Anthropic filed an amended S-1 with the SEC on November 2, 2026. It adds the third quarter to the numbers in the original draft. The headline figures: revenue near $17 billion in Q3, a net loss of $23.29 billion, and about $895.4 billion in purchase commitments and leases that are almost all compute. The source is CNBC's report on the filing. We have not read every page of the amendment ourselves.
Earlier coverage on explainx.ai covered the first draft: see Anthropic's draft prospectus: $16.3B H1 revenue and $450B compute. This post covers what is new.
TL;DR: the numbers and who reported them
| Question | Answer | Source |
|---|---|---|
| Q3 2026 revenue | About $17 billion | CNBC |
| Nine-month revenue | $33.24 billion, up about 1,200% | CNBC |
| Q3 net loss | $23.29 billion | CNBC |
| Nine-month net loss | About $104 billion, mostly fair-value changes on convertible notes and derivatives | CNBC |
| Purchase commitments and leases | About $895.4 billion, substantially all compute and infrastructure | CNBC, quoting the filing |
| Compute expense growth | Up $14.4 billion, or 329%, year over year over nine months | CNBC |
What is new compared with the October draft?
The first filing covered only the first half of 2026. The amendment adds July through September. That is why the figures moved so much between filings.
The October draft, as covered in our earlier post, showed $16.3 billion of H1 revenue. Q3 alone is now reported near $17 billion. In other words, one quarter came in larger than the whole first half. That is the single most important fact in the amendment.
On commitments, the draft pointed to roughly $450 billion of compute. The amended filing gives about $895.4 billion for "purchase commitments and lease agreements". The two figures are not like-for-like. The larger one includes leases and other infrastructure, and CNBC quotes the filing as saying "substantially all" of it is compute and infrastructure. Our claimed-vs-verified post on the $800B figure explains why headline commitment numbers need care.
Why is the net loss $23 billion on $17 billion of revenue?
Read this carefully, because the loss is not simply "spent more than earned".
CNBC reports that most of the nine-month net loss of about $104 billion came from changes in the fair value of convertible notes and derivative liabilities. Those are accounting marks. When a company's valuation rises, the value of instruments that convert into its shares also rises, and the difference is booked as a loss. It does not mean $104 billion of cash left the building.
The operating picture is separate. CNBC reports that compute and infrastructure expenses rose $14.4 billion, or 329%, over nine months, and quotes the filing: "As our compute and infrastructure capacity grows, the associated compute and infrastructure expenses have increased." The coverage we could read does not give a Q3 operating-loss figure, so we do not state one.
The gap between the reading levels matters:
- Net loss ($23.29B in Q3): includes non-cash fair-value effects.
- Operating costs: closer to the real cost of running the business. Check the filing's operating line.
- Commitments ($895.4B): future obligations, paid over many years.
What does it mean for people who build on Claude?
A green chip in front of server towers, showing how compute commitments drive Anthropic S-1 costs
The filing is not a price list. But it shows where pressure comes from.
Compute is the cost center. Almost all of the commitments are compute. When nine-month revenue grows about 1,200% in a year, the company has to buy capacity ahead of revenue. That is the logic behind usage limits, rate limits and plan changes at every frontier lab. For context, see how Anthropic's IPO losses and OpenAI usage limits interact.
Growth is concentrated. A roughly 1,200% year-over-year jump in nine-month revenue means a large share of revenue is recent. Developer and enterprise spend on Claude Code and the API is the likely driver, though the sources we read do not give a product split. We do not state one.
Roadmap risk is public now. The S-1 runs more than 90 pages of risk factors, per CNBC, covering political headwinds, infrastructure needs and the balance between safety and commercial goals. One of those risks is supply: our post on the TSMC and Taiwan risk factor shows how chip supply appears in the filing.
If you build on Claude, the practical steps are small but real:
- Budget for price and limit changes. A company with this cost growth will keep tuning plans.
- Keep a second model option in your stack, so a limit change does not stop your product.
- Track cost per task, not just cost per token. See our token-economics coverage of the first S-1 for the framing.
What is confirmed and what is not
Confirmed by reporting that cites the filing: the Q3 revenue figure near $17 billion, the $23.29 billion Q3 net loss, the $33.24 billion nine-month revenue, and the $895.4 billion commitments figure, which CNBC quotes directly.
We did not independently open the amended SEC document, so all figures rest on CNBC's reading of it.
Not confirmed: an IPO date, a price range, or the number of shares. An S-1 amendment is a step in registration, not a listing. Do not read the filing as a signal that trading starts on any given day.
You can read the original registration statement on the SEC's EDGAR site and compare it with the amendment when it is indexed.
Politics and governance around the filing
The numbers land in a political week. House progressives have drafted a resolution against AI bailouts that cites the prospectus; our post, House Progressives Draft a "No AI Bailouts" Resolution, lays out what is confirmed. A $895 billion commitment figure will feed that argument whichever way it goes.
Governance is the other thread. Experts debated whether the company's structure can resist public-market pressure; see Anthropic IPO governance.
Timeline: how the filing story unfolded
The S-1 story has moved in steps, and each step changed the numbers people quote.
- The first draft. The original registration statement covered only the first half of 2026. It showed $16.3 billion of H1 revenue and a large loss, and it pointed to roughly $450 billion of compute spending. Our draft prospectus post records that stage.
- The commitments debate. Commentators argued over whether an $800 billion figure meant cash owed or capacity reserved. The claimed-vs-verified post separates the two.
- Politics. Members of Congress began citing the prospectus in a push against federal support for AI companies.
- The amendment. On November 2, the filing added Q3. The commitments figure rose again, to about $895.4 billion, per CNBC.
The pattern is clear: every update moves the compute number up. If you follow this story, track that number first. It tells you how much capacity the company has promised to buy, and so how hard it must push revenue per unit of compute.
Questions readers ask about the filing
Is a 1,200% revenue jump normal? No. Even for a fast-growing software company it is extreme. It also makes year-over-year comparisons fragile, because the base quarter in 2025 was small. Look at quarter-over-quarter change as well when the full tables are available.
Does a $23.3 billion loss mean Anthropic is running out of money? Not from this evidence. The reported loss is driven largely by fair-value accounting. The filing's liquidity section is the place to check cash on hand; we did not find that figure in the coverage we read, so we do not state it.
Will Claude get more expensive? The filing does not say. What it does show is that costs scale with compute, which is why limits and plans keep changing across the industry. Watch the official pricing page, not the prospectus, for actual changes.
How to read the next filing
When Anthropic or any lab files again, check four things in this order:
- Revenue by quarter, to see if growth is still accelerating.
- Operating expenses, which show cash-like cost growth.
- The net loss footnotes, to separate fair-value marks from operating loss.
- Commitments and their timing, to see how many years the obligations run.
A single big number in a headline rarely survives these four checks unchanged. That is the pattern in this filing as well.
Bottom line
The amended S-1 shows a company growing very fast and spending even faster on compute. The $17 billion quarter is as reported by CNBC. The nine-month loss is driven mostly by fair-value accounting, per CNBC. The $895 billion in commitments is the number to watch, because it ties Claude's future prices and limits to compute supply.
Related reading
-
Anthropic S-1/A: "dramatically accelerated progress" threshold, claimed vs verified
-
Anthropic's internal Claude use, at retail prices, now exceeds employee pay
-
Anthropic draft prospectus: $16.3B H1 revenue, $450B compute
-
Anthropic's $800B compute commitments in the S-1: claimed vs verified
Figures are accurate as of November 2, 2026, based on CNBC reporting of the amended filing. Check the SEC filing for final numbers.