Harry Stebbings, Rory O’Driscoll and I went through Nvidia’s ecosystem spending spree, OpenAI’s move from number one to number two, and what happens to CFO budgets when your best people are addicted to tokens.
#1. Nvidia paid $6B for Poolside’s Model Factory, and the trigger was a $2B round that didn’t close
Nvidia is paying $6B for a non-exclusive license to Poolside’s Model Factory, the internal system Poolside used to build its models, and separately investing $1B at a $12B pre-money valuation. 109 engineers got offers to join Nvidia’s open-weight Nemotron effort. The three founders are staying. The investor letter says explicitly that this is not an acquisition and not an acquihire.
Poolside had a six-week window to raise $2B to pay for a 40,000 GB300 cluster coming online in January. They didn’t close it in time, and they lost the cluster.
My read: it read almost depressing. Not being critical, but it showed that infinite capitalism is not as infinite as it looks. Even in the age of AI, and even with Nvidia seemingly funding everyone on planet Earth, the VC gravy train only runs so long. There’s only so much money coming out of big funds.
Rory’s read: it was an excellent letter, and one line does the work. They said they had found themselves on the right side of prediction in a market that scaled exponentially in capital intensity. Translated: we were right three years ago that there was a market for a US open-weight model, we built it, and the capital required for the next turn of the crank is beyond us.
Harry’s read: neo labs are out of favor, and this is universal across every investor he talks to. Nobody was surprised Poolside found it hard to raise.
#2. A $9B outcome sometimes doesn’t clear the bar for seed investing in 2026. Welcome to the AI Era
Harry, a Poolside shareholder, said the outcome is roughly a 15x for them as a seed investor and that it’s pretty great. That’s where I pushed.
My read: 15x sounds good for a late-stage investor. For a true seed fund it isn’t enough. You don’t get a fund return out of a 15x. If you’re deploying a seed fund at today’s entry prices, you need your best deals doing 50x for the math to work, unless you’re running a hyper-concentrated book.
Rory’s math on the dilution: a 15x on a $9B outcome implies an effective entry of $600M, not $60M. The nominal seed price was much lower. The dilution across the capital-intensive rounds is what got you there. To hit 50x, that company needed to exit around $63B.
Rory’s defense of the bet: it was rational at the time. The two closed-source frontier winners are worth a trillion each. The two or three open-weight winners in China are worth $50B to $100B each. The upside case supported a 100x. The thesis broke on capital markets, not on execution, and the team still returned 15x on a bet that didn’t work. If you get 15x on your failures in venture, you’ll die a rich man.
#3. Only four or five companies on the planet can finance a frontier model
Rory’s frame: the VC money ran out on Anthropic and OpenAI a long time ago. That’s why no VC owns more than 1% or 2% of either of them. The only entities capable of financing a state-of-the-art frontier model in the United States are the hyperscalers. OpenAI and Anthropic exist because Microsoft, Google and Amazon wrote checks nobody else on earth could write. Nvidia is now the fifth member of that club, which is why it’s the one financing the open-weight side. VCs have been along for the ride, providing occasional pricing discipline.
The negative implication for everyone else: the next-smartest team going for it just hit the capital wall. Everyone behind them hits the same wall.
#4. Nvidia’s plan is to spend the entire free cash flow on the ecosystem
Nvidia is also in talks to back Mercor in a round at a $20B valuation led by General Catalyst, doubling the $10B mark from October. Harry, an investor, said Mercor is crossing $2.5B in run-rate revenue and that he never thought it would get this big this fast.
Rory’s question: the other Nvidia bets are all TAM expansion. Fund a neocloud, they buy more chips. Fund Poolside, more open-weight inference runs on Nvidia. Fund Mercor, and you get more training data, which doesn’t obviously sell more chips.
My read: there’s no thesis here. They have something on the order of $70B of free cash flow this year, the strat team and top VPs get in a room, everyone brings their best ideas, and the budget gets allocated. Somebody in that room thinks data labeling matters, and his best idea was Mercor. Adding cash to the balance sheet does nothing for a CEO of a profitable company beyond defense. If Wall Street lets you spend it, spend it.
Rory’s one caveat: four years ago Nvidia’s cash flow was a tenth of what it is today, and they only carry about $50B of cash and investments. There’s a version of him that would keep more for a rainy day. He agrees the highest-return use is buying ecosystem viability, especially when the OpenAI-style deals give you a customer, revenue and ecosystem insurance in one transaction.
#5. Kroll’s data: gross margins above 30% no longer earn you anything at exit
Kroll published a look at M&A and large transactions over the last six months.
My read of it: above 30% gross margin, you get no additional credit in M&A or other exits. There’s a penalty below 30. There is no premium above it. Value Mercor on classic multiples for 80% gross margins and growing, and $20B isn’t expensive.
Rory’s caution: don’t overfit on the sample of deals where negative gross margins resolved. Cursor started with negative gross margins, subsidized unlimited usage, had no way to defend it, capped it, built its own models, and got to a $60B outcome. He can also name plenty of deals that started with bad gross margins and ended with bad gross margins. The question is when it’s rational to underwrite massive improvement. His doubt on the training-data companies: Cursor had hundreds of thousands of customers to work margin against. The training companies have three to five.
#6. OpenAI’s reacceleration story is existential, not promotional
OpenAI’s CFO Sarah Friar told employees the company will be public by 2027.
Rory’s read on why the narrative shifted when it did: Q1 to Q2 was roughly $5B to $6B in GAAP revenue, about 18% quarter over quarter. That annualizes to just under 100% and would put them somewhere under $30B in GAAP revenue this year against about $12.5B last year. In isolation that’s a spectacular business. Against a competitor at a $60B run rate mid-year and growing faster, anyone can model it into irrelevance in two years.
Then there’s Broadcom, Nvidia and everyone else planning to sell $200B of chips to OpenAI. They run the same math and ask whether that demand is real. So sitting on the number for three months was never an option. A concerted “Q2 was an anomaly, Q3 is exploding” push was the only move available.
My read: I was stunned by the 18%. At almost any other company that number is a celebration. Relative to what everyone had priced in, it was a disappointment.
#7. Being number two is worse when there are seven candidates for number two
My read: at the start of the year there were effectively two choices, and everyone wanted multi-model anyway, so you got a clean oligopoly bake-off. You hire a few Salesforce people, they put up a slide, it’s us versus them, and that motion sells. Today there are eleven or twelve competitors running inference on open weights with performance close enough to matter. It’s hyper-competitive for number two, and especially hard if you’re the premium product, because you can’t compete on price and you’re left with brand and security.
Rory’s version: when a low-cost competitor with nearly equivalent functionality is entering your industry, you would far rather be number one than number two. Number one says buy us because we’re number one. Number two says please buy us, and also please don’t buy the cheap guys. It’s bad enough to be number two in an industry you invented. Being number two with a wall of ankle biters underneath is a different problem.
Harry’s data point: OpenRouter’s own data showed roughly 68% of tokens going to open weights and climbing. Rory’s caveat still holds: the majority of tokens will be open weight, and the significant majority of revenue will still be frontier, because frontier commands more value than the price of inference.
#8. “It’s all about code” is the only sentence that mattered
I said on the pod I’ve lost the plot on what OpenAI’s differentiated mission is today. Not the evals, not the benchmarks. The mission. These have been the most mission-driven organizations of our lifetimes, and I couldn’t tell you what I’d rally around as an employee or a shareholder now versus the alternatives.
Harry’s counter: the consumer brand is the asset. To large majorities of the world’s population, AI is ChatGPT. He’s in awe that Sam isn’t standing up and saying we are the next Google, our business is advertising, and wait until you see the usage when there’s a consumer hardware device.
My read: that was the plan. Sora was part of it. It just wasn’t the highest-ROI use of limited compute.
Rory’s framing: sometimes when you look back at an outcome you realize only one sentence mattered. If you’d been on the Yahoo board for ten years and all you’d done was scream “it’s all about search,” you’d have made them a hundred billion dollars. Anthropic’s sentence is: it’s all about code. Coding is the fastest-adopting market, the highest-ROI market, and the mother lode. OpenAI focused here and Anthropic focused there.
The economics underneath make it worse. On the power-user consumer plans you’re selling $8,000 to $12,000 worth of tokens for $200. That works when the chat app is a proof of concept for a platform. As a standalone it’s one of the worst business models of our lifetimes. Rory’s fairer long-run view: as cost to serve falls and advertising gets built, that consumer position could become a business within striking distance of Google’s consumer business over a decade. The adoption S-curve for consumers was steep. The propensity to pay was near zero. On the coding side, the propensity to pay is high.
#9. If you’re an enabling technology for open weights, this is your window
Hugging Face is reportedly in play at around $13B on roughly $150M of revenue.
My read: I’m not smart enough to underwrite $13B on that revenue. The timing I do understand. This is the moment when open weights went from experiment to mainstream, so the numbers of any company riding that transition look their best right now. Eli Gil said sell if you have an AI asset. I’d go further: if you have an AI product benefiting from the transition to open weights, there is not a better window than plus or minus 90 days from today. If I could get north of $7B or $10B, I’d sell. Even at 15x.
Rory’s strategic case for the buyer: if you’re an IT company watching two frontier labs claim TAMs larger than US GDP, you conclude they’re coming for everything and you’d better own something relevant on the non-frontier side. It started with Satya saying every enterprise needs its own knowledge rather than handing it to the frontier. If you’re Microsoft or IBM building a counterweight, Hugging Face is one of the assets on the list.
The condition on any deal: if you buy it, you don’t touch it. Promote anything, put a single ad at the top, and you break the neutrality that makes a 10,000-model marketplace worth owning. Anyone spending $3B, let alone $13B, leaves it 95% alone for 24 to 36 months.
#10. Stripe accelerating to 41% rewrites the public software leaderboard
Stripe’s letter reported growth accelerating to 41%, with billings up 71%, and a share count lower than it was three or four years ago because they’ve been buying back stock.
My read: at Stripe’s scale it’s become a derivative of AI, the way a chip manufacturer is. You practically have to argue with an agent to get it not to use Stripe. And when OpenAI and Databricks go public at 80% growth, and Stripe is compounding at 41% with 71% billings growth, nothing in public software except Palantir approaches it. Not even Cloudflare. Almost everything below that line reads as a distant memory.
Rory’s version: if these were public, they’d be so far up the rankings that everyone else gets pushed down. He’d rather get it over with, because then you could actually figure out whether $300M to $400M revenue companies can exist in the public markets. Right now they’re the unattainables, which is why so many public investors are doing crossovers.
The Stripe letter also has the most useful line of the week: intelligence is like capital. It’s fungible, there’s demand for it, and it has to be managed and allocated.
#11. Token addiction is 2027’s CFO problem
My read: we don’t realize how addicted we are. This year started with token-maxing and performative AI, where spending more tokens made you a better employee. Then the $20,000-per-employee bills landed and everyone whiplashed into budget management, caps at $200 or $500 for non-engineers and $10,000 for engineers. The back half of this year is budget management. Next year is the backlash, and the backlash runs the other direction. I see it in my best portfolio companies. People cannot go back. I need ten sub-agents running 24 hours a day to do my job, or I quit.
Rory’s harder version, via the Stripe letter: seat-based B2B was a one-time allocation decision. You bought five seats, you were done, there was no follow-on work. Uncapped intelligence is not that. You have spending controls on money, but you can’t tell employees not to spend money, because money is the lifeblood of the business. Intelligence gets managed the same way, which means pricing it and allocating it per person. Give it to the person who’s wildly productive. Cut off the person asking dumb questions.
If you’re a mainstream US corporate, the token spend has to come out of something. You cannot introduce automation and tell Wall Street the net result of your wild new productivity is EPS down 10%. So someone has to decide which three of the ten people you no longer need.
The counterweight I raised: retention is what CFOs are actually talking about. Empowered CFOs are terrified that their stock is down and the AI leaders have so much stock comp that their best people get vacuumed up. If you don’t give people what they need for AI, you don’t just lose 30% of the company, you lose 90% of the ones who matter, and you’re left with the AI skeptics. The CFO is caught between an uncapped token line and a talent base that will walk without it.
#12. Agents still cannot be trusted, and the 2.0 generation hasn’t fixed it
Instinct, the stealth-launching next-generation personal agent a lot of VCs were posting about, got attention for data security problems around giving it access to everything. Grokbot leaked information. OpenClaw leaked information before that.
My read: none of this should surprise anyone working with agents daily, and the striking part is that a year of better guardrails and better harnesses hasn’t solved it. It wasn’t solved with OpenClaw, it isn’t solved with Grokbot, it isn’t solved with Instinct. Open-weight models have fewer guardrails, which adds a vector rather than removing one. These are goal-seeking probabilistic systems, and they make mistakes with your data the way a junior engineer or a personal assistant would, except a thousand times more often. When I was running the Mac Mini setup, it tried to buy six AP watches for the team. $360,000. It just didn’t complete. I would happily invest at $600 pre in the company that actually solves this. It has to actually solve it.
Harry’s counter: remember when we’d never put our credit cards online. We will trust agents with credit cards, financial data and passwords. This feels inevitable.
Rory’s question: is an individual’s idiosyncratic workload the right place to spend agentic budget, versus boring corporate work like loan processing where there’s less discretion, more expense, and an actual budget line? Silicon Valley falls in love with personal productivity tools because everyone here is hyper-productive. Ben Thompson’s line covers it: Silicon Valley forgets every three years that the average American is not trying to be efficient. Nobody wakes up wanting to grind down a to-do list. The category is real. It’s also historically nichy and hard to get right.
#13. The dumbest category we’re funding right now
We each named one.
Harry: customer support. A huge amount of money burns there, because support doesn’t exist as a unique surface anymore. One or two players take a large share rather than the market staying distributed, and every sophisticated technology company he knows is building its own.
Rory: defense, not because the products aren’t needed, but because venture returns at the margin will be hard. That business requires account control, and the two or three largest players end up scooping up the rest over the next two decades. It takes a portfolio of products to survive the interaction with the Pentagon, not a single killer product.
Both of them, and I agree: humanoid robotics. Rory’s on the board of Locus with 15,000 robots in the field, and that works because it’s a specific-purpose robot. On the humanoid running a 100-meter dash faster than Usain Bolt: if you want a machine to cover 100 meters quickly, buy a Tesla.
Mine: venture money poured into accounting firms and law firms expecting them to turn into the next Mercor. The sister-company structures where some of the people hold ownership are too convoluted. It makes too much sense on a spreadsheet. I’m still waiting to see the $20B outcome that comes from converting a room full of Ivy League grads working 100-hour weeks into AI-driven services.
Rory’s correction on all of it: in every one of these categories, he still finds himself looking at an individual founder and thinking maybe this is the deal that acknowledges the issue and transcends it.
Quotable Moments
Jason Lemkin
“$9 billion doesn’t clear the bar for seed investing in 2026.”
“I need my ten sub-agents running 24 hours a day to do my job, or I quit. Take away my agents, I quit. I won’t do the job.”
“If 30% of my company leaves to go work for Harvey, I’m dead in the water.”
Rory O’Driscoll
“If you get 15x on your failures in venture, you’ll die a rich man.”
“Sometimes when you look back at outcomes, you realize only one sentence matters. Today, what Anthropic is, is: it’s all about code. That’s the only sentence that matters.”
“You can’t introduce automation and say the net result of automation is reduced profits.”
Harry Stebbings
“Neo labs are just going out of favour, and next-generation model providers too. That’s universal from everyone I speak to.”
“To most of the general population in large majorities of the world, AI is ChatGPT. I am in awe that Sam is not going: we are the next Google.”
“Do you guys remember when it was ‘we’ll never put our credit cards online’? It was unthinkable. This feels inevitable.”
This post is part of the ongoing 20VC x SaaStr collaboration with Harry Stebbings and Rory O’Driscoll. The Jason’s Takes companion posts Sunday.
