Benchmark’s Jack Altman joins Harry Stebbings, Rory O’Driscoll and Jason Lemkin: AMD buys World Labs for $8.2B, Meta hires away MongoDB’s CEO, a mock Investment Committee on Jev at $10B, and Oura pulls its IPO two days before pricing
This week’s episode had a guest in the fourth chair: Jack Altman, General Partner at Benchmark, whose firm just co-led Instinct’s $1 billion round. That gave the group someone who could explain from the inside why one of venture’s most disciplined early-stage firms is now writing ten-figure growth checks into companies that launched in August. The conversation ran from Anthropic’s leaked S-1 to the collapse of the traditional seed round to whether open-weight models have already peaked. Below are the takeaways, followed by three quotable moments from each speaker.
#1. Anthropic’s Draft S-1: $4.6B Revenue, an $8B Operating Loss, $518B in Compute Commitments
Anthropic’s draft S-1 leaked this week. The headline numbers: $4.6 billion of 2025 revenue, an $8 billion operating loss, and $518 billion in cloud, computing, and infrastructure commitments in the coming years, according to the prospectus reviewed by Reuters. There are also reports that Anthropic is delaying its IPO, likely until after the US midterm elections.
Rory’s read: there’s almost nothing actionable in the leak. Most of the scary-looking net loss is non-cash accounting. The rest is what everyone expected: about $4.5 billion in revenue, $8 billion in compute, $5 billion in other expenses. The one genuinely new data point is that two customers drove 25% of revenue, meaning someone spent roughly half a billion dollars on Anthropic last year. Rory’s point is that 2025 is stale, everyone knows Q1 and Q2 of 2026 were huge, and the Q3 revenue number is 90% of the data you need to price the IPO.
Jason’s worry is optics, not fundamentals. The S-1 reads negative to the general public, from a risk factor about existential risk to humanity to the jaw-dropping losses. He sees a Facebook-IPO setup: oversubscribed at pricing, then a drift below the IPO price a month or two later as anti-AI sentiment, data center backlash, and default noise around compute deals pile up.
Rory pushed back on the direct impact: most of America may dislike data centers, but they’re not the ones running Fidelity’s growth funds. The indirect risk is real, though. When your own legal document says the product may be unsafe, every state attorney general now has a hook to sue. This will be the most publicly participatory IPO in memory.
Jack’s counter: public investors may end up more long-term oriented than private ones. Private investors are hand-wringing month to month over the Anthropic vs. OpenAI narrative. Public holders will think in quarters, maybe years, and focus on market structure.
On the separate news that Anthropic’s founders are moving to lock in 50.1% voting control, Rory’s explanation was mechanical: founders often have effective control in the preferred-stock structure pre-IPO, and it evaporates when everything converts to common. Jason’s take was simpler: if we’re trusting them not to blow up the world, we can trust them with the votes.
#2. Instinct Raises $1B at $10B, 33 Days After Raising at $2.5B
Sequoia Capital, Benchmark Capital and Coatue participated in the financing round, and the company said last month it was fundraising at a $2.5 billion valuation. Instinct only launched its invite-only service in August 2026, and founder Noah Shinn says the platform is approaching a billion dollars in annual transactions, with more than 50% of transactions on the platform for travel.
Jack’s thesis (Benchmark co-led): consumer agents that act across the entire third-party internet are a paradigm shift on the level of chat and coding. AI moves from something you talk to into something that does things for you. And when a category is that important, multiple players win, just like coding produced Cursor and Cognition alongside the labs. He gave real credit to Meta’s Muse as a great product that reinvigorated the company.
Jason’s test for any agent category: do we use it all day long? He runs coding agents 10 hours a day. Legal AI works because lawyers run it constantly. OpenClaw was generation one, which nobody could safely run. Instinct and Muse are generation two. The open question is whether either becomes an eight-hours-a-day app.
Jack agreed it’s early, and pointed to how Harvey and Legora reference calls sounded at $1 million ARR (“the thing barely works”) versus today (“I run my whole life out of it”) at hundreds of millions in ARR.
Harry’s admission: he used to be a heavy ChatGPT user and now runs his life through Instinct. Jack’s view on cannibalization: the form factor takes share from chat because screen time is finite, whether or not the labs build it themselves. He also flagged how short the windows are now. Cursor had a long runway before the labs caught up. Instinct’s window is much shorter.
#3. Early-Stage Risk at Growth-Stage Prices: How Benchmark Sized It
Rory’s frame: this cycle has produced investments with early-stage risk that require super late-stage capital. Benchmark didn’t really launch a growth strategy. It expanded its early-stage strategy because early-stage now requires bigger checks.
Jack confirmed Benchmark invested at both $2.5 billion and $10 billion, and internally treated it as an early-stage bet. Between now and Christmas there will be three more evolutions of the category, so you don’t pretend to underwrite it like a public company with $2.5 billion of ARR.
On sizing, Jack’s two constraints: enough shots on goal so the fund has multiple chances at something great (the floor on diversification), and enough dollars to matter in the round given the valuation (the ceiling on ownership). Rory framed it as Kelly betting: edge over odds, then deciding whether you have the stomach for full Kelly.
When Harry said he’d put 5% of his fund into Instinct at $10 billion, Rory pointed out that if you truly believe the risk profile Harry described (50x upside, roughly 1x downside via acquisition), full Kelly says closer to 30% of the fund. Harry’s response was the real-world constraint: at his stage of career, a 30% position would end his LP relationships.
Then Jack made the line of the episode: in a market where valuations are extraordinarily high and outcomes are also extraordinarily high, every firm is investing either way too fast or way too slow. When both sides of the equation are this out of whack, the odds of getting the speed right are zero.
Rory added a story from his firm’s annual meeting: economist Tyler Cowen, asked about venture’s future, said returns will be highly skewed, variance will go up with AI, and many of you will fail. Then moved on to the next question.
Harry’s concern isn’t this round. Instinct’s cohort data is almost certainly exceptional. What worries him is three rounds in three weeks with no material change in between.
#4. AMD Buys Fei-Fei Li’s World Labs for $8.2B in Stock
AMD will acquire Fei-Fei Li’s World Labs for $8.2B, the first big exit for a neolab, about two and a half years after founding.
Rory started with the founder. Fei-Fei Li built ImageNet, the project whose 2012 competition winner first showed deep learning would blow past everything else. She didn’t monetize that, worked at Stanford and Google, then started World Labs mid-career and nailed it. Another immigrant success story that started from nothing.
Jason’s math on AMD: the stock is up 279% this year to roughly $1 trillion. Spending about 0.8% of market cap on a world-class team to keep that run going and stay relevant against Nvidia is cheap. If AMD were up 3% this year, it wouldn’t be spending $8 billion.
Rory admitted he’d been nervous about world-model companies because they lacked the obvious path to revenue that chat and coding gave OpenAI and Anthropic. In retrospect, every big foundation model company is likely in the market to acquire a robotics or world-model team.
Harry cited a report from his partner tallying 102 neolabs that have raised over $70 billion. Ten or twelve exits is fine. 102 is a different question.
Jack’s answer: there are now roughly 10 companies that can do $10 billion acquisitions and want to. That’s never been true before, and it’s faster and easier than going public. He’s hesitant to be too skeptical given what’s happening. Jason added that credentials matter more at this layer: a world-model company needs proven technical leadership both to deliver and to be the acquisition candidate everyone wants.
#5. The $3-6M Seed Check Has Mostly Disappeared
Harry opened with a top CIO’s challenge: can you really play in venture with less than a $1 billion fund now? His own team told him they couldn’t find anything under $100 million. Not valuation. Round size.
Rory (running a $900 million fund) pushed back: there are two kinds of companies. Neolabs and semiconductor companies need hundreds of millions just to ship. But the companies building on top of those models are often shipping product on $3 million, then raising $50 million because they can.
Jason’s view: venture prices have inflated roughly 2.5x since 2010, but three or four people with free token credits sharing an $8,000-a-month apartment can still get 18 months down the road on $2-3 million. That’s still the natural seed round when nobody is throwing money at you.
Jack drew the sharpest line. Two forces killed the traditional seed round: labs that can’t do anything for less than $200 million, and well-networked founders who skip the “$6 million at $40 million” round entirely and go straight to a $50 million raise. Writing $3-6 million checks for 8-15% ownership is, in his words, broken in many lanes. But slow compounders still exist: a company raising $3 million at $30 million today that quietly compounds selling to police departments, fire departments or libraries until 2041.
Harry called that an anomaly, not an industry. Rory estimated the table is 80% fast-action and 20% slow, and said a compounding play only makes sense if it comes with low risk. Jason’s objection from the boardroom: quiet compounders aren’t stable anymore. At one portfolio company north of $100 million in revenue, the last board meeting’s competitor slide showed nine competitors, eight of which he’d never heard of.
#6. Meta Hires MongoDB’s CEO and the Stock Drops 18%
MongoDB shares fell nearly 20% in morning trading Monday after Chief Executive Chirantan CJ Desai left the database company to lead Meta’s new enterprise AI business. Meta says it will focus on bringing its full technology stack, including Muse, Meta Business Agent, Muse API, Muse Code, and more to businesses and developers. MongoDB brought back Dev Ittycheria as interim CEO and reaffirmed guidance.
Rory’s frame: a public company CEO less than a year into the job quit the top seat to run a division. Former CEOs almost never go back to not being CEO. The only explanation is an earth-shattering offer, and that’s what momentum lets you do: take whatever talent you want.
Jason put it bluntly: if the offer is 10x your current package, you hit the bid.
Jack said it’s not only the money. It’s the attention, the zeitgeist, the product everyone’s family is talking about. And he found the zoomed-out positive: talent is unusually unstuck right now. Great people are actually moving to the most important opportunities, which is good for society even when individual moves look strange. Jason added that California’s lack of garden leave makes that liquidity possible.
#7. Investment Committee: Jev at $10B, Two Weeks After a $200M Seed
TypeSafe AI’s Jev model launched on September 15, and the company had raised just $40 million in seed funding led by DCVC, a round that valued the company at $200 million. Now TypeSafe AI is in talks to raise more than $1 billion at a valuation above $10 billion. Jev isn’t a chat model: if you input a set of data and a group of structured questions, it returns typed answers and calibrated probability values, priced at approximately $0.042 per million input tokens, with output tokens provided at no cost.
Jason’s pitch (with tongue partly in cheek): do it, and go 20-30% of the fund. He cited Jev at 17% of traffic on OpenRouter and 20% through Vercel’s router, at a 70th of the price and 100x faster. Frontier model costs at 10-12 hours a day of agent usage aren’t sustainable, even as Sonnet and Opus get cheaper. Downside protection is a team of ex-OpenAI researchers someone will buy no matter what.
Rory’s math said the joke isn’t crazy. Take roughly $100 billion of current model spend, assume 20% is relevant to Jev’s use cases, compress it 5:1, and you get about $4 billion of accessible revenue. Developer adoption has been lightning fast, so a billion-dollar revenue line by saving customers 80 cents on the dollar is credible. His only note: spend 30% of the fund every week and you’re out of business in three weeks.
#8. Modal Hits $15B, Baseten Eyes $26B, and the Open Source Peak Debate
Modal tripled its valuation to $15 billion and Baseten is in talks at $26 billion.
Jack’s read: there was a period when the right trade was to keep buying the labs, every round, even when it looked expensive. For the last 18 months, the right trade has been to keep buying inference: Modal, Baseten, Fireworks, Fal, Together. His partner Eric Vishria’s line: “It’s all going to work.” Inference has been an index bet on everything outside the labs.
Jack’s mechanism: costs aren’t sustainable, and more tasks are hitting intelligence saturation. Once your tax return is filed correctly, more intelligence doesn’t help. A gold-plated $30,000 hammer doesn’t drive the nail better. That pushes workloads to open source and inference clouds. But the labs remain heavily cost-advantaged through compute access, user access and the ability to subsidize.
Jason’s contrarian call: open-weight market share has peaked. Two reasons. First, Anthropic and OpenAI can price their non-frontier models wherever they want; Sonnet 5.5 came out about 20% cheaper in one week, and they can go further. Second, at Dreamforce he couldn’t find a single enterprise customer comfortable running open-weight models, most of which are China-origin today.
Rory pressed on whether the objection is to open weights or Chinese open weights, which is why a US-based low-cost alternative like the Poolside/Nvidia effort matters. He also framed the lab’s real trade-off: every marginal GPU is a choice between training a model for biology and serving someone’s tax return.
Jack’s closing frame: enterprises will increasingly post-train their own models on open weights. But zoom out and it comes down to who owns the compute. All the inference clouds combined are on the order of a gigawatt; OpenAI and Anthropic are each in the high single digits. Rory agreed: compute share roughly proxies token share, within about 2x of revenue share.
#9. OpenAI Reopens Its $200 Plan at Half the Usage, and Jason’s Sonnet 5.5 Token Data
OpenAI reopened its $200 plan, which it had paused after running out of compute for its latest model, but halved what $200 buys.
Jack’s point: the core equation has variables nobody can pin down. How many tokens does a task need as models get smarter? What’s the utility per token? More tokens are consumed at fewer dollars each, but no one knows how many it takes to file the tax return.
Rory tried to model it and gave up: you’re multiplying three numbers, each with a large error bar. The only actionable signal is buying behavior. Some customer spent half a billion dollars on Anthropic last year; assume they ran the numbers.
Jason’s real data: he ran his own evals for SaaStr Connect on Sonnet 5.5. Input tokens were 42% higher than the prior model and output tokens 44% higher. Quality went up and it passed more of his blind tests. Net cost went down about 10%, not the 30% the price cut implied. Rory’s takeaway: if it’s that hard to predict for one person’s specific task, it’s 10-100x harder in aggregate, so watch what the buyers do.
#10. Oura Pulls Its IPO Two Days Before Pricing
Oura had marketed a share sale of 50 million priced in an indicated range of $40 to $44 apiece, a deal that would have raised $2.2 billion at a $15.62 billion valuation. Just last week, Bloomberg reported that investors were clamoring for shares, with the offering four times oversubscribed. Oura cited IPO market uncertainty, and CEO Tom Hale said “we have the luxury of choosing our moment.”
Rory (who has an ownership interest but no inside information) was genuinely surprised. Oura had Morgan Stanley, Goldman and JPMorgan, says it’s profitable, and has a brand consumers know. His best explanation: Forerunner had said up front it would sell its entire position in the IPO, which he’d never seen before. Heavy secondary makes investors extremely price-sensitive. A company taking 10% dilution doesn’t care much about leaving money on the table. An investor selling everything at $18 instead of $22 just cut its entire return by 20%.
Jason didn’t buy that Tom Hale pulled the IPO for Forerunner’s pricing. He pointed to the backdrop: per Dan Primack, the S&P was 1.4% off its all-time high and the Shiller PE at a record. Lots of M&A liquidity, yet you can’t get an IPO done.
Rory’s bobsled analogy: once the S-1 goes public, you’re in the sled sliding to the bottom. Pulling a night or two before is the hardest move there is. It’s survivable here because Oura is consumer and already profitable. Harry noted employees had a $534 million tender a couple of months ago, which softens the blow.
Jack’s forward look: no AI-native application company has gone public yet. Many could go tomorrow; it’s a question of price. Nobody wants to be first, or to go before the labs do. If the market holds, expect quite a few in 2027.
#11. Nubank in Talks to Buy Monzo for $8-12B
Nubank is reportedly in talks to acquire UK neobank Monzo in a range of $8 to $12 billion. Nubank’s stock was down 12.6% over the week, to a market cap of about $47 billion.
Harry was surprised. Nubank is focused on winning the US, and Monzo is a strong asset, but in the UK. Rory’s reframe: it’s far more surprising that Nubank wants to buy than that Monzo wants to sell. Monzo is too small to matter in US public markets, and European public markets don’t reward fintech.
The steelman from the pod: the US banking market is too efficient to leave fat margins for neobanks, while Brazil and Europe are full of overcharging incumbents. Nubank may simply be targeting less efficient markets, though in the UK it runs straight into Revolut.
Jason’s general point: acquirers buy time, not just revenue or customers. That’s essential for venture to work, because it’s the only way A-minus and B-plus assets get bought. Rory flagged the board drama: the founder had stepped back, a hired CEO turned the business around, then the chairman replaced him and investors pushed back. The UK-style non-exec chairman model makes sense for mature public companies and none for venture-backed ones. When you have that much instability and someone offers to buy you, you hit the bid.
#12. Bessemer Raises $5.75B While NFX Goes GP-Capital Only
Bessemer raised $5.75 billion, including a $1.75 billion seed fund. NFX went the opposite direction, investing only GP capital and taking no new LP money.
Jason’s read on Bessemer: it’s the same lesson Benchmark learned. Bessemer’s growth team scaled up and won. And if a seed round is now $30 million, a $1.75 billion seed fund with reserves isn’t large. Even $1 billion starts to sound small.
Jack said both moves make sense. Bessemer has decades of LP trust, so why not go bigger. Homebrew did the NFX move before and it worked. Investing only your own money changes the texture of the job: no explaining yourself to anyone, and the freedom to write a $100K check that’s hard to defend from a seed fund. But firm-building needs outside capital for salaries, careers and partnerships. And it matters who you’re making money for: hospitals and endowments, not one principal’s family office.
Rory pointed to Peter Thiel funding roughly 30% of his own fund: being the largest LP lets you tell LPs you’re comfortable putting 20% into SpaceX in 2008.
Jason said he’s considered going direct, and the Homebrew founders recommended it when he started. The constraint is check size. He wants to write $5 million checks to be relevant, which is too much balance sheet risk alone, and he doesn’t enjoy 60-second $100K checks.
Quotable Moments
Jack Altman
“We’re definitely not investing at the right speed. We are either investing way too fast or way too slow. When both sides of the equation are this out of whack, the odds of having it right are zero.”
“There are like 10 companies that can do $10 billion acquisitions and want to. That’s just so different. It’s much easier than going public, and it’s quicker.”
“There was a period where the correct answer was just keep buying the labs. In the last 18 months, the correct answer was just keep buying inference. Inference has been a really great way to get an index bet on everything outside the labs.”
Rory O’Driscoll
“In this cycle we’ve had investments with early-stage risk that require super late-stage capital, which is just definitionally a strange time to be playing.”
“Money is a signal. Price is a signal. And price is sending a signal: everybody go right here. And everyone will go right here, because that’s the job of price.”
“70% of America might think data centers suck, but they’re not running Fidelity Growth. A small number of highly compensated managers are the buyers of the stock.”
Harry Stebbings
“This is not the round that worries me. What worries me is when you have three rounds in three weeks with no material movement in between and no data suggesting anything is different.”
“My team said to me the other day, we can’t find anything under 100 million. I said, wow, seed prices are expensive. They said, no, no. $100 million round size.”
“If you want to go for that model, your numbers will be crap for quite a long time. We always forget that we’re in an opportunity cost game.”
Jason Lemkin
“The ones that win are the ones we use all day long. Will we run Instinct or Muse eight hours a day? If we do, I guarantee it wins.”
“No one wants to run on open-weight models on the floor at Dreamforce. Not a single customer I talked to is comfortable with it. I think open source has reached its maximum as a market share.”
“I ran my evals on Sonnet 5.5. Input tokens 42% higher than before, output tokens 44% higher. Quality went up. Cost went down about 10%, not 30%. This stuff is hard to predict.”
This post is part of the ongoing 20VC x SaaStr collaboration with Harry Stebbings and Rory O’Driscoll.
