Anthropic Draft Prospectus: $16.3B Revenue, $80.7B Loss, $450B Compute

The Wall Street Journal says it has seen Anthropic's draft IPO prospectus, and the numbers are enormous: $16.3 billion of revenue in the first half of 2026, up from $1.1 billion in the same period of 2025, against an $80.7 billion net loss that is mostly an accounting artifact. About 20 percent of that revenue reportedly came from just two customers. This post separates what is reported from what is verified, because as of today none of it comes from a public SEC filing.

Short version: the revenue story is real and consistent with earlier leaks, the loss headline is misleading without its non-cash context, and the compute commitments are the number that matters most for anyone building on Claude. For the earlier figure-by-figure trail, see our post on Anthropic's reported Q3 revenue of $16.8 billion.

Update (October 20, 2026): the cloud-capacity problem in the filing

TechCrunch adds a detail from the same leaked S-1, as reported by the WSJ: Anthropic is having trouble securing enough capacity from AWS, Google and Microsoft, which increasingly see it as a competitor rather than a customer. The filing reportedly says Anthropic could spend more than $750 billion on cloud contracts, some still under negotiation, and that this is still not enough. Anthropic is leasing its own data centers and using third-party operators, but the filing reportedly worries those operators have a "limited operating history" and may be less reliable than the hyperscalers. TechCrunch also repeats the reported warning that liability for harmful AI behavior could exceed insurance coverage, and that Anthropic is targeting a November IPO. Anthropic did not immediately respond to TechCrunch's request for comment, and none of this is from a public filing yet, so it stays in the "reported" column.

TL;DR: the numbers and their status

The source for the WSJ figures here is a syndicated summary of the WSJ report, which describes a draft prospectus; the page also lists a ticker that the company has not listed, so we do not repeat it. We could not open the WSJ original, so every WSJ-attributed number below is "as reported by the WSJ."

What the draft reportedly says about revenue

Revenue of $16.3 billion in six months is a step change. The same draft reportedly puts H1 2025 at $1.1 billion, a roughly fifteen-fold jump year over year. That is broadly in line with what leaked earlier in the quarter: a Reuters-sourced report carried by CNBC on September 28 described a draft with 2025 revenue of about $4.6 billion, and Fortune listed Q2 2026 at about $11.5 billion, up sharply from roughly $4.73 billion in Q1 (Fortune attributes the Q1 figure to CNBC reporting, not the draft).

Do the arithmetic and the pieces fit: about $4.73 billion plus $11.5 billion is roughly $16.2 billion, within rounding of $16.3 billion. That consistency is the best evidence we have that outlets are looking at the same underlying document. It is still evidence about consistency, not accuracy.

Note that the third-quarter number we covered earlier, $16.8 billion, is a different period. If Q3 alone is near $16.8 billion, then one quarter is already larger than the whole first half. That is plausible given the growth described, but the Q3 figure came from investor-meeting accounts rather than a filing, as our earlier claimed-versus-verified post explains.

Why an $80.7 billion loss is not $80.7 billion burned

This is the number most likely to be misread in headlines. The WSJ-reported draft says about $72 billion of the loss stemmed from non-cash accounting adjustments related to convertible notes and financing instruments.

In plain terms: when a company's valuation rises sharply, instruments that convert into shares (or are marked to fair value) can generate large paper charges. Nobody wrote a $72 billion check. The same pattern appeared in the 2025 figures: the Reuters-sourced reports described an accounting net loss of roughly $42 billion for 2025, with about $34 billion of it a non-cash charge, against an operating loss of just over $8 billion.

Subtract the $72 billion and you are left with roughly $8.7 billion of loss from everything else, which is the number to compare with operating performance. Even that is an inference from two reported figures, not a line item we have seen. Investors will care more about operating loss and cash burn than the headline, and so should readers.

A loss of this kind also does not say anything about whether Claude is cheap or expensive to run. Inference margins, the cost to serve each token, are not broken out in the reporting we found.

Compute: $450 billion, $518 billion, or $600 billion?

The most consequential reported figure is the future spend on cloud capacity. The WSJ summary says Anthropic has committed to roughly $450 billion in cloud compute agreements over the next decade, with specific amounts of $111 billion for Google Cloud, $110 billion for AWS and $31 billion for Microsoft, and that these are noncancelable. It also reportedly says Anthropic "sought more capacity than hyperscalers were willing to supply," which is why it is pursuing third-party data center developments.

Other numbers are circulating. Reuters-sourced coverage cited $518 billion of planned spending on cloud services and data centers, which we covered in Anthropic's reported $517 billion compute commitments. A third-party aggregator headline we saw today rounds the plan to $600 billion. We could not trace $600 billion to any primary report, so we treat it as unverified.

Why would the totals differ? Plausible, unconfirmed reasons: one figure may count only named hyperscaler contracts while another includes data center builds and other partners; one may be a ten-year window and another undefined. Until a public S-1 lists the contracts and their terms, do not average them or pick the largest.

Actual spending so far is smaller and easier to grasp: nearly $9 billion on computing capacity and infrastructure in the first six months of 2026, versus $7.3 billion for all of 2025, per the same WSJ summary.

Customer concentration: two customers, about 20 percent

The draft reportedly says two customers made up about 20 percent of first-half revenue and that Anthropic expects a limited client base to keep supplying a large share. Fortune's account of the Reuters-sourced draft put it at about a quarter of revenue from two clients, so the exact share depends on the period and the document version.

Who they are is not stated in the reporting we found. Large cloud partners and big enterprises that resell or embed Claude are the obvious candidates, but that is speculation. What is documented is that big customers are actively managing their exposure: see our report that Meta and Microsoft cut internal Claude use and the Azure Foundry billing dispute.

Governance and risk disclosures

The WSJ summary also relays several non-financial items:

• The CEO and President each drew base salaries of about $655,000 in 2025.
• Co-founders Dario and Daniela Amodei reportedly pledged 80 percent of their personal equity to philanthropy, and a 5 percent public benefit fund is planned. See our explainers on the public benefit fund and the Long-Term Benefit Trust.
• Risk factors include model alignment, catastrophic incidents, capital intensity, and the warning that "potential damages from severe AI-related security events could exceed available insurance coverage limits."

Fortune, citing the Financial Times, adds that more than a third of the document is risk factors, including existential risks. That framing matches the NYC comptroller's letter urging a pause on the IPO over governance and our longer look at the IPO path and banker meetings.

What this means for what you build or pay

Finance news only matters here if it changes your decisions. Three practical readings:

1. Pricing pressure runs both ways. A company with $450 billion of fixed compute obligations needs high utilization. That can mean discounts and credits to fill capacity, or tight limits when demand spikes. Neither is announced; both are plausible. Track your own cost per task rather than list prices.
2. Concentration risk is a vendor-risk question for you too. If two customers are a fifth of revenue, their decisions can reshape product priorities. Keep your integration portable across providers behind a thin abstraction.
3. Wait for the audited numbers before drawing conclusions. Anthropic has not published these figures. A public S-1 would define revenue recognition (gross or net of cloud-partner resale), the adjusted-profit bridge and the compute contracts.

How to check these claims yourself

• Search the SEC's EDGAR full-text search for a public Anthropic registration statement. As of this writing we could not confirm one.
• Read the originals: the WSJ report, the Reuters draft story as carried by CNBC, and the Financial Times piece Fortune cites.
• Compare the arithmetic: Q1 plus Q2 should approximate H1, and non-cash charges should reconcile the gap between operating and net loss.
• Watch for restatements. Numbers relayed from drafts can be revised before filing, and different outlets may be looking at different drafts.

What is still unknown

• Whether the document is a confidential draft or a public filing, and its exact date.
• The identity of the two largest customers and whether revenue is reported gross or net of partner resale.
• Gross margin, inference cost per token, and cash burn.
• The offering size, share classes, and governance mechanics in detail.
• Why outlets quote $450 billion, $518 billion and $600 billion for compute.

We will update this post if a public filing appears.

Related reading

• Anthropic Q3 revenue of $16.8B: claimed vs verified
• Anthropic's reported $517B compute commitments
• Anthropic IPO path: S-1, banker meetings, and what changes for builders
• Anthropic files confidential S-1
• NYC comptroller letter on pausing the Anthropic IPO
• Anthropic public benefit fund
• OpenAI's reported $750B compute spend through 2030

Figures are as reported on October 20, 2026 and may be revised once a public filing exists.