Newsletter · · Ashutosh Agarwal

Anthropic's Leaked IPO Financials Put Its 2 Trillion Dollar Price on Trial - The VC Read - Week of October 1, 2026

The VC Read for the week of September 24 to October 1, 2026. Podcast synthesis on the leaked draft of Anthropic's S-1, with 4.6 billion dollars of 2025 revenue, a 42 billion dollar net loss and 518 billion dollars of compute commitments, Paul Kedrosky's earnings-before-bad-things critique, Chamath Palihapitiya's call for a sub-2-trillion IPO, the bulls' case for business quality, plus OpenAI's new raise, Oura's pulled IPO, tripled seed prices and the end of the private credit boom.

The VC Read

Week of October 1, 2026: Earnings Before Bad Things

Anthropic's numbers finally leaked. Venture's most-anticipated IPO now has to survive its own spreadsheet. The week of September 24 – October 1, 2026.

For months, every venture podcast has argued about Anthropic's IPO without any numbers to argue over. This week the numbers arrived.

On Monday night, September 28, Reuters reported that it had seen a draft of Anthropic's IPO prospectus. That is the S-1, the long disclosure document a company files with the SEC before it sells stock to the public. To be precise about what happened: the S-1 has not been filed publicly. It is a confidential draft that Tech Brew Ride Home said had been "circulated with a small group of partners in recent days." Prof G Markets described it as a draft Reuters "obtained," and Monetary Matters called it "the new S1 that leaked." Still, it is the first real look anyone outside the company has had, and by Tuesday morning it was the lead story on almost every business podcast.

Here is what was reported, per Tech Brew Ride Home, "We Have The Anthropic IPO Numbers" (Sep 29, 2026) and Squawk on the Street, "9AM HOUR: Anthropic's IPO Prospectus, AI Summit at the White House, Oura Postpones IPO" (Sep 29, 2026):

  • Revenue: about $4.6 billion in 2025, up roughly 12x from the year before.
  • Operating loss: more than $8 billion. That is what the business lost running day to day.
  • Net loss: $42 billion. About $34 billion of that is a non-cash accounting charge, an on-paper revaluation of earlier financing deals. No cash actually left the building for it.
  • Compute: $7.33 billion spent on computing and infrastructure in 2025, three times 2024's figure and more than half of its $12.65 billion in total operating costs.
  • Commitments: $518 billion in future cloud and infrastructure obligations. Brew Markets reported that about 80% of that cannot be cancelled.
  • Customer concentration: close to a quarter of 2025 revenue came from just two customers.
  • Risk factors: nearly a third of the filing, including warnings that the technology may pose "catastrophic or existential risks to humanity."
  • Price tag: a possible valuation above $2 trillion, against roughly $965 billion in its last private round in May. The listing date is likely after the November midterms.

Keep one thing in mind as you read the debate below. The $4.6 billion is a 2025 full-year figure. Anthropic's more recent numbers are much larger: The AI Daily Brief, citing the New York Times, said the company is on track for $100 billion in annualized revenue by year-end, up from $65 billion at the end of July. Much of the argument this week comes down to which of those two pictures you think the S-1 represents.

The Big Debate: Is Anthropic a $2 Trillion Company, or a $2 Trillion Story?

The Bear Case: "Finance as Theater"

The sharpest critic of the week was Paul Kedrosky on Prof G Markets, "Anthropic's Financials Revealed: The Losses Are Stunning" (Sep 30, 2026). His argument is about how Anthropic will ask you to read its own numbers.

The labs have been telling investors they are already profitable "on an inference-only basis." Inference means the cost of actually answering customer requests. What gets left out is the cost of training new models. Kedrosky's view is that training is not a one-off:

"This is the point where the finance as theater begins... It's this old joke we used to call it in my analyst days is earnings before bad things. If you let me get away with characterizing my earnings before bad things, my earnings look really good."

His point is simple. A frontier lab trains a new model every 6 to 18 months, so training is not like a building you pay for once and depreciate over decades. "The training costs are just the day-to-day running of the business," he said. When host Ed Elson concluded that training is what's being stripped out of the "profitable" figure, Kedrosky agreed: "I'll take that bet all day long."

He was careful not to overstate it. He called WeWork's infamous "community-adjusted EBITDA" "frankly fraudulent," and said Anthropic's framing is not that: "Is it fraudulent? No. Is it poor accounting? Yes. Should it be supported by the auditors? No. Should Wall Street punish them for it? Yes." He also pointed to a trap in the bull story: "Either training costs are integral to the business, or they're the thing you're going to cut so that you can be fantastically cash flow positive going forward... they're trying to have it both ways."

On customer concentration, he cited other leaks suggesting six customers make up about 60% of revenue. To use "the technical term," he called these "bananas numbers." The worry is not only that a few clients matter a lot. It is that those early clients (he named Meta as one of the biggest) are "wildly unrepresentative" of how a Goldman Sachs or an industrial company will use AI. That makes the next wave of customers a real unknown.

His verdict on the price: "It's a ridiculous price." He added, "Anyone who thinks that being the buyer at these kinds of prices in a very late-stage IPO of what amounts to a relatively mature company, when you look around the poker table and wonder who the sucker is, it's you."

Tech Brew's Brian McCullough raised the same concentration point more pointedly: "Who are those two customers that make up a quarter of Anthropic's revenue? I think we can be fairly confident one is Meta, right? But what happens to that when Meta builds out its own AI infrastructure enough to be able to eat its own dog food?" This matters because Meta spent the same week launching Muse, its own AI assistant, which hit No. 1 in the app stores.

The macro bear was Jeff Snider on Eurodollar University, "ALERT: Did Anthropic Just Pop AI Bubble?" (Sep 30, 2026). For him the real problem is the gap between the revenue and the commitments: "These people have 4.6 billion in revenue last year, and they've already committed half a trillion dollars. So you can see why this is a debt story." His rough math: suppose Anthropic raises about $100 billion in IPO equity and gets revenue to a generous $50 billion. That still leaves roughly $350 billion to be borrowed just to cover the commitments already known. Every quarter that revenue falls short of the hockey-stick forecast "adds a couple hundred billion more to your overall debt pile." His conclusion: "It's not that the AI bubble is priced to perfection. It is priced to ultra-perfection."

The All-In Twist: A Great Business That Will Have to IPO Cheaper

The most interesting take actually came three days before the leak, on All-In, "Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails" (Sep 26, 2026). Jason Calacanis framed it as a paradox: "The leading model companies need to go public to get capital to do their build out, and now they're delaying their own IPOs." He noted Polymarket odds of an Anthropic IPO in 2026 had dropped from a 96% peak to 76%.

David Sacks put the problem in the risk-factor section. "You have current leadership in the company saying that there's a greater than 10% chance of causing human extinction, and their own product is unsolved in this area... How do you IPO with your own management of the company saying that?" He called the gap between Dario Amodei's essay urging the industry to "pace the frontier" and the launch of a new frontier model, Claude 5.5, days later "corporate schizophrenia." He also flagged the governance setup. The founders own only about 2% each but would hold super-voting control, which "requires a tremendous degree of trust in those founders because they effectively can't be removed no matter what they do."

Chamath Palihapitiya then made what may be the most useful prediction of the week. He would keep Dario as CEO ("he's built essentially the greatest business ramp in all time"), but he expects the market to make Anthropic pay for its own disclosures:

"Even if you think that they're on $100 billion run rate, what would otherwise in a very clean sheet IPO maybe would have been a $2 trillion market cap. Now, you know, you're at one or less."

His logic: pension funds and mutual funds have duties to their own investors, so they will "justify all the risks by essentially asking for a lower and lower price." He sees that as healthy: "It forces them to just be a corporation and then run the business. And I think that they'll do great." He also offered a striking aside about Anthropic's grip on talent: "I've tried to recruit against Ant. I can't... I have not won a single bake-off."

David Friedberg identified what he thinks is the real threat, and it isn't doom. It's open source. "In the last 12 weeks alone, token use has flipped from 80-20 closed versus open to 80-20 open versus closed. I don't think we've ever seen a tidal wave of technology shift like this." (A "token" is the unit AI usage is metered and billed in, roughly a word fragment.) Friedberg's view is that the market is splitting in two. Premium closed models win the hardest work, like science, math and engineering, where buyers will pay "close to infinity dollars". Cheaper open models take most everyday enterprise work, including much of coding. Chamath turned that into the number every S-1 reader should look for: if 60–70% of a lab's tokens are doing tasks open-source models could handle, "you have to discount that revenue. If the answer is only 10%... they're going to be fine."

The Bull Case: Confidence, Not Stress

The bulls had good arguments too.

On 20VC, "Meta's Muse Hits No. 1. ChatGPT Finally Has a Rival | Menlo Sounds the AI Bubble Alarm | Factory Triples Its Valuation to $5 Billion..." (Sep 24, 2026), Rory O'Driscoll read the slip from October to November as a sign of strength. Anthropic had "an amazing Q2. They finally became bigger than OpenAI." Bankers would rather go out with a clean, audited quarter in hand than spend a roadshow explaining numbers they can't yet share. "Clearly, the fact that they didn't take that approach means they decided it's not hell or high water. They're confident... willing to take another month of timing risk probably for some significant valuation pop." His odds: "90% certain that this is a good decision," with "always that 10% chance the world goes crazy and you look back and you go, damn, should have taken the 100 billion."

He also dismissed the fear that scary risk disclosures could sink the deal: "The most important part of securities laws is not that you sell stock that has no risk. It's that you sell stock where you disclose the risk." Anyone buying the stock who doesn't know management thinks AI is dangerous "hasn't been reading their tweets." On the idea that no insurer will cover the labs: "This is a $2 trillion market cap company. It can self-insure."

On CNBC, Jim Cramer (Squawk on the Street, Sep 29) said the 2025 figures are old news: "We don't know enough because this is 2025 and maybe if it grows even faster, it's obviously better, maybe that's $2 trillion." He thinks the IPO is imminent because "people just they want AI plays that are other than NVIDIA."

The broadest bull case came from a16z's markets team on The a16z Show, "The $1 Trillion AI Buildout | State of Markets" (Sep 30, 2026). Their argument is that this doesn't look like 2000. "Stocks are up about 20% while multiples are down about 20%," meaning prices are rising because earnings are rising, not because investors are paying more per dollar of profit. The S&P 500 trades below 20 times earnings. In the dot-com era, top stocks traded around 100 times. Hyperscaler capex (spending by Alphabet, Amazon, Meta, Microsoft and Oracle on data centers) is about $780 billion in 2026, up from $416 billion in 2025, and expected to pass $1 trillion a year from 2027. Microsoft, Google and Amazon together hold roughly $1.7 trillion in cloud backlog, meaning customer contracts already signed. One a16z partner pointed out that Sam Altman and Sarah Friar "got a lot of flack a year or so ago for their massive compute commitments... at this point, like everyone would say, they're incredibly prescient." In other words, the $518 billion that scares Snider looks like foresight to a16z.

Where This Leaves Us

Strip away the noise and the debate comes down to three numbers the public S-1 will have to show:

  1. Training cost as a share of revenue. If it is shrinking fast, the "inference profitable" story starts to hold up. If not, Kedrosky's "earnings before bad things" sticks.
  2. How much revenue is frontier-only work versus work open models could do. This is Chamath's 10% versus 60–70% test.
  3. How many top customers are labs' own competitors. If Meta is one of the two, Muse makes that revenue look less durable.

Our read: the bears have the better argument on price and the bulls have the better argument on business quality. Notably, nobody this week argued Anthropic is a bad company. Even Kedrosky expects it is cash-flow positive on inference. The real question is whether public investors will pay private-market prices for it. Chamath's "one [trillion] or less" was the most specific prediction anyone made, and it is the one to check when this prices.

Signals

OpenAI won't IPO. It's raising again instead. Bloomberg Tech reported OpenAI is seeking at least $30 billion at a $1.4 trillion valuation, with "no clear timetable" for an IPO and possibly a bridge round. A bridge round is interim money raised to carry a company to its next big event. CFO Sarah Friar framed it as wanting "maximum optionality": "We could use more compute if we had it today. We could use more compute in 2027." Bloomberg's Seth Figerman put the scale in context: OpenAI "raised $122 billion in March, which exceeds pretty much any IPO that we've ever seen." On 20VC, the hosts discussed reports that OpenAI forecasts $278 billion of net burn by 2030 against $122 billion of cash, with total capex, much of it on partners' balance sheets, near $700 billion. Jason Lemkin's reply: "It doesn't have a history of the burn coming in less than plan. They may need 400." Rory O'Driscoll's summary: "Intelligence is not cheap." (Bloomberg Tech, "OpenAI Seeks $30B as Trump Backs AI Safety Audits" (Sep 30, 2026); 20VC, "Meta's Muse Hits No. 1..." (Sep 24, 2026))

The IPO window is open for some companies but not others. Oura just walked away. Smart-ring maker Oura postponed its IPO, citing market uncertainty. That is despite being profitable, with revenue expected to grow 90% this year and 5.7 million users. It had been looking to price at the top of its range, roughly a $15 billion valuation. CEO Tom Hale said Oura "has the luxury of choosing its moment," and the hosts noted "they are not choosing this one." Their read: investors may be "spooked by the long line of private capital that might be looking for an exit." The backdrop matters too. Prof G Markets noted the 30-year Treasury yield hit its highest level since 2002 and the 10-year was climbing toward 5.3%. (Brew Markets, "Oura Ring Pauses IPO & Netflix's 'Worrying' Engagement" (Sep 29, 2026))

Exits are back, but only at the very top. On This Week in Startups, Dave McClure (Practical VC) said his firm has "sold about $1.5 billion worth of stuff in the last three, four months, which hasn't happened for a while," after "a pause for five years." He described a split market. For "the top 30 names" (Anthropic, Anduril, Stripe, Databricks) "it's a seller's market... they can set the price." Outside that group, companies "doing a couple hundred million in revenue, growing 40%, 50%, profitable" may have no market at all. One CEO with over half a billion in revenue told McClure he won't go public "until they're doing $10 billion in revenue." The tool McClure is pushing for funds that need to show LPs cash returns is what he calls a "single asset continuity SPV." The idea is selling a slice of a winning position into a new vehicle to "generate DPI," or distributions to paid-in capital, the actual cash returned to investors. (This Week in Startups, "VCs Would Bet on Open-Source AI Over OpenAI and Anthropic | E2341" (Sep 23, 2026; one day before this week's window))

Seed prices have more than tripled, and the fund math is getting harder. That same episode showed cap-table data on seed valuations. The median went from $8 million in 2017 to $28 million now, the 75th percentile from $12 million to $54 million, and the 95th percentile from $25 million to $209 million. Jeff Clavier of Uncork laid out what that means for a $300 million fund. To return half the fund on one deal ($150 million) from a $5–6 million check, "we need to make a 30x... Like right now, unless you sell something for $10 billion, you barely sort of make money, which is insane." His firm's response is uncomfortable: "continue investing at higher and higher valuations. And we don't want that, but that's what we have to do in order to win those deals." The only deals he loses are to multistage funds writing "10 or $15 million seed check[s] at 100 million or more valuation." Jenny Fielding added that Series A and B fundraising has become "totally unpredictable." (This Week in Startups E2341 (Sep 23, 2026))

Instinct at $10 billion, four months after a $50 million seed. On 20VC, Rory O'Driscoll offered a margin-of-safety test for chasing the hottest names: "You should do the Seed and Instinct all day, every day, but you should think long and hard at 10 billion four months later." At $50 million pre-money with a world-class team, "your margin of safety... is infinite." At $10 billion, "maybe you get a 1x, but there's risk." Lemkin answered with a pragmatic point (it might still work "because... OpenAI might buy it at 50") and, earlier in the episode, the line that explains a lot of today's pricing: "It's other people's money. You got to put it to work." (20VC, "Meta's Muse Hits No. 1..." (Sep 24, 2026))

Chasing returns: foolish in public markets, rational in venture. In the same episode, O'Driscoll pointed out that in mutual funds, last year's winners rarely repeat. "Persistence is almost nothing." In venture, "persistence is not infinite, but it's quite high... because you get these increasing returns to success." So LPs piling into last cycle's top seed managers is "not crazy," even though, as Lemkin noted, it makes being an LP harder because "you will screen out" the next great unknown manager. (20VC (Sep 24, 2026))

Coding is "the mother lode," and enterprises don't want to buy it from the labs. Fresh from Dreamforce, Lemkin said C-level executives "don't trust Anthropic and OpenAI with their data. They don't trust it. This is not something manufactured on X." That is why he backed Factory's new round, which tripled its valuation to $5 billion on $200 million raised. O'Driscoll agreed: coding is "10X everything else in terms of value being created from AI today," and corporations will want their coding tools "from someone who's not also selling you the model." With Cursor "swooped off the table," he sees Factory and Cognition as the two independents still selling hard to large enterprises. (20VC (Sep 24, 2026))

Is SaaS dead? a16z says the data shows a "renaissance." a16z's markets team cited Stripe's data on its SaaS customers showing growth accelerating into 2026 across both young and mature businesses, which Stripe calls "the renaissance." Their framework for who wins and who loses: does AI increase the customer's need for the product (security and observability, which have "really stood out") or eat the workflow (horizontal apps)? Vertical software "has generally held up much better." They also put concentration in perspective. The top six private companies (Anthropic, OpenAI, Databricks, Stripe, Waymo, Revolut) are worth about $2.4 trillion at their last rounds, more than the $1.7 trillion combined market cap of every US IPO in the past decade excluding SpaceX. (The a16z Show, "The $1 Trillion AI Buildout | State of Markets" (Sep 30, 2026))

The SaaS debate is also a credit problem. James Elbaor of Marlton LLC told Monetary Matters flatly: "The private credit boom is over... There's no new money, no new capital flowing into that strategy right now." (Private credit means loans made directly by investment funds rather than banks, often to companies owned by private equity firms.) Part of his reasoning: "The estimated exposure to SaaS by private credit is over half a trillion dollars. And not every single one of those SaaS companies is going to make it." He described the core flaw as "a classic duration risk issue, similar to like a bank where you're borrowing short and you're lending long." In plain English, the funds promised investors their money back faster than the underlying loans pay off. The evidence: Blue Owl stopped quarterly redemptions at one fund (OBDC) to wind it down, and a BlueRock vehicle (BPRE) traded at a 38% discount to its stated asset value the moment it listed. "That window is absolutely closed." He doesn't expect catastrophic defaults, but does expect them to be "higher than what other institutions are pricing." (Monetary Matters, "The Private Credit Boom is Over: Redemption Requests Exceed Liquidity | James Elbaor | Marlton LLC" (Sep 27, 2026))

The retail side door keeps getting louder. Prof G Markets' Anthropic episode opened with an ad for VCX, "a public ticker for private tech," pitched on the idea that "our most innovative companies are staying private longer, which means everyday Americans are missing out." The products letting retail investors into late-stage private companies now sponsor the shows covering those companies' IPOs. That's worth noticing in the same week Kedrosky asked who the sucker at the table is. (Prof G Markets (Sep 30, 2026))

Quote of the Week

"When you look around the poker table and wonder who the sucker is, it's you."

Paul Kedrosky, on buying Anthropic at $2 trillion in a late-stage IPO (Prof G Markets, "Anthropic's Financials Revealed: The Losses Are Stunning," Sep 30, 2026).

Runner-up:

"Even if you think that they're on $100 billion run rate, what would otherwise in a very clean sheet IPO maybe would have been a $2 trillion market cap. Now, you know, you're at one or less."

Chamath Palihapitiya (All-In, Sep 26, 2026).