or orchestrate models around that domain

7. Build the system that gets better every time it runs

Key insight:
The next strong AI companies may look less like SaaS products and more like domain-specific reinforcement learning environments.

8. Vibe coding raised the floor. Agentic engineering raises the ceiling

Karpathy makes a clear distinction.

Vibe coding is when anyone can build software by describing what they want.

It raises the floor.

A non-technical person can now create apps, websites, tools, automations, prototypes, and internal systems.

But agentic engineering is different.

Agentic engineering is about using agents while preserving professional quality.

That means:

• no security holes

• no messy architecture

• no broken abstractions

• no fragile systems

• no random code nobody understands

• no “it works somehow” production software

The human is still responsible for quality.

Key insight:
The best engineers will not be the ones who write every line of code. They will be the ones who can direct agents without letting quality collapse.

9. Hiring also has to change

Karpathy makes a very sharp point:

Most companies are still hiring engineers with old-world tests.

Small coding puzzles made sense when the job was writing code line by line.

But if the job is now agentic engineering, the test should change.

A better test might be:

Build a large project with agents. Make it secure. Make it work. Then let other agents try to break it.

That tests the real skill:

• problem decomposition

• taste

• architecture

• security thinking

• agent management

• quality control

• debugging

• product judgment

Founder implication:
The best AI-native engineers may not look best in traditional coding interviews.

They will look best when asked to ship large, real systems with agentic tools.

10. Human taste, judgment, and understanding become more valuable

Karpathy’s strongest point near the end:

You can outsource thinking.
You cannot outsource understanding.

Agents can write code, generate drafts, call tools, and execute tasks.

But humans still need to decide:

• what is worth building

• what good looks like

• what trade-offs matter

• what the system should optimize for

• what should be simple

• what should be secure

• what should not exist at all

This is where taste matters.

Karpathy gives the example of an agent making a strange product decision: matching Stripe payments to Google accounts through email addresses instead of persistent user IDs.

The code may work.

But the design is wrong.

11. The agent-native world is coming

Karpathy’s future world is simple:

Everything gets rewritten for agents.

Today, most software assumes a human user:

• read this doc

• click this button

• go to settings

• configure DNS

• copy this API key

• paste this value

• deploy here

That is painful for agents.

The agent-native version is different:

• agents get clear instructions

• tools expose machine-readable interfaces

• permissions are explicit

• workflows are decomposed into sensors and actuators

• docs are written for agents first

• agents talk to other agents

This is where the world is going:

“I’ll have my agent talk to your agent.”

12. The final lesson: intelligence gets cheap, but understanding stays scarce

End with the education point.

When AI gets better, the temptation is to learn less.

Karpathy argues the opposite.

Understanding becomes the bottleneck.

You still need enough depth to direct the system. You need to know what to ask, what to inspect, what to reject, and what matters.

AI can help you think.

But it cannot replace the need to understand.

Hope this was valuable!

Cheers,

Guillermo

Here are the facts:

AI is no longer just a software market.

It is becoming an industrial market.

Every AI agent consumes tokens.
Every token consumes compute.
Every unit of compute needs chips, memory, networking, data centers, power, cooling and capital.

That is why compute is becoming the new oil.

But the opportunity is not just Nvidia.

The opportunity is the entire compute supply chain.

A CRM shouldn’t obfuscate your data. It should remove entire layers of busywork. Attio is the AI CRM built to do exactly that.

Connect your email and calendar, and Attio turns every customer interaction into context it can reason over and act on. Whether you’re running product-led growth or enterprise sales, Attio gives you the building blocks to run your ideal GTM system.

Then Ask Attio to:

• Analyze sales calls and surface patterns across hundreds of conversations

• Prep for any meeting with full context on what was said and what’s at risk

• Draft personalized post-call follow-ups in seconds

Ask more from CRM. Ask Attio.

Try Now!

Most of the headlines we see are about the obvious part of the AI boom: another model launch, another giant funding round, another AI startup raising at a crazy valuation.

But there may be a safer and potentially huge opportunity underneath all of that.

The real question is not only who builds the best AI app.

It is who owns, finances, powers, connects, rents or optimizes the compute those AI apps need to run.

So in this newsletter, I’ll break down the compute market in a simple way: the main layers, the companies exposed to each layer, and where the biggest investment opportunities may be.

Let me know what you think!

👋 Just a quick note before you continue!

Building great AI systems and agents is quickly becoming one of the most important advantages a company can have. So we’ve been building non-stop and this will become a much bigger part of AIMF and PMF going forward.

From now on, as a PMF + AI subscriber, you’ll get access not just to the best resources on fundraising and growth (160+ Playbooks, 65+ Success Stories, 30+ Pitch Deck Collections, 25+ Investor Resources), but also:

• 🤖 3+ New AI agent Templates Every Week

• 📚 A Growing Library of 20+ AI Agents

Here, few examples:

• The McKinsey Slide Playbook for Claude

• Claude x Clay Playbook

• The Ultimate Guide to Building Skills for Claude

• Claude For Finance: The CompIete Guide

• The Claude Playbook for Reddit Prospecting

• Claude for AI SDR: The Complete Guide

• Linkedin X Claude Agent Playbook

• The Claude Code Founder Guide

We’re already 70,000+ founders and operators inside.

Get 7 day free trial

The market map

Read more

The agentic stack has six greenfield lanes. Here’s where to build, where to invest, and where the next category leaders get founded in the next 18 months.

McKinsey published a report this month on what they call the “agentic organization.” The headline number: more than 80% of enterprises investing in AI aren’t seeing bottom-line impact. Their frame is a five-pillar model for executives running transformation programs.

Read as a transformation report, it’s fine. Read as a market signal, it’s something else: a map of where the money is about to move.

Eighty percent of enterprises are stuck. They have ChatGPT licenses and a few pilots and no idea how to get from there to the organization McKinsey describes. That gap is the opportunity. Every problem McKinsey names is a startup waiting to exist or an investment waiting to be made.

Here’s the map.

Six layers in the agentic stack:

• what’s already won,

• what’s still open, and

• where the next $10B company gets founded in the next 18 months.

• At the end, five questions to separate real agentic companies from wrappers.

Let’s get to work

👋 Just a quick note before you continue!

Building great AI systems and agents is quickly becoming one of the most important advantages a company can have. So we’ve been building non-stop and this will become a much bigger part of AIMF and PMF going forward.

From now on, as a PMF + AI subscriber, you’ll get access not just to the best resources on fundraising and growth (160+ Playbooks, 65+ Success Stories, 30+ Pitch Deck Collections, 25+ Investor Resources), but also:

• 🤖 3+ New AI agent Templates Every Week

• 📚 A Growing Library of 20+ AI Agents

Here, few examples:

• The McKinsey Slide Playbook for Claude

• Claude x Clay Playbook

• The Ultimate Guide to Building Skills for Claude

• Claude For Finance: The CompIete Guide

• The Claude Playbook for Reddit Prospecting

• Claude for AI SDR: The Complete Guide

• Linkedin X Claude Agent Playbook

• The Claude Code Founder Guide

We’re already 70,000+ founders and operators inside.

Get 7 day free trial

Read more

The quote went viral. The takes wrote themselves. The ladder is disappearing. The bottom rungs are dissolving. Run.

I want to offer a different read.

Subscribe now

As a AI Opportunity + PMF subscriber you’ll also get access to all the other Claude resources we put together, such as:

Access now

• 1️⃣ The Claude Code Founder Guide

• 2️⃣ want to keep your job? read this.

• 3️⃣ claude just made outbound fully automatable.

• 4️⃣ 🚨 The Ultimate AI Fundraising Copilot (Built With Claude)

• 5️⃣ look what anthropic just released

• 6️⃣ The Ultimate Guide to Building Skills for Claude 🔥

• 7️⃣ The Prompting Skills Guide

• 8️⃣ The Garry Tan (YC) Playbook for Building with Claude

• 9️⃣ Claude. md Full guide

• 🔟 Understanding Claude Dispatch (Complete Guide)

• 1️⃣1️⃣ Claude For Finance: The CompIete Guide

• 1️⃣2️⃣ How I Rebuilt My SEO Strategy Inside Claude (The Playbook)

• 1️⃣3️⃣ The McKinsey Slide Playbook for Claude

We’re already 70,000+ founders and operators inside!

and just for this week you can get AI + PMF at 30% Discount!

Upgrade now to AI + PMF with 30% OFF

What’s actually happening

Amodei is right that the work junior analysts, associates, and consultants do today.

Drafting from templates, running comps, synthesizing 200 pages into 20 bullet points is exactly what language models are built to do. He’s not wrong about the capability curve. He’s not wrong that enterprise buyers are deploying Claude to do this work right now.

But “the work is being automated” and “the role is disappearing” are two different claims.

The work is being automated. The role is being rewritten.

The new job description

Here’s the version of the story that nobody went viral with: a junior lawyer who knows how to direct Claude through discovery, draft memos, cross-check citations, and pressure-test arguments doesn’t do less work than the 2019 version of that role. They do ten times more. They do it at a quality that used to require three years of experience. And they do it while the partner is still on the first call of the morning.

That person isn’t being replaced. They’re being promoted on day one, whether their title reflects it or not.

The juniors who disappear are the ones who treat AI like a threat to hide from, or a novelty to dabble with. The juniors who thrive are the ones who treat it like the most important tool they’ll ever learn to use — more important than Excel was for finance in the 90s, more important than Bloomberg terminals, more important than legal research databases.

What the bar looks like now

If you’re early in your career in one of these fields, the honest expectation has shifted. It’s not enough to be smart, willing, and good at following instructions. That was the old deal. The new deal is closer to:

Know your craft well enough to spot when the model is wrong. Know your tools well enough to get ten times the output without ten times the errors. Know your clients and colleagues well enough that your judgment — not your typing speed — is what people are paying for.

This is a higher bar than the one that existed five years ago. It’s also a more interesting job. The parts of junior work that were genuinely soul-crushing — the document review marathons, the deck-formatting death marches, the comps-building drudgery — those parts are going away. What’s left is the thinking.

What seniors should actually be doing

If you manage or mentor junior people, the Amodei quote isn’t a reason to stop hiring them. It’s a reason to hire differently and train harder.

The junior who can run an AI-augmented workflow at partner-review quality is worth more than three juniors who can’t. That’s a real shift in how teams should be structured, but it doesn’t mean teams shrink to zero. It means the ratio changes, and the training gets more deliberate.

The firms that quietly figure this out in the next eighteen months will pull away from the ones still arguing about whether AI belongs in the workflow at all.

The part worth sitting with

Amodei’s prediction reads as a warning because it sounds like mass unemployment. Read it again with fresh eyes and it’s something else: a forecast that the floor of what counts as competent professional work is about to rise dramatically.

That’s not a career-ending event. It’s a career-defining one, for people who take it seriously now rather than in 2029.

Five years is a long time. Long enough to get very good at something. Long enough to become the person on your team who everyone else quietly asks for help. Long enough to build the kind of leverage Amodei is describing — not to be displaced by it.

The question isn’t whether your job survives. It’s whether you do the work now to become the version of yourself that the new job requires.

Starting, as always, this week.

Cheers

Guillermo

OH PS: My friends from Newcomer just announced their annual AI Summit in London: Cerebral Valley. I went last year and it was amazing. Check it out!

So I thought it could be very valuable to break it down for you.

💭 Remember!

As a AI Opportunity + PMF subscriber you’ll also get access to all the other Claude resources we put together, such as:

Access now

• 1️⃣ The Claude Code Founder Guide

• 2️⃣ want to keep your job? read this.

• 3️⃣ claude just made outbound fully automatable.

• 4️⃣ 🚨 The Ultimate AI Fundraising Copilot (Built With Claude)

• 5️⃣ look what anthropic just released

• 6️⃣ The Ultimate Guide to Building Skills for Claude 🔥

• 7️⃣ The Prompting Skills Guide

• 8️⃣ The Garry Tan (YC) Playbook for Building with Claude

• 9️⃣ Claude. md Full guide

• 🔟 Understanding Claude Dispatch (Complete Guide)

• 1️⃣1️⃣ Claude For Finance: The CompIete Guide

• 1️⃣2️⃣ How I Rebuilt My SEO Strategy Inside Claude (The Playbook)

• 1️⃣3️⃣ The McKinsey Slide Playbook for Claude

We’re already 70,000+ founders and operators inside!

and just for this week you can get AI + PMF at 30% Discount!

Upgrade now to AI + PMF with 30% OFF

Inside you will find:

1. Why Jensen thinks Nvidia’s real moat is much bigger than GPUs

2. Why he believes AI will increase software usage, not destroy it

3. Why supply chain coordination is one of Nvidia’s biggest strategic advantages

4. Why he still thinks GPUs beat TPUs and custom ASICs

5. Why he sees China as an ecosystem battle, not just an export control debate

1. Nvidia’s real business: turning electrons into tokens

The best line in the interview was also the simplest:

The input is electrons. The output is tokens. In the middle is Nvidia.

That is probably the cleanest way to understand the company.

Nvidia is not just trying to build the best chip.

It is trying to build the best system for transforming energy into useful AI output.

That includes:

• silicon

• packaging

• networking

• interconnects

• software

• libraries

• optimization

• algorithms

• and the ecosystem around all of it

This is why Jensen pushes back so hard on the idea that Nvidia could get “commoditized” just because parts of the stack are outsourced or because software is becoming easier to generate.

His point is simple:

the hard thing is not writing code

the hard thing is making the whole stack work together at extreme scale.

That is where Nvidia lives.

And that is why he does not think the company is becoming a commodity.

2. The moat is not the chip. The moat is the system.

A lot of people still talk about Nvidia as if its advantage is mostly having the fastest GPU.

That is too narrow.

Jensen’s argument is that Nvidia’s moat has four layers.

A. CUDA

CUDA is still the center of gravity.

Not because customers cannot write their own kernels.

They can.

But because CUDA gives developers the richest base layer to build on:

• the deepest tooling

• the most mature libraries

• the broadest compatibility

• the largest install base

That matters more than raw benchmark slides.

If you are building an AI company, you want your software to run everywhere. Nvidia gives you that.

B. Install base

Jensen repeatedly comes back to this.

Nvidia is everywhere:

• all major clouds

• enterprise data centers

• startups

• research labs

• robotics

• edge systems

That means software built on Nvidia has immediate distribution.

That is an underrated advantage.

C. Full-stack co-design

Nvidia is not just making chips in isolation.

It can change:

• the processor

• the fabric

• the network

• the system design

• the libraries

• the kernel layer

• and sometimes even the algorithmic implementation

That is a different game from building a narrow accelerator.

D. Supply-chain orchestration

This may be the least appreciated part of Nvidia’s moat.

Jensen makes it clear that Nvidia’s strength is not just invention.

It is coordination.

The company can align upstream suppliers, downstream buyers, packaging partners, memory providers, foundries, OEMs, clouds, and application builders around a future that it sees earlier than most of the market.

That is not a normal chip-company capability.

That is industrial power.

So the real takeaway is this:

Nvidia’s moat is not one thing. It is compute architecture + software + distribution + supply chain + ecosystem.

That is why it is so hard to dislodge.

3. Jensen’s contrarian view: AI may make software companies bigger

This was one of the most important parts of the conversation.

The current market narrative says AI will commoditize software.

Jensen’s view is almost the opposite.

He thinks AI will cause tool usage to explode.

Why?

Because the number of software users is about to stop being limited by the number of humans.

In the old world:

• one engineer used one set of tools

• one analyst used one stack

• one designer used one interface

In the new world:

• each person may have multiple agents

• each agent may use multiple tools

• each workflow may generate far more tool interactions than before

That means the number of software “users” could increase massively, even if many of those users are agents rather than people.

He gave examples like:

• EDA tools

• design compilers

• floor planners

• engineering systems

His point is that companies are not going to use fewer tools.

They may end up using far more of them, because agents will become tool users too.

This is a very important idea.

If Jensen is right, AI does not simply compress software spend.

It changes the unit of demand.

From human seat count to human + agent output.

That could be one of the biggest shifts in software over the next few years.

4. Why Nvidia keeps scaling while others get stuck

Another core message from the interview:

Nvidia’s advantage is not just that demand is high.

It is that the company has learned how to prepare the system around that demand.

Jensen describes a world where Nvidia is constantly:

• forecasting demand years ahead

• identifying future bottlenecks

• informing suppliers before the bottleneck hits

• getting them to invest early

• helping shape the ecosystem so capacity is there when needed

That is how you get from “everyone is bottlenecked” to “we are still shipping at scale.”

He mentions this directly across:

• CoWoS

• HBM

• photonics

• logic capacity

• packaging

• testing workflows

• and even basic labor constraints like electricians and plumbers

His broader point is:

most bottlenecks are not permanent

they are usually 2 to 3 year problems if the demand signal is strong enough and someone coordinates the response.

That is a very bullish view on compute scaling.

But he does highlight one bottleneck he takes more seriously than chips:

5. Energy is the real bottleneck

Jensen sounds relatively calm about semis bottlenecks.

He does not sound calm about energy.

That came through clearly.

You cannot build:

• AI factories

• reindustrialization

• robotics

• EV production

• advanced manufacturing

• new data centers

without energy.

His message is that America can solve packaging, memory, and foundry bottlenecks faster than it can solve energy bottlenecks.

That matters because it shifts the real strategic question.

The future of AI infrastructure may depend less on who can tape out the next chip, and more on who can secure the power to run it.

That is a much bigger systems question.

6. Why he still thinks GPUs beat TPUs and ASICs

This was the most technical part of the interview, but the core argument was clear.

Jensen does not deny that TPUs and ASICs can be strong for specific workloads.

What he denies is that the future of AI is stable enough for narrow specialization to dominate.

His reasoning:

AI is not just one repeated matrix multiply forever.

The field keeps changing through:

• new attention mechanisms

• new architectures

• new routing methods

• new parallelization patterns

• new model designs

• new inference patterns

• new system-level tradeoffs

In that world, programmability wins.

That is the heart of the Nvidia case.

Jensen’s claim is that most of the big leaps in AI performance do not come from pure semiconductor scaling.

They come from co-design across the stack.

That is how Nvidia got the kind of gains he references from Hopper to Blackwell.

Not from Moore’s Law alone.

But from:

• numerics

• system architecture

• software

• model design

• distribution of workloads

• network and fabric improvements

• and deep optimization at every level

That is why he keeps returning to CUDA.

Not because CUDA is some magical piece of software.

But because it is the programmable layer that lets Nvidia adapt as AI changes.

That is the strategic point.

ASICs can be good at a frozen problem. Nvidia is betting AI will not stay frozen long enough.

7. Nvidia’s strategy is to do only what nobody else will do

Jensen repeats one idea again and again:

do as much as needed, as little as possible

This is one of the cleanest explanations of Nvidia’s strategy I’ve heard.

It explains why Nvidia:

• builds core platform layers

• supports neoclouds instead of becoming one

• invests in ecosystem players

• helps labs scale

• but avoids absorbing every adjacent business

His test seems to be:

If Nvidia does not do this, will it happen anyway?

If the answer is no, Nvidia steps in.

If the answer is yes, Nvidia prefers to partner.

That is why it built CUDA and NVLink.

That is why it helps enable CoreWeave, Crusoe, and others.

That is why it invests in model companies.

But it is also why it does not want to become the entire stack.

This is not empire building for its own sake.

It is selective control around the hardest irreplaceable layers.

That discipline is part of why Nvidia has stayed so coherent while expanding so aggressively.

8. China: Jensen’s real fear is ecosystem loss

The China section was the most heated, but Jensen’s underlying point was consistent.

He thinks many people frame the issue too narrowly.

The standard debate is:

Should America sell advanced compute into China or not?

Jensen’s framing is different:

What happens if America forces the world’s second largest market to build around a non-American stack?

That is the part he keeps coming back to.

His concern is not only short-term revenue.

It is long-term ecosystem drift.

He believes that if Nvidia is pushed out:

• Chinese developers optimize elsewhere

• open-source models may optimize elsewhere

• the global software ecosystem may tilt elsewhere

• and America weakens one of the most important layers of its own AI stack

You can disagree with him on policy.

But the strategic logic is important.

He is not defending China sales only as transactions.

He is defending them as a way of keeping the global AI ecosystem anchored to the American compute stack.

That is a much bigger argument than “we want to sell more chips.”

It is really an argument about platform control.

9. The hidden worldview behind the interview

Step back from the specific topics and you can see Jensen’s broader worldview.

He believes:

• AI is an industrial revolution, not just a software cycle

• the winners will be the ones that build systems, not features

• software usage will rise because agents will become users

• supply chains are strategic weapons

• energy matters more than most people realize

• programmability matters more than narrow optimization

ecosystems matter more than products

• and America should lead by scaling, not by retreating

This is why Nvidia feels less like a chip company every quarter.

And more like the operating layer for the AI economy.

Hope this was valuable!

Cheers,

Guillermo

Share

hey everybody,

I’ve been diving really deep into AI for the last 4 months, to the point I’m now dedicating 70-80% of my time to coding.

AI is just blurring the lines between what a developer and a non …

Read more

Last months have again been really fast, and it doesn’t look like AI is losing speed.

I’m now dedicating 80% of my time to building AI Agents. And I have a 100% conviction they are going to be the future of work.

Now, my thesis is the following: every company will hire agents to work for them, but at some points agents will be like humans, there’ll be good agents at some tasks and not so good ones.

The unlock for the following years is building the good ones. And those are going to come from forward deployed ai consultant/engineers developing agents that behave exactly how comopanies need them to.

The value: the proprietary know how and strategy companies have.

Inside this piece:

• Why the biggest AI opportunity is now an execution gap, not a model gap

• Why “forward-deployed” may become one of the defining delivery models of the agent era

• Why OpenAI and Anthropic are both pulling consulting and implementation into the center of the market

• What a real agent deployment playbook looks like in practice

• Why the biggest AI opportunity may sit inside the “99%” of companies outside Silicon Valley

💭 Remember!

As a AI Opportunity + PMF subscriber you’ll also get access to all the other Claude resources we put together, such as:

• 1️⃣ The Claude Code Founder Guide

• 2️⃣ want to keep your job? read this.

• 3️⃣ claude just made outbound fully automatable.

• 4️⃣ 🚨 The Ultimate AI Fundraising Copilot (Built With Claude)

• 5️⃣ look what anthropic just released

• 6️⃣ The Ultimate Guide to Building Skills for Claude 🔥

• 7️⃣ The Prompting Skills Guide

• 8️⃣ The Garry Tan (YC) Playbook for Building with Claude

• 9️⃣ Claude. md Full guide

• 🔟 Understanding Claude Dispatch (Complete Guide)

• 1️⃣1️⃣ Claude For Finance: The CompIete Guide

• 1️⃣2️⃣ How I Rebuilt My SEO Strategy Inside Claude (The Playbook)

• 1️⃣3️⃣ The McKinsey Slide Playbook for Claude

We’re already 70,000+ founders and operators inside!

and just for this week you can get AI + PMF at 30% Discount!

Upgrade now to AI + PMF with 30% OFF

1. The biggest AI bottleneck is no longer intelligence. It is implementation

Read more

Andrej Karpathy says he barely writes code anymore.

Since December, he says his workflow flipped from mostly coding himself to mostly delegating to agents.

That is a much bigger shift than most people realize.

The important part is not that AI helps engineers write code faster.
It is that the interface to building software is changing.

The job is moving from:

• writing functions

• debugging line by line

• manually pushing work forward

to something much stranger:

• assigning work to multiple agents

• managing parallel execution

• designing better instructions

• building systems that keep going without you

Karpathy describes it almost like a new kind of operating mode. Not coding in the old sense, but “expressing your will” to a set of agents for 16 hours a day.

That sounds extreme.

But he may be early, not wrong.

Because underneath all the jokes about “AI psychosis” is a serious idea:

The next generation of software work will not be about doing the work yourself. It will be about setting up loops of intelligence that can work on your behalf.

And that has implications far beyond software.

I broke down his latest interview for you, inside you will find:

1. What changed

2. Why the real skill is leverage

3. From copilots to autonomous loops

4. The first big opportunity: rebuilding software for agents

5. The second big opportunity: auto-research

6. Why this still does not fully work

7. The third big opportunity: the collapse of traditional UX

8. Digital will move first. Physical comes later.

9. Education is about to be rebuilt too

10. What this means for founders

11. My view

Deel just launched The Pitch, a global startup tournament, for the best founder in the world. The prices:

• 10 startups will each receive a $1M SAFE
• 100 regional winners will each receive $50K SAFEs
• $15M+ total capital deployed

And we’re officially partnering with them to push this to the Market Fit audience 🔥

This isn’t a gimmick. It’s a real funding vehicle.

Here’s how it works:
1️⃣ Submit a 5-minute application
2️⃣ Top startups get invited to regional in-person finals
3️⃣ Regional winners get funded
4️⃣ Global finalists compete for $1M

Apply Here

What changed

According to Karpathy, something broke open around December…

From here on, this is reserved for premium subscribers.

If you want to really understand where AI is going and how to stay ahead of it, you can subscribe here:

Subscribe now

Read more

Y Combinator just published its requests for Spring 2026 and, among all the things you can find in there, there’s one idea that’s very easy to skim past if you’re not paying attention: AI-native agencies.

At first glance it sounds familiar: every agency already claims to use AI, everything is “AI-powered,” so it’s tempting to dismiss this as just another variation of the same trend.

But that’s not really what YC is pointing at.

If you read it more carefully, what’s emerging is a deeper shift:

the next wave of meaningful AI companies might not start by selling software, but by actually doing the work.

That connects directly to what we saw in the Sequoia piece I wrote (read it here) and that YC is confirming: the next trillion company might be an services company, not software.

So I decided to follow my research into this and see why:

1. YC’s AI Native Agencies Thesis

2. Why to start an AI Native Agency

3. The Difference Between AI in Agencies vs AI-Native Agencies

4. Which Agency Categories Will Be Disrupted First by AI (And Why These Ones Win)

5. Top 15 AI Startups Moving Toward the Agency Model

Apply to The Pitch by Deel now

1. YC’s AI Native Agencies Thesis

For the past decade, the dominant playbook was clear and almost unquestioned: build software and sell it to agencies or service providers.

Agencies were seen as:

• inefficient

• messy

• hard to scale

YC is now suggesting something that moves in the opposite direction.

What if, instead of building tools for agencies, you build the agency itself but redesigned from the ground up?

At the same time, the real shift is in the economics.

Historically, agencies scaled linearly:

more clients → more people → thinner margins → more complexity → inconsistent quality.

With AI, that relationship breaks. You can:

• increase output per person

• reduce the marginal cost of delivery

• standardize parts of quality that used to be unpredictable

But the most important part is this: you can still sell the same outcomes at high prices while your internal cost structure changes dramatically.

That’s where the real advantage comes from:

not doing things better, but doing them on a completely different economic base.

From here on, this is for paid subscribers only.

If you want to keep reading, you can subscribe here:

Subscribe now

2. Why to start an AI Native Agency

Read more

Most people glanced at it, recognized a few names, maybe saved a couple to check later and moved on.

That’s a mistake.

Because that slide is not just a list. It’s a blueprint of where value in AI is being built today and where it will be created next.

And if you don’t know how to read it, you’re just looking at logos.

If you do know how to read it, you start seeing:

• where capital is concentrating

• where new players can still emerge

• and where the next category-defining companies will come from

What you’re about to get (and why it matters)

We didn’t just go through the list.

We rebuilt it from scratch.

We ignored Nvidia’s categories. And we asked a more useful question:

What role does each of these companies play in the AI economy?

That one shift changes everything. Because instead of a flat list, you get a map of the system. And once you see the system, you start seeing:

• where value accumulates

• where bottlenecks form

• and where opportunities actually are

What you will get:

1. What AI-Native Companies Really Are and How to Spot the Real Ones

2. Why Nvidia’s map is misleading (if you read it wrong)

3. How to Reclassify the 103 AI Startups to See Where Value Actually Sits

4. The Definitive Reclassification of All 103 AI-Native Startups

5. The insight Nvidia didn’t spell out

6. The two areas you cannot afford to ignore

Most people stop at the list.
The real insight starts here.

From this point on, this is for subscribers only.

If you want to understand where value is actually being created and how to read signals like this properly you can subscribe here:

Subscribe now

Read more

A few days ago, Sequoia partner Julien Bek made a very big claim:

The next $1 trillion company will be a software company masquerading as a services firm.

At first glance, that sounds like one of those bold VC one-liners people repost because it sounds smart.

But the more I looked into it, the more I thought:

wait, this might actually be one of the most important ideas in AI right now.

Because if this is true, then a huge number of founders are building the wrong thing.

They are building copilots.
They are building tools.
They are building software that helps someone do the job better.

But the real opportunity may be much bigger:

build the company that does the job itself.

Not software for accountants.
The company that closes the books.

Not software for lawyers.
The company that reviews the contracts.

Not software for insurance teams.
The company that handles the claims.

That is the shift Julien is pointing to.

In his piece, he argues that if you sell the tool, you are stuck competing with better models every few months. But if you sell the work, every improvement in AI makes your product faster, cheaper, and stronger.

And there is another reason this idea matters so much:

for every $1 spent on software, around $6 is spent on services.

So the real prize may not be the software budget at all.

It may be the labor budget.

What makes this even more interesting is that this is not a totally new idea.

For years, investors have tried to find ways to turn service businesses into something more scalable and more software-like.

That is part of why strategies like rollups have become so attractive: the belief that fragmented service industries can be consolidated, improved, and made far more efficient.

What AI changes is the magnitude of that opportunity.

Before, software could help professionals do the job better.

Now, it is getting close to doing meaningful parts of the job itself.

That is the real reason Julien’s thesis feels so important right now.

Not because nobody has seen the opportunity before.

But because this may be the first moment when the technology is finally good enough to make it work at scale.

So I decided to go deeper.

I looked into the companies Julien and Sequoia point to as early examples of this shift, what they are actually doing, how they are doing it, and why some of these categories may produce massive companies over the next few years.

In this deep dive, you’ll learn:

1. Who is already doing this: 5 founders building right now (mentioned by Sequoia)

2. How these companies are structuring the model so customers trust them

3. Where the biggest opportunities still are right now

This is the newsletter that surfaces what’s really happening in the AI market before it hits the front page. Every week you’ll get two things, fast:

1. What’s happening now (and where the money’s headed).

2. Who’s already building it (so you can invest, partner, or steal their playbook).

Each issue is short enough to read with your morning coffee but dense enough to move your roadmap.

However, most of the content from this newsletter will be only accessible to the paid subscribers.

If you are serious about building or investing in AI, you can subscribe here:

Subscribe now

Read more

Some people get lucky once.

Leopold Aschenbrenner looks like someone who engineered luck.

In a very short window, he went from being a young OpenAI insider with a reputation for sharp thinking to publishing a 165-page essay laying out a decade-long AI thesis (“Situational Awareness: The Decade Ahead”), and then building an investment vehicle that expressed that thesis in markets.

The numbers are hard to ignore: by the end of 2025, his fund’s public 13F long U.S. equity book alone was in the multi-billion range (this is not total AUM, but the SEC-reported public equities snapshot).

So I did what I always do when someone prints that kind of outcome:

I reverse-engineered the playbook.

In this analysis you’ll learn:

1. The one simple idea that explains his whole portfolio (so it stops feeling random and starts feeling obvious).

2. How the bets evolved over time as AI went from “hype” to “massive capex” to “real constraints” (power, data centers, networking).

3. A clean breakdown of the portfolio into buckets, so you understand what each part is doing

4. A framework you can reuse, so you can look at any “AI stock” and quickly tell:

The goal is not to idolize a manager.

It’s to extract a framework: how to turn a macro AI thesis into a coherent, compounding portfolio of bottlenecks and why that approach can outperform when everyone else is chasing shiny apps.

Let’s get to work

Apply to The Pitch by Deel now

This is the newsletter that surfaces what’s really happening in the AI market before it hits the front page.

Read more