Fortune reported October 21 that it had viewed a draft S-1 showing Anthropic posted $16.3 billion in revenue against an $80.7 billion net loss for the first half of 2026 — numbers that Dealroom, ZeroHedge, and The Next Web each corroborated within days, all citing the same Wall Street Journal review of the document. That’s two days after this site covered Replit walking most of its AI spend to OpenAI because Anthropic wouldn’t match a 30%-plus discount. Read the two stories back to back and the second one stops looking like a coincidence. It looks like the balance sheet behind the first.

Here’s the thing about an $80.7 billion loss: most of it isn’t what it sounds like. But the part that is real — compute spend nearly quadrupling, hyperscalers reportedly starting to treat their biggest AI customer like a rival — is a harder problem than a bad headline number. It’s a disclosed risk that the IPO’s growth math depends on three suppliers who have their own reasons not to cooperate.

Quick Summary: What’s In the Draft S-1

Detail	Info
Reported	Oct 21, 2026, by Fortune, citing a WSJ-reviewed draft filing
H1 2026 revenue	$16.3 billion (up from $1.1 billion in H1 2025)
H1 2026 net loss	$80.7 billion (up from $13 billion in H1 2025)
Non-cash portion of the loss	~$72 billion, tied to convertible-note/financial-instrument revaluation
H1 2026 compute spend	~$9 billion, versus $7.3 billion for all of 2025
Future compute commitments	Up to $750 billion, including $111B to Google Cloud, $110B to AWS, $31B to Microsoft
Risk factor disclosed	Hyperscalers increasingly treat Anthropic as a competitor, not just a customer, for compute access

Bottom line: Strip out the accounting noise and Anthropic is still burning roughly $8.7 billion in six months to run a business whose own suppliers — the three companies it’s paying $750 billion to — are starting to see it as a rival.

What’s Actually In the Draft S-1

Start with the scary number, because it’s the one every headline led with. $80.7 billion in net loss against $16.3 billion in revenue, for six months. Read alone, that’s a company burning through cash at a rate that makes even OpenAI’s widely-reported losses look tame.

It’s also mostly not cash burn. About $72 billion of that $80.7 billion comes from a non-cash accounting charge — the kind of thing that happens when a private company with a fast-rising valuation has to mark its convertible notes and other financial instruments to market as part of going public. Anthropic’s valuation has moved from roughly $965 billion in June to a target near $2 trillion for the IPO itself. Every dollar of that paper gain on instruments issued to early investors gets booked as a loss under the accounting rules that govern this stuff. It’s real on paper. It isn’t a check anyone wrote.

Strip that out and the operating loss for the half is closer to $8.7 billion. Still a loss. Still nine figures a month. But a fundamentally different number than $80.7 billion, and conflating the two — which plenty of the breathless coverage this week did — tells you more about the headline-writer’s incentives than about Anthropic’s finances.

The Number That Isn’t an Accounting Artifact

Compute spend is the one line in this filing that’s just… real. Anthropic spent nearly $9 billion on compute and infrastructure in the first six months of 2026 alone. For comparison: the company spent $7.3 billion on compute in all of 2025. Six months this year cost more than twelve months last year, and revenue growing from $1.1 billion to $16.3 billion over the same stretch doesn’t change the fact that the input costs are growing just as fast, maybe faster.

That’s the part of this filing that should worry anyone modeling Anthropic’s path to profitability. Revenue scaling 14x year-over-year is the kind of number that makes a $2 trillion IPO target sound almost reasonable. Compute costs scaling in lockstep — or faster — is the reason “reasonable” has an asterisk on it.

Why Is Anthropic Struggling to Secure Enough Compute?

This is the part of the filing that got less headline attention than the loss number but matters more for anyone trying to understand Anthropic’s pricing behavior. The draft reportedly includes a risk factor admitting the company can’t always get the compute it wants, for a specific reason:

1. AWS, Google, and Microsoft are Anthropic’s primary compute suppliers — the same three companies it’s locked into tens of billions of dollars in noncancelable cloud contracts with.
2. Those same three companies increasingly compete with Anthropic directly — through their own foundation models (Amazon’s in-house efforts, Google’s Gemini, Microsoft’s stake in and distribution of OpenAI) — which changes their incentive to prioritize Anthropic’s capacity requests.
3. Anthropic has reportedly sought more computing power than its hyperscaler partners were willing to provide, forcing it toward third-party and “neocloud” operators with shorter track records to fill the gap.
4. That dependency compounds the commitment risk — Anthropic is locking in up to $750 billion of future spend partly because it can’t fully trust that its biggest, most established suppliers will keep prioritizing it as their own competing products mature.

That’s a genuinely unusual position for a company to disclose in an IPO filing: our suppliers are also our competitors, and that’s starting to show up in whether we can get what we ordered. It’s the kind of sentence that sounds abstract until you connect it to something concrete — like a discount negotiation Anthropic chose to lose rather than win.

Why This Matters: A Risk Factor That Outlasts Any Single Deal

As covered in Why Replit Dropped Anthropic for OpenAI’s Discount, Anthropic let a $100 million account walk rather than discount. That’s a pricing dispute, and pricing disputes have a shelf life — contracts get renegotiated, accounts come back, and a CFO can reverse the decision next quarter if the math changes.

The hyperscaler clause in this filing doesn’t have that kind of exit. Anthropic has reportedly locked in up to $750 billion in compute commitments with Google Cloud, AWS, and Microsoft, structured as noncancelable obligations running seven to ten years — money owed whether or not those three companies prioritize Anthropic’s capacity requests over their own competing models. A customer walking away is a revenue line that can recover. A decade-long dependency on suppliers who are also your most capitalized rivals is a term baked into the valuation itself, one that doesn’t reset when a quarter goes well.

That’s the distinction worth sitting with: one risk is about whether Anthropic can hold a price. The other is about whether the $750 billion it’s already committed to spend buys the compute it’s counting on, or buys leverage to three companies with every reason to meter it out.

The Bigger Picture

None of this is new information about whether Anthropic can build good models. Nothing in a leaked S-1 changes how Claude performs against GPT on any benchmark that matters. What it changes is the read on Anthropic’s negotiating posture going forward, right as it’s trying to close out an IPO that reports have pegged at a $2 trillion target — a number that assumes investors look past an $80.7 billion headline loss and buy the “$72 billion of it is non-cash” explanation.

That’s not a crazy thing for investors to accept. OpenAI’s own S-1 filing carries its own eye-watering loss figures, and nobody seriously expects either frontier lab to be profitable before the back half of this decade. The compute race is expensive by design; both companies are betting the market they’re building is worth financing at a loss until it isn’t. The difference this filing surfaces is that Anthropic’s version of that bet now includes a disclosed risk that its own infrastructure partners might not fully cooperate — which is a problem OpenAI, backstopped by its own dedicated chip and data-center buildout, is positioned somewhat differently against.

Anthropic is reportedly targeting a listing as soon as November 2026, after the U.S. midterms clear the calendar. Whether this leak changes that timeline, or changes how bankers price the deal, isn’t public yet. What’s public is that the filing landed in the same week as the clearest evidence yet that Anthropic’s enterprise customers are price-shopping — and the filing itself explains, in its own risk factors, why Anthropic has less room to win that fight on price than its rival does.

Our Take

I don’t think $80.7 billion is the number anyone serious should be panicking over — $72 billion of it evaporates the moment you understand convertible-note accounting, and any analyst treating the headline figure as cash burn either doesn’t understand the filing or is choosing the scarier number on purpose. The real number is $8.7 billion in six months, and even that deserves context: it’s the cost of running a company whose revenue just grew 14x year-over-year.

What I’d weight more heavily than either loss figure is the hyperscaler clause, because it’s a claim about Anthropic’s ceiling, not its current burn rate. A $2 trillion IPO target is a bet on a growth curve — revenue and compute scaling together, indefinitely, with margin eventually catching up. A disclosed risk that the three companies supplying that compute have their own reasons to ration it puts a cap on that curve no quarter of strong revenue can lift, because the constraint isn’t demand or capital. It’s whether Google, Amazon, and Microsoft keep choosing to prioritize a competitor’s capacity requests over their own products as those products improve. That’s a harder number for bankers to price than $80.7 billion, because it doesn’t show up on an income statement at all. It shows up the first time Anthropic asks for more compute than it gets.

Frequently Asked Questions

Did Anthropic actually lose $80.7 billion in cash?

No. According to the draft S-1 reported by Fortune and corroborated by Dealroom, roughly $72 billion of the $80.7 billion net loss is a non-cash accounting charge tied to revaluing convertible notes and other financial instruments as the company’s valuation has risen. The underlying operating loss for the half is closer to $8.7 billion.

How much is Anthropic spending on compute?

Nearly $9 billion in the first half of 2026 alone, according to the leaked filing — more than the $7.3 billion it spent across all of 2025, per ZeroHedge’s reporting on the same draft document.

What does “up to $750 billion in future compute commitments” actually mean?

It’s the total value of cloud and chip contracts the draft S-1 discloses Anthropic has signed or is negotiating, including $111 billion with Google Cloud, $110 billion with AWS, and $31 billion with Microsoft, much of it reportedly structured as noncancelable, pay-regardless-of-use commitments stretching seven to ten years.

Why would Anthropic’s cloud providers compete with it?

Google, Amazon, and Microsoft each sell their own AI models or have major stakes in companies that do, while also being the companies Anthropic pays tens of billions of dollars to for the compute Claude runs on. The draft S-1 reportedly discloses this as a risk factor — that those suppliers increasingly treat Anthropic as a rival rather than purely a customer when allocating scarce capacity.

Is this related to Replit and Rippling switching to OpenAI?

Not directly reported as connected, and the two issues are distinct. Replit and Rippling moved significant AI spend to OpenAI over a pricing dispute, which is a reversible, account-by-account decision. The hyperscaler risk factor in the S-1 is a different kind of problem — a structural dependency on the same three companies Anthropic has committed $750 billion to, which doesn’t resolve itself the way a single customer renewal can.

When is Anthropic’s IPO expected?

As soon as November 2026, based on reporting cited alongside the S-1 leak, with a valuation target that’s been reported near $2 trillion — more than double the $965 billion mark it held in June.

Is a draft S-1 the same as a final IPO filing?

No. A draft S-1 is a confidential or near-final working document that can still change before Anthropic files publicly with the SEC. Numbers, risk factors, and even the listing timeline could shift between this leaked draft and whatever Anthropic ultimately registers.

---

Last updated: October 26, 2026. Sources: Fortune — Anthropic’s S-1 leaks (again) · Dealroom — Anthropic’s draft S-1 reveals $16.3B revenue, $80.7B loss and up to $750B compute bill · ZeroHedge — $80.7 billion in red ink and a $600 billion tab · The Next Web — Anthropic could spend $750bn on compute.

Related reading: Why Replit Dropped Anthropic for OpenAI’s Discount · Anthropic’s $965B IPO: What It Means for Claude Users · Anthropic’s IPO Could Top SpaceX · Anthropic’s $45B Nscale Deal · OpenAI Files for IPO