Anthropic Bought $5B of Burst Compute and Flipped to a Loss. Match the Lease, Then Watch the Minimum.

Anthropic's amended S-1 is the week's number, and two people on the roster read different pages of it. This is a Ship Story: the decision, the result, and the line both readings leave out.

The short version

Tomasz Tunguz's November 3 reading of Anthropic's S-1/A: $5.0 billion of short-term burst compute in Q3, 51% of a $9.86 billion quarterly compute bill, at $7 to $13 per GPU-hour against $2 to $4 on multi-year contracts; blended cost per GPU-equivalent hour from $3.10 to $5.65, up 82% in two quarters; Q2's $974 million operating profit turned into a $567 million Q3 loss while revenue grew $5.40 billion to $16.96 billion. His 2027 rule is to sell the Fortune 500 buyer a multi-year minimum commitment at a locked-in rate that matches the GPU lease underneath it. Jason Lemkin's November 4 reading, with Ramp's AI Index: the quarterly dollar add fell from $6.84 billion to $5.40 billion, Ramp's top 1% of firms spent $9,568 per employee in September against a July peak of $9,873, that 1% drives 80% of spend on OpenAI and Anthropic, and Ramp's economist Ara Kharazian blames price cuts, with Opus 5.5 at about 40% lower cost per task. The decision was to buy the burst; the result is the loss and a dollar add shrinking under the price cut. Tunguz's fix is right for the vendor and moves the mismatch to the buyer, and I have sold that contract: in the per-seat sunset I ran, 12 accounts paid their committed minimums while usage ran under them, the dashboard read on plan, and two of them left when the legacy surface went away. Match the lease, then watch utilization against commitment per account per month, write the pass-through decision before the minimum is sold, and run Lemkin's top-1% check in tasks rather than dollars.

The compute pages

Tunguz's post is short and the footnotes are the filing. In the third quarter Anthropic spent $5.0 billion on short-term compute alone, 51% of its $9.86 billion quarterly compute bill. The burst market, as he reports it, runs $7 to $13 per GPU-hour, against the $2 to $4 a lab locks in on multi-year contracts, and when the contracted capacity is exhausted the lab buys burst at those prices. The effect on the blended number: cost per GPU-equivalent hour went from $3.10 to $5.65, within a $3.45 to $8.29 confidence band he derives from the contract commitments and public capacity disclosures, up 82% in two quarters. Q2's $974 million GAAP operating profit became a $567 million Q3 operating loss, even as quarterly revenue grew $5.40 billion sequentially to $16.96 billion, and long-term obligations reached $895.4 billion.

His reading of the response: a doubling of GPU cost has to be met by a doubling of model efficiency, a doubling of price, or some of each, and raising token prices in a price war cedes share. So labs push on three fronts. Efficiency, with Opus 5.5's roughly 40% reduction in cost per task over Opus 5. Tighter allowances, with OpenAI pooling ChatGPT and Codex into one shared token limit from December. And moving enterprise buyers from serverless consumption into multi-year committed contracts.

Then the rule, which is the post's title. Financial engineering in 2027 means closing the duration mismatch: selling a Fortune 500 buyer a multi-year minimum commitment at a locked-in rate to match the duration of the GPU lease underneath it. In 2026 the race was to find GPUs at any price. In 2027 the winners are the labs that match the duration of their customers' contracts to the duration of their compute.

The revenue pages

Lemkin read the same filing a day later, next to Ramp's AI Index, which is built from transactions at more than 85,000 US businesses.

Revenue rose to $16.96 billion from $11.56 billion in Q2, but the dollar add fell: Q2 added $6.84 billion, Q3 added $5.40 billion, 21% less. Ramp's top 1% of firms spent $9,568 per employee per month on AI in September, down 3.1% from a July high of $9,873, after an August drop of 13% that September mostly recovered. The top 10% and the median kept climbing. That top 1% drives 80% of spend on OpenAI and Anthropic, at a median of about $9,568 per employee against $15.70 for the median firm, which is why a few hundred companies running agents around the clock set the revenue line and new adopters at $16 a head cannot replace them when they pause.

The cause, in Lemkin's telling and Ramp's: price. Ramp's lead economist Ara Kharazian attributes the slowdown mainly to falling prices for top models, and Opus 5.5 cut customer cost per task by about 40%. If a heavy user's workload stays the same size, its bill falls by up to 40%, and the top 1% did not add enough new work in August and September to offset that. Summer explains less; by his arithmetic an August vacation dip would have to be about two and a half times the size of Christmas to close the gap.

His check for any B2B AI company: rank customers by monthly spend, take the top 1%, and compare their September total to July. Flat while the customer count grows means a handful of accounts carry the line and have stopped expanding.

The Ship Story

The decision: buy the burst. Anthropic had more demand than contracted compute and paid $7 to $13 an hour for the difference, half its compute bill in one quarter, which is what you do when the alternative is turning customers away in a price war.

The result: the loss, and a dollar add that shrank under the company's own price cut while the heaviest customers held their workloads flat.

Tunguz's fix is right for the vendor, and I want to say that plainly before I argue with it. A multi-year minimum at a locked-in rate against a multi-year lease at a locked-in rate is the match, and it turns the burst bill into a planning number. The SaaS to AI Business Models thesis is that the margin structure moves with the pricing model, and this is the margin structure moving.

The part I have lived is what the matched contract does on the buyer's side. It closes the vendor's duration mismatch by moving it to the customer, who now carries a minimum for three years against a workload that was sized in the quarter the contract was signed.

I sold committed minimums

Field Report: What Broke When We Killed Our Per-Seat Tier has the record. By month 18 of the migration, 12 accounts were classified as on plan. They had migrated on paper and were paying the new committed minimums. Their actual usage was below the minimum, so they generated no overage, and their account teams had not escalated because the dashboard said on plan. When the legacy tier sunset, those 12 did not notice. Three weeks later they found the old surface gone, the escalation was angry, and we lost two of them.

The lesson I wrote down then was that migrated is not the same as using the new pricing, and that the tracker should have flagged low utilization despite commitment as its own class. The dashboard measured the contract. The behavior moved first and the dashboard never did.

Now put that next to a 40% drop in cost per task. A buyer who signed a three-year minimum sized on the old cost per task is now carrying a commitment that the same workload no longer fills. The vendor's revenue line reads on plan for 36 months. The buyer's utilization against the minimum reads under, and it reads under more each time the lab passes another efficiency gain through. That contract is not revenue. It is the sunset escalation with a date on it, and the date is the renewal.

Match the lease, then watch the minimum

Three lines, none of which either post has.

Utilization against commitment, per account, per month. The number the vendor keeps on itself and reads before the renewal conversation, not during it. Under for two consecutive months is the escalation my dashboard did not raise.

The pass-through decision, written before the minimum is sold. Anthropic's S-1 Has No Gross Margin Line. Your AI Product Needs One. argued that when a cost-per-task cut arrives, someone has to decide whether it becomes a lower price, better quality, or margin, and most companies decide by accident. A committed contract makes the accident expensive: a commitment sold at the old cost and renewed at the new one is where the trough starts, and ServiceNow Fell 27% on a 70.5% Gross Margin. Name the Trough and the Month. is what a trough with no stated bottom does to a stock.

Lemkin's top-1% check, run in work rather than dollars. Tasks per employee in the top 1%, September against July. A 40% price cut makes dollars fall while work holds, and the question a vendor needs answered is whether the work held.

Match the lease. Then watch the minimum. The lease is the vendor's problem and Tunguz solved it. The minimum is the customer's, and it comes back to the vendor on the day it expires.

This week, if you sell committed contracts: pull utilization against commitment for every account on a minimum, for the last three months, and count the ones under. Then find the pass-through decision for your last cost cut in writing. If it is not written down, that is the accident, and the minimum is where it lands.

Related answer: Should an AI vendor sell multi-year minimum commitments to match its compute leases?

Sources: Tomasz Tunguz, "The Financial Engineering Essential for AI in 2027," November 3, 2026, read in full; every compute, cost, profit, and obligation figure is Anthropic's S-1/A as Tunguz reports it, and the per-hour rates and confidence band are his derivation as described in his footnotes; the filing was not opened for this post. Jason Lemkin, "The Top 1% of AI Spenders Stalled in Q3. So Did Anthropic's Dollar Growth.," SaaStr, November 4, 2026, read in full; the Ramp AI Index figures are Ramp's as Lemkin reports them, Ara Kharazian's attribution is as Lemkin quotes it, and the 40% cost-per-task figure is Anthropic's as both posts carry it. Falk Gottlob, "Field Report: What Broke When We Killed Our Per-Seat Tier" (the 12 accounts and the two lost), "Anthropic's S-1 Has No Gross Margin Line. Your AI Product Needs One." (the pass-through decision), and "ServiceNow Fell 27% on a 70.5% Gross Margin. Name the Trough and the Month.."