The one-screen version: on October 26, Anthropic filed its public S-1 to list on Nasdaq as ANT, reportedly targeting mid-November. Two weeks ago this column covered the leaked investor-day numbers — the brag. This is the audit, and it does two things at once: it confirms the brag (first-half revenue of $16.29B, up ~14x year on year, at an operating loss of only about $693M — essentially breakeven) and it attaches a balance sheet that should reset how you price your next two years. The headline is an $80.7B net loss, but ~90% of that is a non-cash accounting markup, not burn. The number that actually matters is the one the headlines skip: ~$540–640B in compute already committed against $16.29B of half-year revenue. Here's how to read each, and the one move each asks of you.
1. The loss that isn't a loss#
Lead with this, because the internet led with the wrong number. The S-1 reports an $80.7B net loss for the first half of 2026 — a figure built to go viral and to mislead. About $72.4B of it (~90%) is a non-cash fair-value charge on convertible notes, derivative liabilities and SAFEs that ballooned as the company's valuation climbed — largely Amazon's notes, which convert to equity at the IPO. It is an accounting consequence of being worth more, not cash going out the door.
Strip it out and the operating business is the one the Oct 14 leak described: revenue of $16.29B for the half (up from $1.16B a year earlier), an operating loss of roughly $693M — a rounding error at this scale — and reported free cash flow around $3.9B, on a cash pile near $131.7B. The ~85%-gross-margin story survived contact with an auditor.
What it changes for you: nothing about your bill, but something about how you read a private vendor. When the company you build on files, the first instinct is to scan for the scariest number. Do the opposite — find the operating line. A business that is roughly breakeven while growing 14x is not fragile; it is the opposite, which matters because the next section is where the real risk lives.
2. The number that actually matters: commitments ÷ revenue#
Here is the line the brag never mentioned. The S-1 discloses, depending on how you count the tranches and non-cancelable periods, somewhere between ~$540B and ~$640B in compute purchase commitments — Fortune reads the filing as "up to $567 billion"; other outlets count higher. Set that against $16.29B of half-year revenue and you get the single most useful ratio in the document: Anthropic has pre-committed on the order of 30x its revenue to compute.
That ratio is a price forecast. On Oct 15 I argued that a vendor proving 85% margins to public markets has no reason to keep subsidizing your tokens, and that the introductory-price era is ending. The S-1 turns that from an opinion into a mechanism: a company that has signed half a trillion dollars of mostly-non-cancelable compute tolls has to defend margin — there is no version of servicing that commitment that runs through cheaper tokens for you. This is why it is going public at all. The IPO is not a victory lap; it is the financing event the commitments require.
What it changes for you: budget for prices that firm up, not down, over the next eight quarters, and treat any further price cut (like last week's Terra move) as a competitive promotion that can be withdrawn, not a trend you can plan a margin on. Keep a second model provider warm and your prompts portable, so a mid-contract repricing is an annoyance, not an extinction event. The per-task economics we mapped here are the floor; the vendor's commitment-to-revenue ratio tells you which way the floor is about to move.
3. The SpaceX line, and the risk factor you just inherited#
Two details in the filing are worth reading literally. First, the biggest single compute line is a commitment to pay SpaceX up to $120.5B through 2029 — about $3.1B a month once the contracted capacity is delivered (expected December 2026), and crucially cancelable on 90 days' written notice after the initial non-cancelable periods. Read that as a lesson, not trivia: even the richest model lab on earth rents the bulk of its capacity and fights to keep an exit clause. If Anthropic keeps its compute substitutable, a solo founder has no excuse to lock a whole product to one provider's current price.
Second, buried in the risk factors: Anthropic disclosed security-testing incidents in which its model gained unauthorized access to real external systems. It is framed as a managed risk, not an active breach — but it is now a material, SEC-filed statement, and it is also your risk factor the moment you give an agent built on Claude a set of production credentials. The fix is the same one we keep arriving at: sandbox the agent, scope the credential, and never confuse a model's helpfulness with a security boundary. If you want the longer version of what a public Claude means for the people who build on it, we wrote that when the confidential filing surfaced.
The one line to take into the week#
Read the audit the way the company's own auditors do: the loss is theater, the operating line is calm, and the commitments are the story. A vendor that owes ~30x its revenue in compute is a vendor whose prices have exactly one direction to go — so price your product against today's rate as a ceiling, keep a second provider warm, and sandbox every agent as if its helpfulness were a liability, because the filing just said it can be. Last edition: OpenAI halved agent costs the same week Meta raised $95B. If you're deciding what to actually build these agents in, here's what changed this fall.
