Anthropic Opens Its Books To Wall Street As OpenAI’s $20 Billion “Gross-Net” Hole Hangs Over AI Trade

One number.

That is what an entire equity complex, by now worth tens of trillions across public and private markets, is waiting for today in San Francisco. As Bloomberg first reported, Anthropic invited a select group of institutional investors to its headquarters for a closed-door session with senior leadership ahead of what could be the largest IPO in history. According to CNBC, the company will use the meeting to release updated financials, and one figure matters far more than the rest: the revenue run rate.

“When word gets out about the revenue run rate, it will move the entire market,” said JonesTrading’s Mike O’Rourke. “The investment community wants to know whether Anthropic is going public from a position of strength or weakness.”

Goldman’s trading desk agrees. In this morning’s “What Matters Today” (available to pro subs), Goldman’s Nelson Armbrust boiled the setup down to one line:

Anthropic: investor event in San Fran today. Anthropic will release updated financials to a group of institutional investors gathering at its headquarters this week…. Watch for ARR numbers after OpenAI lowered theirs last week from $70bn to $50bn.

Over in London, Goldman TMT specialist Sean Johnstone (available to pro subs) called today “a test just how bearish the market has become on AI and semiconductors,” adding that “the other thing that will drive sentiment in the space today” is the Anthropic event in SF, “where there is fixation on its ARR.” JPMorgan’s Market Intelligence desk lists “headlines regarding frontier IPOs” alongside tomorrow’s TSMC print as “the biggest near-term drivers,” and earlier this week JPM’s TMT specialist Brian Heavey flagged a “lot of focus on the pre-IPO investor meeting in SF on Wed (which may be creating a bit of anxiety).”

The two markets that trade Anthropic in real time are already voting. Polymarket odds that Anthropic goes public by Nov 30 have slid from around 70% earlier this month to a coin flip, as we pointed out this morning:

Today Anthropic is meeting prospective investors on Oct. 14 in preparation for its initial public offering. Meanwhile odds of its IPO going live before the end of November are sliding. https://t.co/vxkjo9Y9Km

— zerohedge (@zerohedge) October 14, 2026

At the same time, the price of the thing Anthropic actually sells keeps falling. According to Ornn, the Anthropic Token Price Index just dropped to a new all-time low of $0.84. That is roughly half its mid-September level and about 64% below the mid-August spike, and we flagged it as “one reason for the skepticism”:

Put differently, investors are getting less sure the IPO happens on schedule just as the company’s unit price hits a record low. And that is before the run-rate number even lands. Anxiety may be an understatement. Here’s why.

The Last Time A Frontier Lab Showed Its Numbers…

Regular readers will recall what happened last Thursday, when the FT reported that OpenAI had told investors its annualized revenue was roughly $50 billion at the end of September, some $20 billion below the ~$70 billion figure that had been circulating for weeks. The Nasdaq promptly tumbled: the NDX fell 139bps, semis slid 3.4% and Goldman’s desk saw “a heavy wave of selling of Semis / AI / Megacap tech… ($1bn+ in net supply) following the midday OpenAI headlines.” Oracle dropped as much as 6%, and its 5Y CDS closed at a record 261bps.

By Friday morning came the damage control. Bloomberg reported that OpenAI now expects to “reach or exceed $70 billion” of annualized revenue by year-end and blamed the confusion on “differences in how OpenAI and Anthropic calculate comparable revenue,” which was enough to spark a tech rebound. As we put it at the time: “it wasn’t actual… it was projected.”

The Goldman TMT desk’s post-mortem (also available to pro subs) is worth quoting at length, because the same accounting question will hang over every number Anthropic shows today:

Gross vs net – the source of the gap: Anthropic reports closer to gross revenue (full customer spend through cloud partners, partner cut booked as a cost). OpenAI has reported closer to net (mainly its own share). When some investors tried to put the two on a like-for-like basis by grossing up OpenAI’s figures, they produced the higher ~$40bn (August) and then ~$70bn numbers. OpenAI’s own investor update used its more conservative framing and landed near $50bn at end-September.

The distinction has been out there for sometime but these are not listed entities… In theory most should know the difference BUT these names are not listed and many do not adjust… The episode highlights how sensitive the complex has become to the precise trajectory of two private companies that use different reporting conventions.

Translation: OpenAI’s underlying progression was “nearer $30bn to $50bn than $40bn to $70bn,” per Johnstone, which is still roughly 70% growth, though well short of what the market had priced. We charted the gap in last weekend’s look at neocloud rents:

Goldman’s Delta-One head Rich Privorotsky shrugged it off as “largely a question of accounting definitions,” but added the part that matters for today: “the reaction suggests more length and increasing sensitivity in these pockets.” And Byrne Hobart of The Diff explained why Anthropic’s numbers carry even more weight than OpenAI’s: “once Anthropic files an S-1, they’ll set a standard for what disclosure looks like, and more analysis of OpenAI will be explicitly benchmarked to how Anthropic reports things.” Note the irony: the gross basis that made Anthropic’s run rate look so much bigger than OpenAI’s is now the yardstick for everyone.

In other words, today Anthropic’s number gets judged against its own history and against the market’s patience with private-company math, both at once.

Seven-Fold In Seven Months

And that history is staggering. Anthropic’s run rate went from $87 million in January 2024 to $1 billion at the end of 2024, roughly $9 billion at the end of 2025, then $14 billion in February, $19 billion in March and $30 billion in April. By the end of July, Bloomberg reported that the company had told investors its annualized revenue had hit $65 billion, seven times where it ended 2025. Last month the NYT reported that Anthropic is expected to finish the year with more than $100 billion in annualized revenue.

Hence the problem. With a print above $65 billion already in the market for over two months, today’s figure has to show the curve is still compounding toward $100 billion, or the “strength” narrative cracks. Recall Brad Gerstner’s mid-September framework, which Goldman’s desk resurfaced last week: the top three labs (Anthropic, OpenAI and SpaceX/xAI) were at a combined ~$100 billion run rate in July and need to reach $180-200 billion by year-end “to keep the AI trade intact.” Anthropic, booked gross, is carrying the biggest share of that load.

The price tag assumes it can. Prospective buyers have floated a valuation of $1.8 trillion to $2 trillion, roughly double the $965 billion post-money of its May round. Reuters Breakingviews did the napkin math: take the $16 billion of revenue Anthropic booked in the first two quarters, hold it steady, and you get “just shy of $40 billion” for 2026, which puts a $2 trillion valuation (less ~$120 billion of net cash) at roughly 47x revenue, above even SpaceX’s ~42x. Independent research shop New Constructs went further, branding the deal “the most ridiculous IPO of 2026.”

Then again, if the year-end run rate really clears $100 billion, that multiple drops to about 19x in a hurry. That is exactly why Anthropic will want to talk about the run rate today, and leave trailing revenue for the footnotes.

The Bear Case Has Charts Too

So what could go wrong? According to Goldman’s Johnstone, plenty, and most of it has to do with the price of what the labs sell:

Competition and technology risks are adding to the uncertainty. New models from Meta and Google, cheaper open-weight Chinese models and falling token prices are increasing pressure on frontier labs and compressing pricing power at the model layer. Bears worry that weaker model economics could eventually reduce spending on high-end GPUs and ASICs…

The competition is visible in real time. Meta’s newly released Muse Wonder reportedly ranks behind only Claude Opus 5.5 on Artificial Analysis’ Intelligence Index, beats it on SWE-Marathon (57.5% vs 46.3%) and is priced at $4/$20 per million input/output tokens. And Google’s Gemini 4 Argon dropped this morning.

“Can’t remember a time when the frontier was quite so competitive,” Goldman’s Delta One head wrote overnight, noting that “you now have Anthropic, OpenAI, Google and Meta all genuinely in contention.” Prediction markets agree: in the Polymarket chart Privorotsky circulated on Monday (first published by ZeroHedge), Google had leapfrogged Anthropic as the favorite to have the best model at the end of 2026, 43% to 42%, after Anthropic spent most of the summer above 60%.

More competition means cheaper tokens, which is exactly what the Ornn index at the top of this post shows. The bulls will say it doesn’t matter, since cheaper tokens mean more tokens (Jevons, etc.). Goldman’s Bobby Molavi flagged Ramp data showing API token usage hit a record above 700 (indexed from 100 in January) in September, yet token spend was down 5% on the month “in spite of it being their peak token consumption month.” Volume up, revenue flat (or down) is a great outcome for users and the infrastructure landlords, but less so for a company whose IPO is priced on revenue growth.

Recall also that 83% of Anthropic’s 2025 revenue came from metered consumption, per the prospectus details reported by Reuters, up from 75% in 2024. As Breakingviews put it, those customers “can cut usage quickly if budgets tighten or cheaper, alternative models emerge.”

The $518 Billion Tab

Which brings us to the other side of the ledger. According to the confidential prospectus seen by Reuters, Anthropic posted a $42 billion net loss in 2025 (about $34 billion of it an accounting charge tied to financing that could eventually convert into shares) on revenue of $4.6 billion, with an operating loss of more than $8 billion and $7.33 billion spent on compute. Going forward, it expects to spend at least $518 billion over roughly a decade with six partners, and “about 80% of that sum is non-cancelable or requires payment regardless of usage”:

“If our actual spend falls short, we must pay Google the difference,” the prospectus says. Amazon gets the same terms. And the single biggest line item, $161.2 billion of Broadcom-related equipment leases, is being financed by the debt market: Broadcom’s $60 billion chip-financing package for Anthropic (“APCHIP 2.0″ in Goldman’s IG credit desk shorthand) began syndicating at the start of the month, with banks marketing a $42 billion senior-secured Class A tranche and Blackstone leading an $18 billion junior piece, as we detailed in “The Crowding Out Effect Is Immense” last Thursday.

Goldman credit strategist Alisha Pasi (available to pro subs) called it “the $150bn question“: the announced chip-financing pipeline now spans “AVGO/Anthropic ($60bn), AVGO/OpenAI ($30–50bn), SPCX/NVDA ($40bn), AMZN/NVDA ($8.5bn) and an additional ORCL/NVDA transaction,” and “AI is increasingly its own credit cycle.” The spread between Goldman’s 30Y hyperscaler basket and its ex-hyperscaler comps widened another 7.4bps last week, to the widest since July. Johnstone put the equity angle bluntly:

Funding and returns is a major concern. Investors are questioning how much chip demand depends on loss-making frontier labs such as OpenAI and Anthropic, while debt-funded data-centre projects, higher borrowing costs and large spending commitments have increased fears around circular financing and credit risk.

How much depends on them? Jefferies’ Chris Wood answered that in his latest quarterly, which we covered on Saturday: of the ~$2.3 trillion in contracted backlog across Microsoft, Alphabet, Amazon and Oracle…

… “about half is owed by OpenAI and Anthropic,” which in Wood’s words amounts to “concentrated unsecured loans to what, for now at least, are cash-burning tenants.”

Hence the one question that matters for strength vs. weakness: is Anthropic raising as much as $100 billion because it can, or because $400+ billion of non-cancelable bills say it must? As we tweeted the night Bloomberg first reported the pre-Thanksgiving timeline, Anthropic is “in a far bigger rush to reload capital” than OpenAI:

BBG report of a pre-Thanksgiving IPO by Anthropic sent odds the company goes public by November to 70%. Ant is in a far bigger rush to reload capital unlike OpenAI which is now happy to wait well into 2027 https://t.co/SebNhAXPXN

— zerohedge (@zerohedge) October 2, 2026

Two weeks later, as shown above, the same market has slumped to a coin flip.

Wrong Time To Be A Price-Taker

The timing doesn’t help either. In Tuesday’s Goldman flow and positioning note, the bank’s Tony Pasquariello forecast “$600bn of primary issuance, which includes $175bn of IPOs,” while “lockup expirations could bring $1.7tr of supply to market” and “net US equity supply could increase the most since the year 2000.” That is the same supply wall Paul Tudor Jones warned about on CNBC this morning, when he said he’ll be “short the whole year” in 2027. One Bear Traps PM pointed out that the Mag 7 fell 15% in the 29 trading days into the SpaceX IPO, and with the S-1 due publicly 15 days before the roadshow, all eyes are on a public filing around Oct 25 for a mid-November debut.

Meanwhile, the cost of money for anyone financing $518 billion of compute just went up again: this morning’s hotter-than-expected CPI put an October hike “firmly on the table.” Privorotsky, who remains “fundamentally bullish on the technology,” spelled out the real risk in his overnight note (available to pro subs):

The issue is whether the cost of capital and physical bottlenecks prevent that demand from translating into actual spending… Competition among the labs may compress model economics without necessarily undermining the infrastructure owners. So fundamentally bullish on the technology and the acceleration towards the next leg of intelligence. But increasingly conscious that the limiting factor may not be demand, but our willingness and ability to finance and deliver the supply.

Even Broadcom’s Hock Tan, whose own credit is paying for many of the chips, warned at the weekend that capital may now be a bigger constraint than chips or power, and that “a lot of investors have no clue what they are investing in.”

Bottom Line

Goldman’s Johnstone laid out what the bulls need: “The clearest positive catalyst would be evidence that total compute demand—token volumes multiplied by compute intensity—continues to accelerate despite greater competition and lower inference costs.” A blowout run rate today, say on track for the $100 billion the NYT reported, would do exactly that, and given how much negativity is priced into semis (Johnstone notes discussions are “increasingly focused on who can make the most negative case”), a squeeze is entirely plausible. JPM’s desk is already flagging “a possible squeeze post-CPI.”

But even the best possible number answers only the demand question. It says nothing about whether ~$400 billion in non-cancelable commitments, the widest hyperscaler credit spread since July, half of a $2.3 trillion backlog and an all-time low token price add up to a company going public “from a position of strength.” Having pounded the table on AI’s debt problem since last October, we lean toward the second half of O’Rourke’s question: a lab growing this fast that still needs $100 billion of new equity, on top of $150 billion of chip debt, to pay its landlords is going public from a position of need. Need and strength can coexist, of course… until the cost of capital decides which one gets priced.

And if anyone is wondering who is on the hook if today’s number disappoints, here is the latest map of who owes whom in the biggest circular financing diagram ever seen: