Fuck It, We're Buying Compute
Anthropic's S-1 shows an $80.7 billion first-half loss. $72.4 billion of it never touched cash. The rest is a real business wearing a fantasy price.
Anthropic filed to go public on Monday morning carrying a net loss of $80,748,000,000 for the first half of 2026, and the reason nobody at the company blushes sits one footnote away: $72,388,000,000 of that figure is a fair-value change on convertible notes and SAFEs, an entry that grows exactly when the company's valuation does. When investors mark Anthropic's paper higher, those instruments become worth more, and accounting rules book the increase as a loss. The headline number is, in a literal sense, a receipt for a successful fundraising run, printed in red.
The instruments come mostly from strategic partners, and the S-1 names Amazon's converts, bought under two agreements dated September 2023 and after. CNBC confirmed the Nasdaq listing under ticker ANT the same morning, with Goldman Sachs and Morgan Stanley lead-managing the deal. Fortune notes the filing leaves price and share count blank, as is customary at this stage, so the most important number in the document is still unwritten, pending conversations with investors who may have already seen a confidential draft.
The company under the headline
Strip the fair-value line and a different animal appears. Second-quarter revenue hit $11.56 billion, up 14.7x year over year and 144 percent sequentially, and that quarter posted $974 million of operating income, the first profitable quarter in the S-1's own table, at an 8.4 percent margin. Operating cash flow for the half ran positive $3.93 billion, against negative $1.51 billion a year earlier. Compute and infrastructure, the cost eating 177 cents of every revenue dollar in mid-2025, cost 48 cents in Q2 2026. Net revenue retention sits near 900 percent, existing customers now spend roughly nine times what they did a year before, and the company ended June holding $131.7 billion of cash and short-term investments.
Growth is real, and rented from a small base: $41 million of revenue in 2023, $386 million in 2024, $4.585 billion in 2025, then $16.29 billion in six months. Over 80 percent of third-quarter revenue came from model versions released in the prior six months, which makes the catalog feel like a restaurant whose entire menu turns over twice a year. One customer represented 11 percent of first-half revenue. Sales, marketing and partnerships spend reached $2.14 billion in the quarter, 22 times last year's level.
The bill for the compute
Yahoo Finance's chart-of-the-day read puts locked-in computing commitments at $430 billion, roughly a fifth of Amazon's and Google's cloud backlogs, while other coverage of the draft ranged to $600 billion-plus, including $111 billion with Google and $110 billion with Amazon; the filing itself leaves the commitments total as of September 30, 2026 blank. What the audited pages do print: $54.6 billion of non-cancelable purchase commitments as of December 31, 2025, before this year's deals, plus a March 2026 equipment lease with Broadcom payable with up to $42 billion of new convertible notes. Anthropic is not exactly buying compute. It is buying compute on margin, against $131.7 billion of cash, at a scale where the financing instrument has become part of the invoice.
What each price needs
Run-rate is doing heavy lifting in this market, so here it is three ways: first-half revenue doubled is $32.6 billion, Q2 quadrupled is $46.2 billion, and the filing's own coy hint, that third-quarter revenue was 86 times the $197 million of Q4 2024, implies about $16.9 billion, a $67.8 billion annualized pace (PitchBook ran the same arithmetic). Those denominators build the table below, and every multiple in it is computed here, on filed numbers and dated coverage.
| Price tag | Anchor | × Q2 run-rate ($46.2B) | × implied Q3 run-rate ($67.8B) |
|---|---|---|---|
| $380B | Series G post-money, Feb 2026 | 8.2x | 5.6x |
| $965B | Series H post-money, May 2026 | 20.9x | 14.2x |
| $1.5T | Low end of IPO chatter (Forbes) | 32.5x | 22.1x |
| $2T | High end (NYT: "could value at") | 43.3x | 29.5x |
Calibration: OpenAI told investors to expect at least $70 billion of annualized revenue by the end of 2026, and one October writeup puts its implied value between $850 billion and $1.4 trillion on roughly $50 billion annualized, a 17x-to-28x band. Every Anthropic price from the Series H to the whisper range hangs 14 to 30 times revenue on a company whose filing books some cloud resale gross, whose share classes give one founders'-company share 50.1 percent of the votes, and whose price range does not exist yet.
The honest crackpipe reading
Which parts are solid business, and which are narrative priced as fact? Solid: the growth curve, the 900 percent retention, the swing from a $1.67 billion operating loss in Q1 to a $974 million operating profit in Q2, compute costs falling from 177 to 48 percent of revenue in four quarters, and a war chest that covers about five years of net building even at double the current pace, since the half saw $17.2 billion invested against $3.9 billion of operating cash flow with $131.7 billion on hand. Narrative priced as fact: that model leadership compounds fast enough to grow into 20 to 30 times sales before the commitments come due, and that customers will keep paying premiums for a catalog that turns over entirely twice a year. A distant third category, the $80.7 billion loss itself, is neither solid nor narrative. It is bookkeeping's echo of the fundraising, and it would shrink if the valuation stopped rising.
Limitations
Third-quarter revenue is an inference from a ratio, not a printed line; the filing blanks it, along with the September 30 commitments total and the October 31 run-rate. An $8.76 billion "other expense" line in the summary table is unexplained in every source checked for this piece. Some cloud sales may be booked gross, which would flatter both revenue and the compute-cost ratio in ways this analysis cannot undo. Fair-value losses move with private marks that the company's own fundraises set, so the decomposition is only as honest as those marks. No price range is filed, so every multiple in the table hangs off chatter or a private round. And this piece did not re-audit the S-1; it read the filing and its dated coverage.
What you can do
If you plan to buy the IPO at whatever range drops, you are underwriting 20 to 30 times unreported run-rate behind a founder share with 50.1 percent of the votes and a trust that elects most of the board, so size it like venture, not like a stock. If you sell into enterprise AI budgets, the filing is your negotiation one-pager: compute fell from 177 to 48 percent of revenue in a year, which means somebody's list price has room to move. If you build anything at all, the actionable number is retention, because nine-times expansion from existing customers is why a company with $430 billion-ish of commitments could sign them.
The strongest case against
The bull case is not that 30x sales is cheap. It is that multiple frameworks misprice companies whose revenue base is 14 times larger than a year ago with margins still improving: Nvidia looked rich at every multiple it ever carried until earnings arrived and made the price ordinary, and Anthropic just printed its first operating profit while growing 144 percent sequentially. If Q3 lands near $17 billion and the 8.4 percent margin scales at all, the trailing multiple compresses faster than any dossier of commitments can grow. The five-class share structure looks like a governance red flag from outside; from inside, it is a machine for ignoring exactly the quarters a doubter would point at, backed by founders who have pledged 80 percent of their own equity to philanthropy.
The Bottom Line
Anthropic's S-1 proves the frontier-lab business can work: real revenue, absurd retention, one genuinely profitable quarter, and costs falling fast enough to matter. It also proves nobody involved needs it to work yet, because $110.8 billion of financing arrived during the same six months the loss did. So the honest verdict on the crackpipe question: the loss is a mirror, the business underneath is real, and the whispered range of $1.5 to $2 trillion is where narrative becomes fact and invoices become destiny. Fuck it, we're buying compute was a financing strategy before it was a ticker. The compute was never the question. The multiple is.