The most viral moment at the 2026 All-In Summit wasn’t a product announcement.

It was Jensen Huang’s phone ringing.

Onstage in front of thousands of people, the Nvidia CEO received a call from President Donald Trump and put him on speaker.

Trump dismissed the growing alarm around artificial intelligence as a “hoax,” declared that robots would not take over and argued that slowing American AI development would primarily benefit China.

But that was only one part of a much bigger conversation.

We watched and broke down the summit’s key interviews, extracting the arguments, disagreements, predictions and clips actually worth your time.

Inside, you’ll find:

• Jensen Huang → AI doom, regulation, China, open source, Nvidia’s capital strategy and its expansion beyond chips

• Satya Nadella → Human control, independent evaluators and why companies must own their AI learning loops

• Donald Trump → The AI race with China, data centers and the politics of slowing down

• Elon Musk → Exponentially increasing AI risk, peer review between labs, Starship, Terafab and a possible SpaceX–Tesla combination

• Gwynne Shotwell → Running SpaceX, direct-to-cell Starlink, reusable rockets and data centers in orbit

Plus, direct links to the most important moments and what these conversations reveal about the growing split inside Silicon Valley.

Hey, welcome to AI Market Fit!

Every week you’ll get two things, fast:

1. What’s happening now and where the money is moving

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

Most of the content from this newsletter is available only to paid subscribers.

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

Subscribe now

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

Let’s begin!

Read more

An Anthropic researcher recently resigned after warning that people inside leading AI labs assign a meaningful probability to AI causing catastrophic harm this decade.

His colleagues put a greater-than-10% chance on AI killing everyone this decade, and within days Sam Altman went on camera to call that exact number "unacceptable" while refusing to name any number of his own.

At the same time, Dario Amodei has been pushing for third-party evaluators, deployment controls, and stronger oversight of frontier models.

On the surface, this is a debate about AI safety.

But there is another consequence receiving far less attention: the safety standards being proposed by the largest AI companies could become a regulatory moat around the frontier.

→ The more expensive it becomes to train, evaluate, audit, and deploy frontier models, the harder it becomes for new competitors to enter.

And one requirement in particular looks increasingly likely to become part of that system: mandatory independent model evaluations.

In this issue:

1. The Resignation That Changed the Debate

2. Why Altman Won’t Name a Number

3. The Emerging AI Compliance Stack

4. Safety Meets Business Incentives

5. The Regulation to Watch

6. The Incumbent Moat

Hey, welcome to AI Market Fit!

Every week you’ll get two things, fast:

1. What’s happening now and where the money is moving

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

Most of the content from this newsletter is available only to paid subscribers.

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

Subscribe now

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

Let’s begin!

Read more

Our verdict up front: do not build another raw TTS/STT model to fight ElevenLabs, Cartesia and OpenAI on quality.

Build in the layers where deployment, evaluation and vertical integration are still unsolved. Here is the founder’s own reasoning, and the investable map that falls out of it.

In this issue you’ll find:

1. Why launching before ChatGPT in the crypto-metaverse era was the unlock

2. The Polish dubbing problem that became an $11B company

3. Why they pivoted away from dubbing

4. The Honda rule ElevenLabs uses to decide what NOT to build

5. What is the real Moat in ElevenLabs (and AI)

6. How ElevenLabs grows

7. Where ElevenLab’s revenue comes from

8. How to create AI voices that stick

9. Four ways to build or invest in the voice AI stack right now

Hey, welcome to AI Market Fit!

Every week you’ll get two things, fast:

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

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

This week a 27-year-old researcher who spent three years doing pretraining at OpenAI and then Anthropic quit his job, posted a thread that passed 100 million views, and put the odds of AI killing every human on earth at more than 10 percent. Then an Anthropic alignment lead replied, on the record, that he agreed.

I read the whole thing twice and sat with it. I want to walk you through what it actually says, why it landed so hard, and where I honestly come down: not on panic, and not on hype, but on faith.

In this issue you’ll find:

1. The resignation that put 10 percent odds on human extinction

2. Why this warning is credible even to those of us on the outside

3. We survived the last weapon that could have erased us

4. If progress can’t be stopped, root for the AI that cures cancer

5. What choosing faith actually changes for you

1. The resignation that put 10 percent odds on human extinction

Jacob Coxon resigned from Anthropic on Tuesday, and he did not go quietly. He is 27, and he spent the last three years doing pretraining research at the two most important labs in the world: a member of OpenAI’s technical staff from 2023 until July 2026, where he worked on GPT-4o, then a researcher at Anthropic. This is not a fringe voice. This is someone who was in the room.

The line that traveled everywhere: “Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.” You can read his full thread here: x.com/hilbertspaess/status/2097476196791709843.

What made it more than a resignation letter was the second claim: “The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.” He pre-empted the obvious dismissal that scary talk is just good marketing, noting that many executives and senior researchers soften their phrasing in the press to sound sensible, but that he hears the same people express fear privately.

By September 9, the thread had passed 100 million views and been picked up by Fortune, Business Insider, Ars Technica, TechCrunch, the Washington Post, and CNBC, which framed the headline number bluntly: more than a 10 percent chance of “killing all humans”.

Strip away the virality, and here is the uncomfortable core. A person whose entire job was building the thing looked at the trajectory and decided the honest move was to walk out and say it out loud. That is worth taking seriously before we decide what to do with it.

2. Why this warning is credible even to those of us on the outside

The strongest reason not to dismiss Coxon is that someone still inside the building agreed with him publicly. Anthropic’s own Alignment Science lead, Evan Hubinger, replied on X: “Jacob is correct here, we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade”. That reframes everything. This is not a bitter ex-employee versus the company. It is a view held inside the company, by the person whose title is literally about keeping the models aligned.

Coxon’s sharpest observation is the difference he draws between the two labs. “At OpenAI, many have not deeply internalized the civilizational stakes,” he wrote. “At Anthropic, the stakes are well understood, but they are locked in a race to get there first, they believe no one else will act responsibly, so they must do it themselves, despite the risk.” That is the trap in one sentence: the people who understand the danger best feel they have to run toward it anyway, because they trust themselves more than the alternative.

He is also not the first to walk. Ars Technica noted that a prominent researcher quitting a frontier AI lab has become a recognizable event, one in a pattern of departure warnings rather than a one-off. When multiple credible insiders keep leaving with the same message, the base rate matters. This is real signal.

But signal is not proof. Coxon is describing a future (”these will soon be superhuman systems that can hack anything”) that has not happened yet. Even Anthropic’s own August alignment report, the one Hubinger pointed to, rates the catastrophic risk from current models as “low,” while warning that future, more capable models “may cause unbounded harm.” So the honest read is this: the people closest to the technology are genuinely scared, they may well be right, and none of us on the outside can verify it either way. All three are true at once.

3. We survived the last weapon that could have erased us

Here is where I have to be honest about my own limits. I am not in the lab, and neither are you. From out here, we cannot run the experiment that tells us whether Coxon’s 10 percent is closer to the truth than someone else’s 1 percent or 30 percent. The same uncertainty that makes his warning credible is what makes it unverifiable. We are not being handed a calculation. We are being handed a choice between fear and faith.

And I keep coming back to one fact that steadies me. Humanity has held the power to erase itself since the 1940s. Nuclear weapons could have ended the whole story many times over. Through the Cold War, through Cuba, through decades of terrifyingly close calls, we did not use them. Not because the danger was fake, it was as real as danger gets, but because enough people, on enough sides, chose not to. World-ending power sat in human hands for eighty years, and we are still here.

That is not naive optimism. It is precedent. It is the single best piece of evidence we have about how our species behaves when it can destroy itself, and the evidence says we can hold that power without pulling the trigger. Coxon points at something similar when he talks about coordination: “warning shots like the Hugging Face attack have made pacing agreements between the US labs more viable.” The people worried about this are also actively trying to build the guardrails, the same way the nuclear age eventually produced treaties.

So I refuse to resolve the uncertainty into doom just because doom is the loudest emotion in the room. Holding “he might be right and we can’t know” is uncomfortable, but it is the mature place to stand. Panic is a decision, not a fact. And I would rather spend my energy on the version of the future I actually want to see.

4. If progress can’t be stopped, root for the AI that cures cancer

I do not think we can stop this. That is the plain truth. Coxon himself, who is more worried than almost anyone, wrote that he doesn’t feel we’re “on track to prevent a global race.” When even the person walking out in protest can’t see how the brakes get pulled, betting your peace of mind on a full stop is a losing bet. So if progress is coming either way, the sane move is to lean hard into the good it can do.

And the good is already showing up. Last month Anthropic’s Claude completed the first formalized proof of a major mathematical result, converting the reasoning into a form the Lean proof assistant could verify, work that used to take human mathematicians years (Anthropic on X, 2026-09-04). The same category of self-improving system that Coxon fears is the one now accelerating real science.

Formalizing math is not a party trick. It is the same muscle that reads through millions of molecular interactions to find a drug candidate, that models how a protein folds, that spots the pattern a human researcher would need a decade to notice. If a system can hold an entire mathematical proof in its head and check every step, pointing it at cancer biology or the tangle of Alzheimer’s is a difference of domain, not of kind.

So here is the near-term version I am rooting for, and it is not science fiction. In the next couple of years, I fully expect AI to compress drug-discovery timelines, to hand oncologists better-targeted treatments, and to help crack diseases that have humbled us for generations. That is the exact same technology in the exact same headlines. Coxon is scared of what it could destroy. I choose to also see what it could heal. Both live inside the same machine, and which one we build more of is not fully out of our hands.

5. What choosing faith actually changes for you

If you build with AI, invest in it, or bet your career on it, this week’s warning is not a reason to freeze. It is a reason to be deliberate about which future you pour yourself into. AI is going to end up being an image of humanity, and humanity is not one thing. It is capable of the bomb and capable of the vaccine. What it becomes depends heavily on what the people building it are actually trying to make.

That means the choice in front of you is more consequential than it feels. Every founder who points these systems at disease instead of manipulation, every investor who backs the lab treating safety as a first-order problem rather than a press line, every operator who insists on understanding what they’re shipping, is casting a vote for which image AI reflects back. Coxon’s whole closing appeal to lab researchers was exactly this: don’t put your head down because it’s happening anyway, use the moment to “call for different conditions.” That applies to all of us, not just the people inside.

I am choosing to trust the good in people. Not because I’ve proven the pessimists wrong, I can’t, and neither can they prove themselves right. I’m choosing it because the last time we held world-ending power we chose not to end the world, and because the same technology that scares a credible insider is the one plausibly about to cure the things that have been killing the people I love.

So hold the warning honestly. Take it seriously. Then decide which side of it you’re going to spend your one working life building. The future is not something that happens to us. It is the thing we root for, and then go make. I’m rooting for the one that cures cancer. Let’s build that one.

Best,

Guillermo

Roughly one in eight American adults reported taking a GLP-1 like Ozempic in 2025, according to a Kaiser Family Foundation poll cited by CNBC. The cheap-drug boom underneath that number runs on a legal loophole that is now closing: compounded semaglutide is only permitted while the FDA lists an official shortage, and those shortages are resolving. Every telehealth reseller arbitraging that gap is standing on a cliff.

So we broke down the whole GLP-1 and peptide market: what the drugs actually are, who is making billions, exactly where the regulation bites, and the vertical-integration playbook that survives the cliff.

The anchor is RonanRX, the YC Summer 2026 company (with $7M+ committed to its seed) that we just invested in, because its structure is the clearest live example of the durable bet.

In this issue you’ll find:

1. What are GLP-1s

2. Can you just sell GLP-1s?

3. Why We Invested In RonanRX

4. Where The Money Already Is

5. The Next 18 Months

6. What This Changes

Hey, welcome to AI Market Fit!

Every week you’ll get two things, fast:

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

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

Sequoia just told a room of 80 portfolio founders that owning your AI models down to the weights is no longer a performance sacrifice, it may be your performance edge.

Sonya Huang’s “Own Your Intelligence” talk laid out exactly which parts of your stack to own versus rent and the four-step roadmap to build your own lab. We went through all 17 minutes, screenshotted every slide, and broke down what Sequoia argues on each one, so you don’t have to watch it yourself.

This matters now because of one claim Huang makes that would have been false a year ago: open-weight models like Qwen and GLM are close enough to the frontier that you can start near it and tune past it. That flips the whole calculus. Owning your intelligence used to mean accepting worse performance. Sequoia argues it no longer does.

In this guide you’ll find:

1. Sovereign AI defined: owning your weights without ripping out Opus or GPT

2. Centralized vs decentralized intelligence: why owning your models is the optimistic bet

3. The four reasons companies own their models: cost, speed, performance, destiny

4. “Not your weights, not your product”: why the battleground moved to the intelligence layer

5. Step 1 and 2: the four-factor own-vs-rent test and how to build the team without shoehorning your platform group

6. Step 3: legibility is the underrated moat because every buyer is choosing their AI champion

7. Step 4: the technical roadmap and why open weights now push you past frontier performance

8. The stack and Pandora’s Box: what owning your intelligence actually costs you

Hey, welcome to AI Market Fit!

Every week you’ll get two things, fast:

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

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

Subscribe now

How Companies Are Building Their Own Intelligence, According to Sequoia

Read more

Then, this spring, they exploded past 600 per month. That is not a blip. It is AI industrializing the discovery of software flaws, and it forces a matching wave of defensive innovation that is now one of the most urgent, investable categories in tech. We pulled the funding rounds, the market numbers, and the exact whitespace so you can decide where to build, invest, or partner before this prices in.

In this issue you’ll find:

1. The 600/month signal: why the vulnerability surge means AI broke cyber for good

2. The symmetry trap: how AI made offense AND defense cheaper, and who wins that race

3. The three defensive plays with room to build (AI-native SOC, securing AI infra, agent security)

4. Where the money already moved: the AI-security rounds and market numbers that prove the thesis

5. The operator playbook: how to position before the category prices in

Hey, welcome to AI Market Fit!

Every week you’ll get two things, fast:

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

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

At 69, after 26 years running Sequoia and sitting on stakes (such as 20% of Nvidia and 10% of Google), Doug Leone came back to work and told his partners he feels “good for nothing.”

Then he gave himself a deadline: if in nine months he still feels irrelevant, he fires himself.

Most coverage of Sequoia’s returns stops at the numbers and skips the engine that produced them, the fear, the founder spec, the hold-forever conviction, the rules he uses to read people. We went through David Senra’s full conversation with Leone and pulled out the heuristics he actually uses to pick founders, hold winners and build boards.

High-signal analysis on markets, capital and the companies shaping the next decade

Inside you will find:

1. Why Fear Beats Vision

2. How to Turn Fear Into a Tailwind

3. Where Real Drive Comes From

4. Why He Never Had a Plan B

5. When the Old King Has to Leave

6. The Three Filters Behind Every Investment

7. Why Every VC Sells Too Early

8. The Founder Spec Leone Wants

9. Heart of a Lion Over Resume

10. Why the Best Leaders Stay Miserable

This issue is available exclusively to paid subscribers!

If you’re not subscribed yet, you can unlock this and our full library by upgrading below:

Get 7 day free trial

Let’s begin!

Read more

Glean just crossed $300M in top-line revenue seven years in, and its founder gave up his real playbook in a five-minute hallway interview over lunch.

Arvind Jain, the 2x unicorn founder behind Rubrik and Glean and a decade-long Google search engineer, broke down who to hire first, when to raise, what investors actually bet on, and how Glean quiewetly became the layer enterprises use to cut AI costs by close to 6x. We went through the entire conversation question by question so you can steal the moves without watching a single minute.

In this guide you’ll find:

1. The origin story: ‘Google for your work life’ (and 20 engineers before anyone else)

2. Arvind’s fundraising rules: raise fast, dilute early, and know investors bet on you not the idea

3. We asked about Uber burning its AI budget: is enterprise AI ROI actually real?

4. The use cases nobody predicted, and who actually buys Glean

5. Openrouter getting acquired by Stripe 💰

6. Will AI erase all jobs as Dario Amodei says?

7. The OpenAI panel: a 3-layer playbook for deploying agents safely

8. The Enterprise AI Playbook by Glean

High-signal analysis on markets, capital and the companies shaping the next decade.

1. The origin story: ‘Google for your work life’ (and 20 engineers before anyone else)

We opened with the classics: how did you end up here, who do you hire first after raising your first round, and how do you communicate the vision to so many engineers?

The wedge was a problem Arvind lived, not one he researched. At Rubrik, the enterprise data security company he co-founded, fast growth quietly wrecked internal productivity. “Everybody would complain about not being able to find things inside the company or connect with the right people to get help,” Arvind told us. He had spent over a decade building search systems at Google, so the fix was obvious to him: a real search product for all of a company’s data and knowledge. That became Glean in early 2019. Before that, he grew up in India and moved to the States in the ‘90s for grad school, and he’s been in the tech industry — and based in San Francisco — ever since.

His answer to “who do you hire first” was one word: builders. “The first 10 people that we hired at Glean, they were all engineers. Maybe even the first 20, I don’t remember hiring anybody outside of R&D,” Arvind said. Office, devices, food, early customers, he handled all of that himself so the product team could stay heads-down until the product hit maturity.

He could hire 20 engineers around one idea because the idea fit in a sentence. “It’s a Google for you in your work life. It’s a place where you go and ask questions when you’re looking for information or answers and we’ll bring the right information back to you,” Arvind said. A pitch this simple removes the friction of explaining what you’re building. But he flags the harder half: people have to believe it will be big. “They’ve got to feel that it’s going to be something big.”

The steal: if you can’t recruit engineers around your idea in one sentence, the idea is too complicated to fund and too complicated to build. Pressure-test your mission on a non-technical friend this week. If they can repeat it back, you can hire around it.

2. Arvind’s fundraising rules: raise fast, dilute early, and know investors bet on you not the idea

Next we asked when a founder should raise at all — bootstrapping versus raising money — and whether raising was easier the first time at Rubrik or the second time with Glean.

Arvind’s capital strategy is deliberately aggressive, and it’s built to protect focus, not valuation. “I like to go big and go and go big fast. So if you can, then I would recommend that, just raise capital. What’s the downside?” he told us. His reasoning is operational: he does not want to spend every month calculating whether the company can afford one more engineer. Raising more removes that recurring distraction.

He directly rejects the most common counterargument. “Sometimes people feel that, oh, is it too early to dilute the company? And I don’t really believe in that,” Arvind said. His view is that the drag of under-capitalization costs more than the dilution.

The deeper insight came from our second question: raising was definitely easier the second time, and the reason is about what investors are actually buying. “Investors are actually never looking to invest in great ideas. They’re actually, they’re looking to invest in people,” Arvind said. His logic: any early idea gets transformed, refined, probably changed into something completely different once you talk to customers, so investors are underwriting your persistence, not your pitch deck. A track record signals you’ll “stay on the mission” and not give up.

He also reads the current market as unusually generous. “There’s a lot of capital that is actually waiting to be deployed” on AI ideas, Arvind said, so founders “should feel really, really, really comfortable” they’ll be able to raise. Note this is his read, not a hard figure.

The steal: frame your raise around the person and the persistence, not the perfect idea, and if you can raise enough to stop thinking about hiring math for 18 months, take it. Run the raise as a distraction-elimination decision, not a valuation-optimization one.

3. We asked about Uber burning its AI budget: is enterprise AI ROI actually real?

We put the reported case of Uber burning its 2026 AI budget in four months without seeing ROI directly to Arvind. He didn’t dodge it. “It’s a wide spectrum, but businesses are definitely getting some value from AI today. If they’re getting none, they would have stopped by now,” he said.

The clearest value, per Arvind, is basic knowledge-seeking that makes every employee “more productive, more knowledgeable,” and “that’s not in contention.” The catch: “it’s hard to measure. The fact that, you know, people can work faster and, you know, oftentimes, you know, that doesn’t translate into like your P&L statement.” The companies seeing real results are larger ones with intentional programs, good metrics, and systematic process-by-process adoption — “it is hard, it takes a lot of time, and AI vendors kind of made it seem like you’ll automatically and instantly get all these value savings.” Reality, he said, is “like any other technology where you have to provide a lot of love and care and investment.”

That’s why Glean sells a long-term engagement, not a switch you flip. “We bring our platform so you immediately get those productivity savings, you get a great coworker, but then the real business value, like, you know, we spend many, many months understanding, you know, their business and helping them build those agents and get them to production level of quality,” Arvind said.

How Glean is using this to win. Arvind’s Glean:GO keynote turned the ROI gap into Glean’s entire product pitch. He named the failure mode most enterprises are living: employees now work for the AI instead of the reverse. “Today employees spend a lot of time coaching AI agents and are feeling overburdened with that. This is bot sitting. We are spending too much time working for AI as opposed to AI working for us,” he said. Botsitting — employees feeding context, supervising, and cleaning up after AI agents — has become a significant drain on worker time: Glean’s Work AI Institute, surveying 6,000 full-time digital workers, found employees spend 6.4 hours a week on it, more time than they spend using AI to actually produce work. In the keynote he laid out the three pains every enterprise names: struggling to bring the right context to agents, AI costs running higher than they can afford, and the proliferation of AI-written documents (SLOP), skills, and agents that are hard to manage and secure. Every one of those pains maps to a Glean feature.

On the OpenAI panel he was blunt about the near-term math: “AI is often becoming more expensive than the business value that it is delivering in the short order... We need to make sure companies don’t run out of money before that happens.” Positioning Glean as the platform that closes that gap — cheaper, better-contexted agents — is the commercial answer to the ROI question we asked.

The steal: stop selling (or buying) instant ROI. Set the expectation that production-quality agents take months of context-building, and instrument the specific business processes where you can actually measure cost-per-task before and after. Pick one high-cost, high-human-involvement process and commit to a multi-month build rather than spraying pilots.

4. The use cases nobody predicted, and who actually buys Glean

We asked four go-to-market questions in a row: have customers used Glean in ways you never imagined, who do you reach out to in the organization, do customers build the agents themselves or do you send in your engineers, and where are your customers?

Because Glean is horizontal, the entry point is the person who owns company-wide AI. “Typically we’ll start with the CIO. You know, the CIO typically thinks about how to bring company-wide AI technologies,” Arvind told us. From there, Glean expands to departmental leaders when the value is concentrated. “Sometimes, you know, we do work with departmental leaders too, because for example, for salespeople, for CROs, we do a lot.”

The footprint is mid-to-large enterprises, weighted to the US. “About 70% of the business is probably US and then 30% is the rest of the world right now,” Arvind said, naming Europe, Asia, Canada, and Latin America as the rest. (Those percentages are Arvind’s stated figures.) Named customers on stage at Glean:GO included General Motors (helping engineers design cars faster), McCarthy (construction management), and Navita (patient support), showing how far the horizontal platform stretches across industries.

The use cases prove the “chief of staff” positioning, and Arvind admits almost none were planned. “Almost all use cases are things that we didn’t imagine people would actually use our product for,” he said. The surprises: a performance-management agent that “writes the first drafts” of annual reviews for individuals and managers, customers “responding to RFPs” with AI, teams “reviewing legal contracts with AI,” and salespeople using AI “to figure out what messages to send to which one of their accounts.” His mental model: “since Glean is so horizontal and it connects to all enterprise systems, it sort of functions as your chief of staff.”

Who builds the agents once Glean is in? The customers do, with Glean enabling them. When we asked whether Glean sends in its own engineers, Arvind said customers build the agents themselves, backed by Glean’s “training and enablement programs,” and “sometimes we actually bring our partners to help build agents for our customers.”

The steal: for a horizontal platform, sell the platform to the CIO and the sharpest single-department win (CRO/sales) in parallel, then let unpredicted use cases pull expansion. Ship analytics that surface which unplanned workflows customers invent, because those are your next feature roadmap and your next case study.

5. Is Glean just a model router? The context moat and the 81% cost cut

Our spiciest question: OpenRouter — reportedly being acquired by Stripe — what do you think of that? It got the most strategically important answer of the whole interview.

Arvind said he’s very familiar with them: “They were one of the first model routing companies, but more sort of model access company than routing.” Glean does some of the same things OpenRouter did, but mostly in the enterprise context for large enterprises, and the business is very different: model access “is actually only one part of our overall solution.” That’s why Arvind resists being called a model router even though Glean does model access. “Largely, like, you know, we think of Glean as the context and intelligence company,” he told us, describing the goal as “bringing the collective human intelligence and the know-how, the knowledge and experience of your enterprise in one place and then make it available to AI systems.”

How Glean is using this to win. The “context company” framing isn’t marketing — it’s the entire thesis, and it rests on one number: structured data is a rounding error of what a company actually knows. “Structured data and data lakes capture only 5 to 10% of all enterprise knowledge,” Arvind said in his Glean:GO keynote, and documents like playbooks and SOPs only go so far. “Much of the real know-how of how work happens lives in people’s heads.” (The 5-10% figure comes from his talk; treat it as his framing.) The good news, per Arvind, is that the other 90-95% “does surface in emails, messages, meeting notes, exception approvals” as people do their regular work.

Capturing that is the moat, and Arvind draws a hard line between Glean’s approach and the connector-first competition. Glean’s context engine “combines advanced retrieval with a powerful enterprise graph” that maps the people, customers, teams, projects, and the relationships between them. “That deeper understanding of your business, how it operates, is what Glean sets itself apart from the industry’s typical approach of delivering shallow context via MCP,” Arvind said, adding: “Glean is the only context and intelligence company that truly understands your enterprise.” Glean CPO Emrecan Dogan sharpened the same point in the company’s Glean:GO press release: “Models are getting better and more interchangeable. Understanding the enterprise is not. We designed Glean around that reality in 2019.”

And context is exactly how Glean wins the cost war. The headline number from Glean:GO: “Glean saves 81% on token costs compared to Claude, with Glean response preferred 78% of the time,” Arvind said. (Percentages are Arvind’s own from the keynote.) Independent observers noticed the token gap too; Rohan Paul flagged the 1.3M vs 4.4M split on X.

As we put it in the interview, “the cost is even like 6 times or close to 6 times less than other options like Claude Cowork.”

The waste Glean attacks is architectural, not technical. “Most systems still point the most expensive models at the simplest of the tasks. That’s not really a limitation of AI, it’s an architecture choice,” Arvind said. Glean’s Pareto frontier analysis makes the spread concrete: across 1,000 enterprise tasks and 37 model/reasoning configurations, the cost spread between the cheapest and most capable options is dramatic for modest quality differences. There is no single best model, so routing is the point. Auto-routing uses a small specialist model to pick the right reasoning effort per task.

The second half of the savings comes back to context. “Quality context stops AI from having to reconstruct it at runtime, which burns tokens and hands back answers it can’t use,” Arvind said. Better context means fewer tokens and better answers at the same time. The delivery layer is the AI Gateway, which pairs model choice with governance: closed and open models routed through one place with Glean Protect controls, so customers can pick models that meet their data retention and regulatory requirements.

The steal: the durable asset in enterprise AI is the years-long graph of how work actually happens, not a thinner MCP connector to the same systems. Then audit which model is handling your simplest, highest-volume tasks — if a frontier model is answering questions a mid-tier model could answer, you’re paying the 36x spread for a 22% gain you don’t need. Build (or buy) a routing layer that matches task difficulty to model cost, and measure it in dollars per task, not tokens.

6. Will model companies eat every startup idea (and what happens to jobs)?

We closed the interview with the two questions on every founder’s mind: Dario Amodei’s much-quoted prediction that AI will erase 50% of white-collar jobs in the next 5 years, and what Arvind would say to founders starting a company right now.

On jobs, he’s a clear skeptic of the doomsday number. “It’s a very bold statement and obviously I don’t think there’s any kind of deep analysis or research that was done to come up with the number 50. The fact that the number is 50 means it’s more figurative... But I kind of disagree with that regardless. I don’t think people are going to be replaced with AI in the sense that there are no jobs for us anymore. I think we’ll all have jobs and I think the nature of our jobs is going to change” — the same way his own work has continuously evolved in tech.

On founders, Arvind spends real time reassuring people who fear the model companies will eat everything. He hears it constantly: “Every idea they come up with, you know, it feels like, you know, you can just use an AI model or the model companies, you know, just come and do that.” His worry is that founders “start to give up too soon” and never build.

His counter is that change is the entrepreneur’s raw material. “AI changes, you know, if anything, you know, that AI changes, it actually creates more opportunity for entrepreneurs. Because it brings change, and whenever there is change, that is great for entrepreneurs,” Arvind said. And he doesn’t think the model layer can cover the surface area: “Model companies won’t be able to meet even 10% of all the needs, you know, that the market is going to have.”

The OpenAI panel added the sharpest version of the upside case. Alexander predicted a step change in what one person can do: “We are now going to enter a situation where the most productive individuals could be running unicorn companies, and folks who are less productive are, well, where they are today.” His prescription: hire high-agency people and structure teams to unlock their creativity, then scale it. Alexander framed it as an old pattern, not a new one; since the history of our species, someone who is good at using tools is faster and more productive than someone who is not.

The steal: pick a specific workflow the model companies will never prioritize (your 10%+) and build the context and product around it. Then staff for leverage, hire high-agency people who can each drive AI across multiple functions, because the ceiling on individual output just jumped.

7. The OpenAI panel: a 3-layer playbook for deploying agents safely

The panel with OpenAI’s Alexander produced the cleanest framework in all three sources for deploying agents without spooking the security team. It starts with a simple rule: an agent acting for you inherits exactly your permissions, no more. “If they’re running on behalf of me, these agents should not be given access to enterprise data and information that I did not have rights to see or use,” Arvind said, adding that they “should not have the ability to make any changes into the system that I could not make.” Then you scope further, because an agent is built for a specific set of tasks and shouldn’t wield all the powers you personally hold.

The third case is the new one: independent agents with their own identity. “Take your customer support team. Maybe you have 10 people in the team today and now this agent is the 11th member of that team,” Arvind said. Once an agent has its own identity rather than impersonating a person, you need a distinct permissions and guardrails model so it behaves like a good teammate, “adding positive value and not actually escalating and, you know, fighting with the humans inside.”

Alexander distilled the whole thing into three layers. “One is the control of what the agent can do. The second is the observability of that. And then lastly, starting to shift away from agents always impersonating us as people and having more agents that have a specific role with exactly the right permission set for that workflow, and that’s an independent agent.” His rollout advice was to start narrow and expand: connect agents to the most important context first, then add more incrementally. “Your security team will be very happy if you add one and then you add the other one. You can actually move really fast if you do it incrementally.”

This connects straight back to the ROI question from our interview: the enterprises that actually reach production-quality agents are the ones whose security teams never got a reason to say no. Glean’s three-layer model is how it sells “safe to deploy” alongside “cheaper to run.”

The steal: deploy agents in exactly this order. Mirror the user’s permissions, scope down to the task, then graduate the most valuable workflows to independent agents with their own identity, observability, and guardrails. Give your security team an incremental rollout they can approve one connection at a time, and you’ll ship faster than a big-bang launch that gets blocked.

The full interview Q&A

Every question we asked Arvind in the 20-minute interview, one by one, with his answers (lightly cleaned of filler words; the sections above expand on each theme).

Q1. How did you end up here? What’s your personal story as a founder?

Glean started in early 2019 while he was still co-founder at Rubrik, where fast growth had quietly wrecked internal productivity: “Everybody would complain about not being able to find things inside the company or connect with the right people to get help.” Before Rubrik he spent over a decade at Google building search systems, so the fix was obvious to him: “We should actually build a really good search product for all of our enterprise data and knowledge. And that’s how Glean got started.” Personally: grew up in India, moved to the States in the ‘90s for grad school, and has been in tech — and based in San Francisco — ever since.

Q2. You just raised your first round — who do you hire first?

“Hire software engineers, people who are going to build the product, because right now that is all we need.” Everything else — office, devices, food, finding early customers — he handled himself: “The first 10 people that we hired at Glean, they were all engineers. Maybe even the first 20, I don’t remember hiring anybody outside of R&D,” until the product reached a certain level of maturity.

Q3. How do you communicate the vision of the product to so many engineers?

“Simple ideas are easy to communicate. So if you think about Glean’s idea, it was actually very, very straightforward. It’s a Google for you in your work life. It’s a place where you go and ask questions when you’re looking for information or answers and we’ll bring the right information back to you.” The second half is belief: “They’ve got to feel that it’s going to be something big.”

Q4. Bootstrapping versus raising money — when should a founder raise?

“I personally have chosen the path of raising capital quickly... I want to focus all of my attention and energy into building the product. I don’t want to continuously think every month about, okay, well, how much can we invest? Can we really hire one more engineer or not? I like to go big and go big fast. So if you can, then I would recommend that, just raise capital. What’s the downside?” And on the classic objection: “Sometimes people feel that, oh, is it too early to dilute the company? And I don’t really believe in that.”

Q5. Was raising money easier the first time or with Glean?

Definitely easier the second time: “Investors are actually never looking to invest in great ideas. They’re actually looking to invest in people.” Ideas get transformed once you talk to customers, so investors underwrite persistence — “do you have the persistence to stay on that journey?” A track record answers that. He also notes today’s market is unusually founder-friendly: “There’s a lot of capital that is actually waiting to be deployed on those ideas.”

Q6. Dario Amodei says 50% of white-collar jobs could be erased in 5 years. What do you think?

“It’s a very bold statement and obviously I don’t think there’s any kind of deep analysis or research that was done to come up with the number 50. The fact that the number is 50 means it’s more figurative... But I kind of disagree with that regardless. I don’t think people are going to be replaced with AI in the sense that there are no jobs for us anymore. I think we’ll all have jobs and I think the nature of our jobs is going to change” — the same way his own work has continuously evolved in tech.

Q7. Uber reportedly burned its 2026 AI budget in four months without ROI — do enterprises actually see ROI from AI?

“It’s a wide spectrum, but businesses are definitely getting some value from AI today. If they’re getting none, they would have stopped by now.” The clearest value is basic knowledge-seeking — “every employee in the company becoming more productive, more knowledgeable” — “but it’s hard to measure. The fact that people can work faster oftentimes doesn’t translate into your P&L statement.” Larger companies with intentional programs, good metrics, and systematic process-by-process adoption are seeing results — “it is hard, it takes a lot of time, and AI vendors kind of made it seem like you’ll automatically and instantly get all these value savings.” Glean’s model is built for that reality: immediate productivity savings up front, then “many, many months understanding their business and helping them build those agents and get them to production level of quality.”

Q8. Have you found any use case you didn’t think of when you started?

“Many. Actually, almost all use cases are things that we didn’t imagine people would actually use our product for.” His list: a performance-management agent that “writes the first drafts” of annual reviews, customers “responding to RFPs” with AI, teams “reviewing legal contracts with AI,” and salespeople using AI “to figure out what messages to send to which one of their accounts.” Why it works: “Since Glean is so horizontal and it connects to all enterprise systems, it sort of functions as your chief of staff.”

Q9. Who do you reach out to to sell Glean in the organization?

“Because we’re horizontal, typically we’ll start with the CIO. The CIO typically thinks about how to bring company-wide AI technologies. And sometimes we do work with departmental leaders too, because for example, for salespeople, for CROs, we do a lot — there’s a lot in Glean that can help you truly improve the productivity of your sales team.”

Q10. OpenRouter is reportedly being acquired by Stripe — what do you think of that?

“I’m very familiar with them. They were one of the first model routing companies, but more sort of model access company than routing. And Glean does some of the same things that OpenRouter did, but we do it mostly in the enterprise context for large enterprises. But our business is actually very different. Model access is actually only one part of our overall solution. Largely, we think of Glean as the context and intelligence company” — the goal being to bring “the collective human intelligence and the know-how, the knowledge and experience of your enterprise in one place and then make it available to AI systems.”

Q11. For your agent builder — do customers build the agents themselves, or do you send in your engineers?

Customers build them, with Glean enabling rather than staffing: “We also have training and enablement programs. And in fact, sometimes we actually bring our partners to help build agents for our customers.”

Q12. What’s the main location of your customers — your go-to-market, roughly?

“We have customers worldwide. About 70% of the business is probably US and then 30% is the rest of the world right now. So we are in Europe, we are in Asia, in Canada, Latin America. And typically we work with mid-sized to large enterprises.”

Q13. To close — anything you’d like to share with founders starting a company right now?

He hears the same fear constantly: founders feel “every idea they come up with, you can just use an AI model or the model companies just come and do that... Is there any value to ideas left? What I fear and what I worry is, will people start to give up too soon and not become entrepreneurs?” His answer: “AI changes, if anything, actually creates more opportunity for entrepreneurs. Because it brings change, and whenever there is change, that is great for entrepreneurs. And model companies won’t be able to meet even 10% of all the needs that the market is going to have.”

What to steal

1. Recruit and fund around a one-sentence mission. If you can’t explain your product in the length of “Google for your work life,” it’s too complex to hire engineers around. Test yours on a non-technical friend this week; if they can repeat it, it’s tight enough.

2. Raise to eliminate distraction, not to optimize valuation. Arvind’s rule is to raise enough that you never spend a month calculating whether you can afford one more engineer. Frame your pitch around your persistence, because investors bet on people, not ideas.

3. Sell (and expect) months, not minutes. The “botsitting” tax is 6.4 hours a week per worker (Glean’s Work AI Institute) because vendors oversold instant magic. Commit to one high-cost process, build production-quality agents over months, and instrument the before/after so value actually reaches the P&L.

4. Sell horizontal platforms to the CIO, expand through surprise. Start with the company-wide AI owner, win one sharp department (CRO/sales) in parallel, then let the use cases customers invent — performance reviews, RFPs, contract review — pull your expansion and write your roadmap.

5. Own context, not the model, and route by task difficulty. Structured data is 5-10% of what a company knows; the moat is the years-long graph of how work actually happens, surfaced from emails, messages, and meeting notes. Then match model cost to task difficulty: the 36x spread for a 22% quality gain is pure waste, and Glean’s benchmark shows routing plus context can cut token costs 81% while winning 78% of comparisons.

6. Deploy agents in three layers so security says yes. Mirror the user’s permissions, scope to the task, then graduate the best workflows to independent agents with their own identity and observability. Incremental rollouts ship faster than big-bang launches that get blocked — and they free high-agency people to run what used to take whole teams.

Hope this was valuable!

Cheers,
Guillermo

AI demand is growing 10x a year while the hardware underneath it grows 20-30%, GPU supply is booked out to 2028, and chips are the one thing in computing history whose price is going UP.

a16z just put $1.1 billion behind that gap with the Machine Age Fund, and we pulled the conversation between Ben Horowitz, Martin Casado, and Raghu Raghuram breakdown into the exact thesis, the hard numbers, and the founder profile they are writing checks to, so you can see where the next decade of AI value is actually moving before you build or raise into it.

Know the panel before you read them.

• Ben Horowitz cofounded a16z and has invested through every compute epoch since the client-server era.

• Martin Casado leads the firm’s infrastructure practice and pioneered software-defined networking at Nicira before venture.

• Raghu Raghuram is an a16z managing partner and former VMware CEO, the one person here who ran data center infrastructure at scale for a living.

Three people with hardware in their bones are telling you the model is no longer where the money is.

In this guide you’ll find:

1. Why a16z Says the Model Is No Longer the Bottleneck

2. The Supply Crunch That Turns AI Into a Physics Problem, Not a Growth Problem

3. Why This Is Nothing Like the Dot-Com Dark Fiber Bubble

4. What Capital Can Now Buy in AI That It Never Could Before

5. The Rebuild: From 5kW Racks to DC Power, Concrete, and Per-Model ASICs

6. The Exact Founder and Check the Machine Age Fund Is Writing

👋 Just a quick note before you continue!

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

Here’s a good place to start:

↳Alex Karp’s Playbook for Which AI Companies Survive

↳19 great new details about Cursor’s rise

↳Perplexity’s CEO 2026 AI Pitch

↳The First 48 Hours of Claude Fable 5

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

Read more

OpenAI ran for four and a half years before launching a single product, spent most of those early days failing to raise money and got hammered by the intellectual giants of the field for saying AGI was even possible.

We pulled Sam Altman’s full conversation with David Senra apart and extracted the exact operating playbook he used to bet on the impossible: how he picks non-consensus bets, survives with no customer signal, decides when to kill good products, and why shipping embarrassing V1s to real users is how you build a company that’s both great and safe.

Inside you will find:

1. Why 2026 Is the Year Every Business Is Up for Grabs

2. The iPhone Moment Hasn’t Happened for AI Yet

3. How to Spot a Real Non-Consensus Bet

4. Why Being an Investor First Made Altman a Better Founder

5. Why Shipping to Real Users Is How You Make AI Safe

6. Why Nobody Has Built for the Small-Business Boom

7. Context Is the Next Unlock

8. Why OpenAI Chose Platform over Product

9. Killing good ideas: Sora and Atlas

10. How Altman Ran OpenAI for 4.5 Years with No Customer Signal

👋 Just a quick note before you continue!

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

Here’s a good place to start:

↳Alex Karp’s Playbook for Which AI Companies Survive

↳19 great new details about Cursor’s rise

↳Perplexity’s CEO 2026 AI Pitch

↳The First 48 Hours of Claude Fable 5

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

Read more

A disclosure up front: I’m an investor in Apollo Atomics, which just came out of stealth with a seed round led by FCVC. I’m writing this because I think the team found the one insight the rest of the nuclear industry walked past. Every other startup is trying to shrink or reinvent the reactor. Apollo left the reactor alone and attacked the largest, hand-built, several-stories-tall component nobody talks about: the steam generator.

In this issue you’ll find:

1. The steam-generator insight: why Apollo left the reactor alone and attacked the part nobody thinks about

2. Not an exotic SMR: how using proven light-water tech is the moat, not a limitation

3. The 3c/kWh math and the 24-month clock: the cost and speed case against natural gas

4. Why now: the AI power crunch that turns firm baseload into the scarcest input in tech

5. The founders and the proof: an MIT reactor that went critical, a factory operator, and a YC batch that’s betting on hard tech

Hey, welcome to AI Market Fit!

Every week you’ll get two things, fast:

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

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, you can subscribe here:

Subscribe now

1. The steam-generator insight: why Apollo left the reactor alone and attacked the part nobody thinks about

The entire thesis rests on one component almost nobody outside nuclear engineering thinks about. Every nuclear startup chasing lower costs is trying to invent a new reactor. Apollo’s founders decided that was the wrong problem. “We focus on the largest component, the steam generator,” CEO Assil Halimi told TechCrunch. “We have been able to shrink that component and make the whole system way more compact than any other reactor in the market.”

A steam generator turns the reactor’s heat into steam to spin a turbine. In conventional plants these are hand-built and several stories tall. Apollo’s Compact Steam Generator is roughly 20x smaller, about the size of a person, and mass-manufacturable. Because that one component drives so much of the system’s bulk, shrinking it lets Apollo build a reactor 40x smaller than one using a conventional generator.

The technical mechanism is the moat. Most steam generators run reactor coolant through large tubes surrounded by the water that becomes steam. Apollo instead threads both fluid loops through what TechCrunch describes as “a compact block of metal that’s laced with needle-thin channels,” transferring heat far more efficiently and letting the whole thing collapse in size. Halimi’s read on why the old designs are so bad is the tell: existing generators, he says, “borrow heavily from those used by coal and natural gas power plants,” fuels with far lower energy density than uranium. The result, in his words: “You have this high-potential fuel, but you’re converting just a small piece of it.”

For decades, nuclear’s cost problem has been treated as a reactor problem. Apollo says it’s a heat-exchange problem, and heat exchange is a manufacturing challenge, not a physics gamble. When you screen the next wave of energy-infrastructure bets, this is the pattern to look for: a team attacking the one component that drives disproportionate cost while leaving the proven parts untouched.

2. Not an exotic SMR: how using proven light-water tech is the moat, not a limitation

The most contrarian thing about Apollo is what it did NOT change. Most small modular reactor startups differentiate with liquid sodium, molten salt, or other exotic fuels. Per tumisangbogwasi’s breakdown, that choice “forces them into a labyrinthine, decades-long regulatory approval process with zero existing supply chains.” Apollo kept everything boring.

Apollo builds a pressurized water reactor using the same licensed light-water technology that powers roughly 80% of the world’s nuclear plants, commercial-grade low-enriched uranium, and existing qualified supply chains. Its own site claims more than 15,000 reactor-years of commercial operation stand behind that fuel and architecture. As PRNewswire puts it, Apollo chose to “not replace the light water, commercial-grade low-enriched uranium fuel and established supply chains that underpin decades of reliable and safe PWR operating experience,” and instead redesigned “the largest and most complex component in the nuclear steam system: the steam generator.”

This is the part investors miss. The exotic-fuel SMRs are betting on supply chains that don’t exist yet. Apollo’s bottleneck is manufacturing a metal heat exchanger, not standing up a fuel industry from scratch. The company is advancing a commercial regulatory pathway with the NRC and has submitted a regulatory engagement plan, a far shorter road when the underlying tech is already the most-operated reactor type on Earth.

The stealable lesson for anyone screening hard-tech deals: proven-tech-plus-novel-manufacturing beats novel-tech-plus-unproven-everything. The reactor being unremarkable is precisely why the timeline can be remarkable. That’s the difference between a 24-month product and a company that spends a decade proving basic physics to regulators before it ever pours concrete.

3. The 3c/kWh math and the 24-month clock: the cost and speed case against natural gas

Apollo’s target is explicit and aggressive: 3 cents per kilowatt hour. “We don’t want to be incremental in nuclear and just improve a small thing,” Halimi told TechCrunch. “Our target is to beat natural gas.” That’s the benchmark that matters, because natural gas is cheap to build and quick to deploy, which is exactly why it has dominated new power generation.

The cost case flows directly from the compact steam generator. Because the reactor and generator are small enough to assemble in a factory rather than on site, Apollo captures labor savings that bespoke megaprojects never can. Halimi expects to build a 300-megawatt power plant in less than 24 months, and predicts the reactor itself will cost four to five times less to produce than existing designs. The whole unit fits on a truck.

Set that against the incumbent path. Gigawatt-scale nuclear takes more than 10 years and roughly $20 billion, delivered as a one-off construction project. Apollo’s roadmap replaces that with three standardized, truck-transportable units: the A-10 (10 MWe), the A-50 (50 MWe), and the A-300 (300 MWe). Factory-built and repeatable beats hand-built and bespoke on both axes that killed nuclear economics: time and cost.

Here’s how to evaluate it yourself. The claim decomposes into three testable pieces: whether the compact generator actually delivers an order-of-magnitude higher power density (the physics), whether a 300 MW plant can genuinely be factory-assembled in under 24 months (the manufacturing), and whether that stack of savings lands at 3c/kWh (the economics). The honest risk is the one TechBuzz names directly: “the nuclear industry is littered with promising startups that couldn’t cross the valley between innovation and deployment.” A 40 kW demonstrator is not a 300 MW commercial plant. The scale-up is unproven, and that’s exactly what the round is meant to fund.

4. Why now: the AI power crunch that turns firm baseload into the scarcest input in tech

The demand side is why the timeline matters more than the technology. As PRNewswire frames it: “The defining constraint of the next industrial era may not be compute or manufacturing capacity. It is the ability to reliably and timely power its infrastructure.” Microsoft, Google, and Amazon have all signed deals to explore nuclear for their data centers, per TechBuzz, because they’ve hit the wall.

That wall is the interconnection queue. When hyperscalers ask utilities to plug in new data centers today, regional authorities hand them wait times “stretching into the next decade,” per the tumisangbogwasi analysis. You cannot build the future of intelligence on a seven-year utility waiting list, and solar and wind can’t provide the round-the-clock baseload a continuous AI training run demands. That leaves firm, carbon-free, fast-to-deploy power as the scarce input, and almost nothing on the market delivers it before 2030.

Apollo has already converted this demand into paper. The company reports 20 GW of signed LOIs in its commercial pipeline, targeting data center operators and hyperscalers, industrial offtakers needing process heat or behind-the-meter power, and utilities evaluating SMRs. It has also joined the AI for Nuclear Energy Consortium led by Idaho National Laboratory to bring AI into reactor design, licensing, manufacturing, and operations.

For the reader sizing this market: the adjacent bets sit around firm baseload for data centers, not around chips or models. Anything that shortens the path from power demand to power delivery before 2030 (behind-the-meter generation, interconnection workarounds, factory-built energy hardware) is where the scarcity premium lives. The 20 GW of LOIs is the demand signal to watch. If those convert to orders, the thesis is validated by customers, not by pitch decks.

5. The founders and the proof: an MIT reactor that went critical, a factory operator, and a YC batch that’s betting on hard tech

The team is built for exactly the two problems this bet requires: nuclear physics and manufacturing at scale. CEO Assil Halimi holds an MIT Nuclear Engineering PhD and brings a decade of reactor design and operations experience working with leading reactor manufacturers and utilities. He’s the reason the reactor side is credible.

COO Drew Walker covers the half of the problem that kills most hard-tech companies: making it in volume. He previously founded an electric boat company, scaled manufacturing and operations at an electric truck company, and led operations at the White House. The factory model only works if someone has actually run a factory, and Walker has.

The proof isn’t a slide, it’s hardware. Apollo built and tested a 40-kilowatt demonstration reactor within MIT’s Department of Nuclear Science and Engineering, and its fuel went critical before July 4th. The roadmap runs from that demonstrator to a 1 MW demonstrator in 2027 and commercial deployment in 2028.

The $26M seed was led by FCVC with Y Combinator, TeleSoft Partners, Alumni Ventures, Robinhood Ventures, Nucleation Capital, Pelion VC, and Duke Capital Partners, plus individuals Paul Graham, Evan Meagher, Matteo Franceschetti, and Ray Rothrock. After office hours with Halimi, Graham called it “very, very impressive,” and said it “brought home what an opportunity we missed by ignoring nuclear power for 40 years” .

Now the contrarian frame. Apollo sits in Y Combinator’s Spring 2026 batch of 196 companies, and that batch is overwhelmingly software. Across the full dataset, agents lead at 48 companies (24.5%), while YC’s “Industrials” label covers just 25 companies (12.8%), down 1.8 points from Winter 2026. The “Hard Tech” tag is new to the batch’s top tags at 4 companies. Against a field of peers like Akkari, Thomas, Clawvisor, Playabl.ai, Mochatrade, Runtime, YouArt, Tolmo, General Aviation, and Hessian, a factory-built nuclear company is a genuine outlier.

Hope you enjoyed this one!

Cheers,

Guillermo

Every enterprise search product before Glean had “universally failed,” and Arvind Jain was cold-DMing strangers on LinkedIn who wouldn’t give him 30 minutes. IGNORE, not a violation, Glean is over $300M in revenue (TechCrunch, May 2026) and valued at roughly $7.2B (Fortune, June 2026), and the rule that got him there never changed.

We broke down the full 12-minute Arvind Jain interview with EO into the exact sequence he ran: how he validated the idea, why he gave the product away free for two years, how he got Sequoia to come to him without a pitch, and the rent the model, own the enterprise layer strategy behind all of it. His edge was building less than his rivals.

In this guide you’ll find:

1. Why Jain walked away from one unicorn to build another (the productivity collapse that started Glean)

2. Validating an idea nobody wanted to buy: the LinkedIn cold-outreach test and the “universal yes”

3. The 2-year free beta that built $7B: the Google-quality bar, 20 design partners, and the user revolt that pulled Sequoia in

4. Rent the model, own the enterprise layer: the build-less strategy behind a pure-play AI company

5. The rule that never changed: conviction is the founder’s job, and don’t be the one who kills your own idea

👋 Just a quick note before you continue!

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

Here’s a good place to start:

↳Alex Karp’s Playbook for Which AI Companies Survive

↳19 great new details about Cursor’s rise

↳Perplexity’s CEO 2026 AI Pitch

↳The First 48 Hours of Claude Fable 5

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

Read more

Ken Griffin puts a number on the scariest risk in tech: if the US loses access to Taiwan’s chips, GDP falls 8% in six months, a great depression in the blink of an eye.

We went through the full most recent Goldman Sachs conversation with Citadel’s CEO, and pulled out the moves that actually matter to founders: why agentic AI compressed six weeks of PhD work into hours, why that collapses moats while raising a compute-cost wall, and the stress-test framework he uses to survive tail events, so you do not have to watch 31 minutes to steal it.

In this guide you’ll find:

1. Why Griffin dragged everyone back to the office 5 days a week and the Fed paper proving remote work hurt youth more than AI

2. The Warp Speed incentive design that turned a few billion into trillions saved

3. Why the S&P hit all-time highs during a Middle East war and why most ‘AI transformation’ was machine learning all along

4. How agentic AI compresses six weeks of PhD work into hours, fills moats at lightning speed, and lets tiny teams beat incumbents

5. The new moat is compute cost: why all compute is used all the time and low-margin businesses get filtered out

6. The 8% GDP tail risk: China leads 67 of 74 technologies, losing Taiwan’s chips triggers a depression, and how to stress-test your own exposure

👋 Just a quick note before you continue!

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

Here’s a good place to start:

↳Alex Karp’s Playbook for Which AI Companies Survive

↳19 great new details about Cursor’s rise

↳Perplexity’s CEO 2026 AI Pitch

↳The First 48 Hours of Claude Fable 5

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

1. Why Griffin dragged everyone back to the office 5 days a week and the Fed paper proving remote work hurt youth more than AI

Read more

Sequoia recently told the market that the next trillion-dollar AI companies will sell work, not software. Instead of charging for software seats, they will complete the work itself and capture a share of the much larger global labor budget.

A few weeks ago, I broke down why that thesis is so compelling.

Then Andrew Côté published a rebuttal that exposes the uncomfortable weakness inside it:

“Using AI for a task trains AI to do that task, so every services niche eventually becomes a niche the model owns outright.”

That is not always technically literal. Enterprise providers may not train on customer data, companies may self-host models, and proprietary workflows do not automatically flow back to foundation-model labs.

But economically, the direction is the same: as foundation models improve, more of the value created by thin AI service layers becomes commoditized.

That breaks the simplest version of Sequoia’s thesis.

So where does the durable value go?

I pulled apart both theses to map where the money may actually go next, including physical markets where a single industrial process can represent hundreds of billions of dollars in annual activity.

In this issue you’ll find:

1. Why “sell work, not software” may train its own killer

2. Software investing is close to over

3. The five enormous atoms markets Côté is pointing toward

4. How AI and robotics could bring software-like leverage to physical industries

5. Who is already building the atoms version of this thesis

6. What founders and investors should start, back or partner with before this becomes consensus

Powerful models are only as useful as the company context they can access!

Make enterprise AI actually work with Glean, the context layer that connects every model, tool and workflow.

Glean connects your organization’s knowledge, people, permissions and activity to give AI the understanding it needs to deliver relevant, secure and work-ready results.

With one governed context and intelligence layer across tools like Claude, Cursor, Copilot, and custom agents, Glean helps your company:

• Give AI the right context: connect knowledge across 275+ enterprise applications while preserving permissions, metadata and access controls

• Get more value from every model: improve retrieval, reduce unnecessary tool calls and use approximately 30% fewer tokens than off-the-shelf MCP tools.

• Stay flexible as AI evolves: access 35+ models (including Claude, GPT and Gemini) without rebuilding your context layer or locking your company into one provider

Try Glean Now!

1. Why “sell work, not software” trains its own killer

The flaw in Sequoia’s “sell work, not software” thesis is captured in Côté’s line:

“Using AI for something trains it to do that task.”

Again, this should be understood as an economic principle, not always as a literal description of model training.

A customer interaction does not necessarily become a training example sent directly to OpenAI, Anthropic or Google. But every successful AI workflow reveals what can be automated, which tools are required, where humans intervene and what a correct output looks like.

That knowledge spreads.

Model providers improve their reasoning and tool use. Competitors replicate the workflow. Open-source models catch up. Customers learn to perform more of the process themselves. What was once a differentiated service gradually becomes a native model capability or a standard feature.

Côté’s conclusion is that “every possible niche where you can use AI to do something better eventually becomes a niche occupied by AI.”

That does not mean every AI services company disappears. It means the service itself is not automatically the moat.

An AI services business that wraps a foundation model and sells its output may initially capture attractive margins. But its advantage often comes from a temporary gap between what the customer can do directly and what the company can orchestrate on the customer’s behalf.

That gap closes on the model’s schedule, not the startup’s.

Sequoia’s own feed illustrates the direction. Promoting Factory AI’s Matan, the firm wrote that almost every token today begins with a human asking a model to do something, but that within 12 to 24 months, 90% of tokens could be machine-initiated.

That human typing “hey, go do this” is precisely the human-in-the-loop signal that teaches the model to remove the human. Sequoia frames the shift from human-initiated to machine-initiated tokens as the bull case for AI labor. Read it Côté’s way and it is the countdown clock on every services wrapper: the closer we get to 90% machine-initiated, the less any middle layer between customer and model gets to keep.

So run this test before you build or back an AI services company: what remains defensible after the foundation model has ingested a year of your usage? If the answer is “our workflow,” “our prompts,” or “our fine-tune,” you are selling the training data for your own obsolescence. If the answer is a proprietary distribution channel, a regulated license, or a physical asset the model cannot replicate, you have a real business. Most services plays fail that test.

2. Software investing is close to over

Côté’s second claim lands harder than the first: “Software investing is over. Software was never more than 2% of the economy anyway.” The trillion-dollar software franchises are already public and obvious. Betting on a “next generation” of $1T pure-software names is betting on 2% of GDP to keep minting the outcomes that a saturated category already produced.

This reframes the Sequoia thesis as a fight over the wrong prize. Sequoia is right that the labor budget dwarfs the software budget, which is exactly why “sell work” sounds bigger than “sell seats.” But labor-as-software still lives inside the services trap from section one: it captures labor spend only until the model absorbs the task. Côté’s move is to skip the trap entirely and go where the spend is both enormous and physically defensible. As he puts it, “the hard business of technology lies ahead, which is remaking the physical world. The other 98% of the economy.”

The number that reframes everything is his growth claim: “Almost all economic growth is in the future.” Software’s growth is largely behind it; the category matured and the winners were crowned. The physical economy’s AI-driven growth has not started. If you are allocating for the next decade rather than trading the last one, a mature 2% category is a strange place to concentrate.

Hardware itself has not suddenly become as scalable as software. What has changed is the intelligence controlling the hardware: AI models, simulation, digital twins and more adaptable robots can be developed once, improved with data and reused across multiple machines and factories.

That means a company can increase output without increasing engineers and workers at the same rate, because one team can supervise more equipment and software updates can improve an entire installed fleet. The economics become more scalable, but materials, energy, machinery, maintenance and deployment costs remain.

3. The five $10T atoms markets

Côté names the destinations explicitly: “Chemicals, Metal, Energy, Transportation, Healthcare.” His scale claim is the part investors keep underweighting: “Each market is 10x the total size of all software. These are not $1T company valuations, they are $10T gross annual revenue markets.”

The unit distinction matters. A $1T valuation is a stock-market number for a handful of names. A $10T gross annual revenue market is the recurring flow of money through an entire sector every year: the actual dollars changing hands for chemicals, steel, electrons, freight, and care. Côté is telling you to stop hunting for the next trillion-dollar market cap and start hunting for a slice of a multi-trillion-dollar annual revenue river that has never had software-grade tooling pointed at it.

He is careful to include the reason these markets have been ignored, giving two in his words: “Lower annual growth rates” and “Lower leverage on value production.” Chemicals and steel grow at GDP-like rates, not SaaS rates, and each unit of value historically required proportional physical input, with no software-style leverage where one engineer serves a million users at near-zero marginal cost. That is exactly why generalist VCs skipped them, and exactly why they are uncontested now.

The bet is that AI plus robotics breaks both constraints at once. Remove the growth ceiling by unlocking demand that was capped by cost and capacity, and add software-style leverage by letting a handful of models and autonomous machines run processes that used to require armies of operators. Do both, and a GDP-growth, low-leverage industry starts to behave like software while sitting on a revenue base ten times software’s size.

For a builder, the actionable move is to pick one of the five and go one layer deeper into a single process inside it, not the whole sector. “Chemicals” is not a strategy. “AI-controlled continuous-flow synthesis for one class of specialty chemical” is a wedge into a $10T sector where nobody has your tooling.

4. How a $400B industrial process gets software-like margins with AI plus robotics

The most quotable and most testable claim Côté makes is the unit-economics one: “You will see software like margins and growth rates in physical industries where the TAM for a single industrial process is $400 billion, where demand is extremely price sensitive and the addressable market is massively under-estimated.”

Unpack “extremely price sensitive,” because it is the engine of the whole thesis. In physical markets, demand is capped by price. When AI plus robotics collapses the cost of a process, you do not just win share, you unlock demand that never existed at the old price. That is how a mature-looking market turns out to be “massively under-estimated”: the visible TAM was measured at today’s cost, and cutting the cost expands the market underneath you.

The software-like margin comes from replacing per-unit human labor with a fixed model-and-robotics cost that then serves rising volume at near-zero marginal increase. That is the same leverage curve that made software gross margins 80-plus percent, now applied to a process that currently runs at industrial margins because every unit of output needs proportional human operation. Automate the operation and the margin structure migrates from steel toward software while the revenue base stays the size of steel.

The tell that this is real and near sits in Côté’s own most recent posts. On July 28 he noted “the kind of visual models you can fit on a drone” (@Andercot): the concrete detail that the perception models needed to run physical processes now fit on cheap edge hardware. The bottleneck was never the ambition, it was models small and cheap enough to sit on a machine in a plant, on a truck, or on a drone. That constraint is lifting now.

The builder takeaway: underwrite the demand unlock, not just the cost cut. Model what happens to volume when your process cost drops 40%. If that cut unlocks 3x the demand, you are not competing for a fixed pie, you are creating the market and you get to price it.

5. Who is already building the atoms version: ATOMS, hard tech, defense, energy, and the playbook to copy

The bits-to-atoms trade is not a forecast; capital is already rotating. Travis Kalanick’s ATOMS is the loudest signal that a top-tier operator is treating the physical economy, not another app layer, as the next platform. Around it sits a widening cluster: industrial AI applying models to plants and supply chains, hard tech building the machines, defense tech where physical autonomy gets funded and battle-tested, and energy, the input every other atoms market depends on.

Côté’s own timeline puts him inside the perception-and-robotics layer of this wave. His July 28 note on drone-scale visual models, referenced above, is a direct read on where the buildable edge is right now: put capable models on cheap physical hardware and point them at a process.

Travis Kalanick's direct quote on what he's building:

"CPU manipulates bits to compute information. Robots and AI manipulate atoms to compute the physical world."

That's not a metaphor. It's a thesis.

Uber was the proof of concept — touch glass, car arrives. Atoms as a programmable layer. Kalanick calls it "sci-fi at the time." It shipped in 2010.

His new company, Atoms, is the follow-through: take that same logic from ride-sharing into every factory, warehouse, and supply chain that still runs on unoptimized physical processes.

The framing matters because it changes the competitive landscape entirely. If the physical world is just an atoms-based computer, then robotics isn't automation — it's infrastructure. The same reclassification that made servers into cloud computing and phones into platforms.

Kalanick's been thinking about this since Uber. That's a long time to sharpen a thesis before building the company around it.

Côté raises the cultural stakes to make the urgency concrete: “Walk through the world with nostalgia and appreciation, you are seeing the last time in history where everything that has been designed and made was wholly made by humans” (@Andercot, July 28 2026). Strip the poetry and it is an investment thesis: the transition of physical making from wholly-human to AI-and-robot-assisted is happening now, which means the winners in each atoms vertical are being seeded in this window.

The playbook to copy from the ATOMS-style plays is consistent. Pick one physical process. Own the machine or the deployment, not just the model, so the foundation-model owner cannot commoditize you. Sell into a price-sensitive market where a cost cut unlocks new demand. Keep the defensible asset physical or regulated, the exact opposite of the services wrapper that trains its own killer.

Defense and energy are the two verticals where this is furthest along, because both fund physical autonomy directly and both are politically insulated from the “just wrap a model” competition. If you want to see the atoms playbook running with real budgets today, those are the two rooms to be in.

6. The builder and investor moves: what to start, back, or partner on before this is consensus

Start here if you are building: pick one industrial process inside chemicals, metal, energy, transportation, or healthcare, and own a physical or regulated asset in it, not a workflow on top of a model. The section-one test is your filter. If a year of your usage would train a foundation model to replace you, do not build it. If your moat is a deployed machine, a plant contract, a license, or a distribution channel the model cannot touch, build it now while the category is uncontested.

Back this if you are investing: companies whose defensibility is physical and whose demand is price-elastic, so a cost cut expands the market instead of just shifting share. Underwrite the demand unlock Côté describes, the “massively under-estimated” addressable market that only appears once cost drops. A $400B single-process TAM with elastic demand and a physical moat is a fundamentally better shape than a labor-as-software play sitting inside a $10T sector but exposed to model commoditization.

Partner on the sensing-and-robotics layer that every atoms vertical needs. Côté’s repeated point about “visual models you can fit on a drone” (@Andercot, July 28) marks the enabling technology: edge-deployable perception models are the shared infrastructure for chemicals, metal, energy, transportation, and healthcare automation alike. If you cannot own a full vertical, own the perception stack that all five need to run.

The one thing to avoid: the pure AI services company whose entire value is completing a knowledge task with a wrapped foundation model. That is the trade Sequoia is talking the market into, and it is the exact business Côté’s one line dismantles. Every task it completes is a training example for its executioner, and the timeline to obsolescence is set by the model owner.

The move this week: take one AI services idea you were excited about and rewrite it as an atoms play. Same customer pain, but instead of selling the completed knowledge work, ask what physical process sits underneath it, whether a model on cheap hardware could run that process, and what asset you would own that the foundation model never can. Answer all three, and you have found a business on the right side of the bits-to-atoms line before it becomes consensus.

What do you think?

Cheers,

Guillermo

Share

Anthropic just deleted over 80% of Claude Code’s system prompt for Opus 5, and the model got MORE intelligent, not less.

Fresh off the Opus 5 launch, Claude Code creator Boris Cherny sat down with YC at Startup School 2026 and laid out the exact playbook for building products when the model keeps outrunning them.

We went through the full 36 minutes and pulled out every stealable move (the ablation method, SKIP - not a real violation, higher-level prompting, dynamic workflows that spawn thousands of agents) so you can rebuild your AI product for Opus 5 without watching a second of it.

In this guide you’ll find:

1. Why Opus 5 Runs for Weeks With No Scaffolding and Ignores Prompt Injection

2. The Delete Playbook: Cut 80% of Your System Prompt and Rebuild It Line by Line

3. Product Overhang: The Model Capabilities Startups Aren’t Shipping

4. Give Claude Harder Tasks Than You Think It Can Do: Bun’s Zig-to-Rust Rewrite in 11 Days

5. Verification Is the Skill Nobody Gets Right: The Two-Week Prompt That Live-Blogged Itself

6. Spawn Thousands of Agents From One Prompt: Dynamic Workflows and Self-Maintaining Codebases

👋 Just a quick note before you continue!

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

Here’s a good place to start:

↳Alex Karp’s Playbook for Which AI Companies Survive

↳19 great new details about Cursor’s rise

↳Perplexity’s CEO 2026 AI Pitch

↳The First 48 Hours of Claude Fable 5

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

Cherny frames the launch with one data point: Anthropic took Arc AGI to 30%, “before the best score was in the low single digits or low teens.” Everything below is his account of how to build products for a model moving that fast.

1. Why Opus 5 Runs for Weeks With No Scaffolding and Ignores Prompt Injection

Read more

SPOILER: SAM BELIEVES WE’VE REACHED SINGULARITY

If your company looks like a startup from 10 years ago, you are already in trouble.

Two weeks before thi

s interview, Altman met a startup that had rebuilt an entire office productivity suite from scratch. Documents. Presentations. Spreadsheets. The kind of product that, until recently, would have required a full team and at least a year of development.

They built it in two weeks.

That is how dramatically the startup baseline has changed.

But most founders are still following the old playbook: hire a large team, spend a year building version one, and launch another “AI agent for vertical X.”

Altman believes that approach may produce revenue but it probably will not produce the defining companies of the AI era.

The biggest opportunities are not in applying today’s models to obvious workflows. They are in building products for the models that will exist two to four years from now.

And according to Altman, almost nobody is thinking far enough ahead.

In a 70-minute conversation with Ti Morse, he explained why startups now have an unprecedented advantage, why the most popular AI startup idea is probably the wrong one, how he identifies founders who will survive, and why he believes we may already have entered the singularity.

We watched the full interview and extracted the five most important lessons for founders.

In this guide, you’ll discover:

1. Why a startup that looks like it did 10 years ago is now “in bad shape”

2. The bet almost every founder is making and why it probably will not define this era

3. Why you should build for the AI model you will not have for another two years

4. The founder skill that cannot be taught, only earned through painful repetition

5. How Altman can tell within minutes which founders are likely to survive

👋 Just a quick note before you continue!

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

Here’s a good place to start:

↳Alex Karp’s Playbook for Which AI Companies Survive

↳19 great new details about Cursor’s rise

↳Perplexity’s CEO 2026 AI Pitch

↳The First 48 Hours of Claude Fable 5

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

Read more

And you, are pro open models or against it?

Leave your opinion here!

Almost 200 Silicon Valley companies, including Y Combinator and Proton, just told the Trump administration that banning Chinese open-weight models would make “hundreds of companies instantly die.” On the same day, 25 firms led by Nvidia, Microsoft, Meta, and Palantir sent their own letter warning against “premature restrictions,” while OpenAI and Anthropic, both gearing up for IPOs, pointedly refused to sign.

Subscribe now

We read every letter, both op-eds, and the actual model releases from the past two weeks, then turned it all into a founder’s playbook for building a model layer that survives whatever Washington decides.

In this guide you’ll find:

1. Why 200 founders say a Chinese-model ban kills their companies overnight (and whether yours is exposed)

2. The cost-discipline playbook: how to match model to task so your unit economics survive at scale

3. Who signed, who refused, and what the OpenAI/Anthropic no-show tells you about the coming lock-in

4. The self-hosting escape hatch: building a model stack no policy shift can shut off

5. The 2026 open-weight shortlist: which models to build on right now

👋 Just a quick note before you continue!

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

Here’s a good place to start:

↳Alex Karp’s Playbook for Which AI Companies Survive

↳19 great new details about Cursor’s rise

↳Perplexity’s CEO 2026 AI Pitch

↳The First 48 Hours of Claude Fable 5

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

1. Why 200 founders say a Chinese-model ban kills their companies overnight (and whether yours is exposed)

If your product runs any high-volume task on a downloadable Chinese model like Qwen or Moonshot because frontier API prices would wipe out your margin, a ban targets your unit economics directly. That is the specific mechanic behind the warning from Suhail Doshi, founder of AI infrastructure startup Particle, who told Business Insider on July 22 that “there’ll be hundreds of companies that instantly die” if Washington blocks downloads of Chinese open weights (businessinsider.com).

The letter came from the newly formed Little Tech Association, which sent its appeal to Trump, Commerce Secretary Howard Lutnick, and OSTP Director Michael Kratsios. It names the models at issue specifically, Moonshot AI and Alibaba’s Qwen, and targets the downloadable weights rather than API access. The founders’ core argument, in their own words: “American leadership requires two things: world-leading American open-weight models and continued access for U.S. builders to open models already available worldwide.” They asked for “targeted safeguards” instead of “broad prohibitions” (politico.com).

Their sharpest practical point is that a ban does not stop proliferation, it only handicaps the domestic side of it. The weights are already downloaded, mirrored, and running on servers worldwide. A U.S. prohibition removes them from the legal stack of American startups while leaving them freely available to everyone else, foreign competitors included.

Run the test on your own product this week. List every model call your app makes, tag each one by provider and by whether the task is high-volume or genuinely hard, and flag anything that would break at frontier pricing. If a Chinese open-weight model is carrying your cheap, high-frequency path, you are in the group Doshi is describing, and the next four sections are your contingency plan.

2. The cost-discipline playbook: how to match model to task so your unit economics survive at scale

Open weights are a survival issue because of cost discipline, not ideology. Microsoft laid out the exact principle in its July 24 essay: open weights “let every organization match the right model to the right job at the right cost, reserving frontier-scale capability for genuine frontier problems and running efficient, specialized models everywhere else.” Microsoft calls that discipline “what will make AI economically sustainable as its use scales into the billions of everyday tasks” (microsoft.com).

Google’s July release cadence is the clearest live demonstration of tiering. In three days it shipped three models at three price-performance points: Gemini 3.6 Flash for general quality work, 3.5 Flash-Lite explicitly “for scaling repetitive use cases like sorting tickets and extracting data,” and a locked-down 3.5 Flash Cyber restricted to governments. The Flash-Lite demo compared it head to head with 3.5 Flash on high-volume tasks:

The lesson: even a frontier lab does not run everything on its best model. It routes ticket-sorting and data extraction to the cheapest tier and reserves the expensive one for hard problems. If a $2T company enforces that discipline, a seed-stage startup burning $8 a call on frontier tokens for classification tasks has no excuse.

Build a routing layer this week: classify each request by difficulty, send the bulk of your volume to a cheap open-weight or Flash-Lite-class model, and reserve frontier calls for the small share of requests that genuinely need them. That single move turns a negative-margin AI feature into a positive-margin one, which is why the ban debate is really a margins debate.

3. Who signed, who refused, and what the OpenAI/Anthropic no-show tells you about the coming lock-in

The signatures tell you who wins under a ban and who loses. The 25-company letter against “premature restrictions” that would “stifle competition” was signed by the infrastructure and diffusion layer, Nvidia, Microsoft, Meta, and Palantir, and reported by CNBC on July 24 (cnbc.com). These companies make money when AI diffuses widely: more chips, more Azure seats, more open Llama downloads, more deployments.

OpenAI and Anthropic did not sign. CNBC notes both are “gearing up for potentially massive IPOs,” and that is the whole story. Closed frontier labs benefit when cheap open substitutes are removed from the legal market, because it pushes every high-volume task back onto their paid APIs. TechCrunch put it bluntly on July 20 in a piece titled “OpenAI is scared of open-weight models” (techcrunch.com). For them, a ban is not a safety measure, it is a moat.

The strongest signal of where the infra giants stand came from Jensen Huang’s first-ever post on X, where the Nvidia CEO shared the letter and argued that open models let organizations “inspect, modify and run AI locally while reducing dependence on foreign providers” (via Stocktwits and Digg). When a CEO breaks a lifetime of X silence for one issue, that issue is existential to his business.

The takeaway for founders: your incentives are aligned with the infra camp, not the frontier labs. Read any pricing or lock-in argument from a lab racing to IPO as self-interested, and design as if the closed labs will not protect your margins, because they have publicly declined to.

4. The self-hosting escape hatch: building a model stack no policy shift can shut off

The one architecture no policy shift can shut off is self-hosted open weights running on your own infrastructure. Once you have downloaded the weights and deployed them in your environment, no ban, price hike, or API deprecation reaches into your stack. Mistral built its enterprise pitch around exactly this, arguing on July 3 that the value from “data, workflows, feedback, and models” should accrue to the enterprise “rather than to model providers” (x.com).

On July 21, Mistral expanded its Microsoft partnership specifically to give “enterprises and regulated industries frontier AI they can control,” pairing open weights with the compute to run them (x.com). A self-hostable model is insurance against every scenario in this fight, including the ones nobody is predicting yet.

Mistral’s Robostral Navigate, released July 8, is the template for a specialized, self-hostable open model. This 8B robotics navigation model runs on a single ordinary RGB camera, hits 76.6% success on the R2R-CE unseen benchmark, and beats multi-camera systems by 9.7 points while using far less sensing. It is small enough to deploy anywhere and good enough to be state of the art at its job.

The concrete move: identify the one or two highest-volume tasks in your product, find an open-weight model in the 7B-to-30B range that handles them, and stand up a self-hosted deployment behind your routing layer as a fallback path, even if you keep hitting an API today. Hugging Face’s “Local AI 201” guide, posted July 21, is a starting point for the mechanics (x.com). If you can flip a config and route around any single provider, no letter to Trump can touch your business.

5. The 2026 open-weight shortlist: which models to build on right now

Build your default stack on models you can legally keep no matter what Washington decides, which means American and European open weights first, with Chinese models isolated behind a switch.

Mistral (French, self-hostable). The safest control-layer bet. Frontier-class enterprise models plus specialized releases like the 8B Robostral Navigate, all deployable in your own environment, backed by the expanded Microsoft compute partnership announced July 21. Use it for anything where control and data residency matter (x.com).

Meta Llama (American, open). The reference open-weight family from a company that signed the anti-restriction letter, which means a strong incentive to keep releasing weights. The default for general-purpose self-hosted workloads that need broad ecosystem support and tooling.

Google Gemini variants (American, tiered API). Not open weights, but the cleanest cost-tiering options: 3.6 Flash for quality, 3.5 Flash-Lite for repetitive high-volume tasks like ticket sorting and data extraction, and 3.5 Flash Cyber restricted to governments. Use these to fill the price-performance tiers your self-hosted stack does not cover (x.com).

Moonshot AI and Alibaba Qwen (Chinese, at risk). The exact models named in the ban debate. They frequently offer the best price-to-performance available, which is why founders depend on them and precisely why they are the legal risk. If you use them, isolate them behind your routing layer so a single config change swaps in a Llama or Mistral equivalent the day a rule lands. The New York Times argued on July 23 that these Chinese open weights are the underlying strength of China’s AI strategy and that “America should take cues” (nytimes.com); treat them as valuable but expendable in your architecture.

Structure the whole stack so no single model is load-bearing: an American or European open-weight default, a tiered API layer for cost, and any Chinese model quarantined behind a switch. Do that, and whatever Washington decides in the coming weeks becomes a config change instead of an extinction event.

And you, are pro open models or against it?

Leave your opinion here!

NOW, He’s Back

He broke down the exact fundraising script, expansion rule, and physical-world framework he used to build Uber and is now reusing to build Atoms.

We pulled out the stealable plays (the uncapped-anchor auction that took Uber’s Series B to a $375M pre, the 98% beachhead rule, and the “atoms computer” pitch) so you can run them without watching all 92 minutes.

In this guide you’ll find:

1. The ‘uncapped anchor’: how Kalanick ratcheted Uber’s Series B from 2x to a $375M pre

2. When the lead backed out: what a $210M reset taught Kalanick about fundraise credibility

3. Why Kalanick blames a missing board seat for losing food to DoorDash at 5% share

4. The ‘atoms computer’: how Kalanick reframes real estate, manufacturing, and logistics into one pitch

5. The stealth tradeoff: how 8 years of silence built a cultural moat but killed the talent pipeline

6. The beachhead rule: why Kalanick won’t start a new business until the last one runs itself at 98%

Subscribe now

👋 Just a quick note before you continue!

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

Here’s a good place to start:

↳Alex Karp’s Playbook for Which AI Companies Survive

↳19 great new details about Cursor’s rise

↳Perplexity’s CEO 2026 AI Pitch

↳The First 48 Hours of Claude Fable 5

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

Read more

• LEARNING OF THE DAY: RETARDMAXXING, if Marc Andreessen is doing it, why shouldn’t you? Learn what it means here!

YC just told founders in one line that the last three years of AI startups are already obsolete: “that’s what most startups from 2023 to 2025 built, tools that help people do their jobs.”

Their Summer 2026 Requests for Startups is a map of where the next wave of money goes instead, and the total addressable market they’re pointing at (services spend) is many times larger than all of software.

We read the full RFS and broke down all six bets into what to build, why the timing just flipped, and which incumbents are exposed, so you can act before the batch does.

In this issue you’ll find:

1. The 2023-2025 playbook is dead: why YC is done funding copilots and wants companies that sell the service

2. The 4 outsourced services YC will fund you to replace (and why outsourced means easiest to kill)

3. Company Brain: the missing infrastructure layer every AI startup is quietly blocked on

4. Why the science bets (agriculture, discovery engines, personalized medicine) all became fundable in the same 12 months

5. The stealable timing signal: how to read an RFS for when a market just opened

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

Here’s a good place to start:

↳Alex Karp’s Playbook for Which AI Companies Survive

↳19 great new details about Cursor’s rise

↳Perplexity’s CEO 2026 AI Pitch

↳The First 48 Hours of Claude Fable 5

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

Read more