Anthropic’s Compute Bill Fell to 48% of Revenue From 251%

Anthropic’s compute and infrastructure expense fell from 251% of revenue in the first quarter of 2025 to 48% in the second quarter of 2026, according to the quarterly tables in the Form S-1 it filed with the SEC on 26 October 2026.

In the same table, the quarter to 30 June 2026 shows operating income of $974 million, after operating losses in each of the five earlier quarters presented. The filing says it “generated operating income of $1.0 billion for the three months ended June 30, 2026, as revenue growth outpaced the growth in total operating expenses.”

Business Pill · FORECAST VS COMMITMENT

A short Business Engineer explainer on the difference between a forecast and a commitment. It teaches the general idea only and says nothing about Anthropic or its filing.

The key insight: As we read it, the new element is the trend inside the year: the S-1’s own percentage table shows compute and infrastructure expense falling from 251% of revenue to 48% in six quarters, and the cost base behind it is a set of multi-year commitments the filing itemises counterparty by counterparty.

The Quarterly Numbers

Revenue rose in each quarter, from $376 million in the first quarter of 2025 to $11.56 billion in the second quarter of 2026, the S-1 shows. Compute and infrastructure expense rose from $944 million to $5.56 billion over the same quarters.

QuarterRevenue ($M)Compute and infrastructure ($M)Compute as % of revenueOperating income (loss) ($M)Revenue per $1 of compute
Q1 of 2025376944251%(935)$0.40
Q2 of 20257871,390177%(1,200)$0.57
Q3 of 20251,3822,047148%(3,089)$0.68
Q4 of 20252,0402,945144%(2,836)$0.69
Q1 of 20264,7313,38772%(1,667)$1.40
Q2 of 202611,5555,55848%974$2.08

Source: Anthropic, PBC Form S-1, Quarterly Consolidated Statements of Operations and the same table as a percentage of revenue, filed 26 October 2026 (unaudited; quarters end 31 March, 30 June, 30 September and 31 December). The last column is our arithmetic (revenue divided by compute and infrastructure expense); it is not a metric the S-1 reports.

The third quarter of 2025 includes the $1.5 billion Bartz settlement charge, the S-1 notes. Net loss for the second quarter of 2026 was $66.31 billion, mostly a $59.60 billion change in the fair value of convertible notes, derivative liabilities and SAFEs, according to the same table.

Compute and infrastructure expense as a percentage of revenue, by quarter, as reported in the Quarterly Consol
Compute and infrastructure expense as a percentage of revenue, by quarter, as reported in the Quarterly Consolidated Statements of Operations of Anthropic’s Form S-1 filed 26 October 2026 (unaudited). The quarter to 30 June 2026 shows operating income of $974 million.

Revenue per Compute Dollar

By our arithmetic, Anthropic recorded $2.08 of revenue for every dollar of compute and infrastructure expense in the second quarter of 2026, up from $0.40 in the first quarter of 2025. The S-1 itself defines its goal this way: “We define compute efficiency as maximizing the intelligence produced and revenue generated per dollar of compute invested.”

The filing gives its own reading of the ratio: “Compute and infrastructure expenses as a percentage of revenue decreased from 251% to 48% over the same period, as revenue growth outpaced the growth of our committed compute cost base.”

It says the cost line moves on its own schedule. “Because we typically procure compute 12 to 18 months in advance of availability, these expenses are generally fixed in the near term and increased as committed compute became available to us rather than in response to in-period demand.”

The Commitments Behind the Cost Line

The S-1 leaves the total of its contractual obligations blank, but its MD&A describes the largest arrangements one by one. They include at least $110.0 billion with Amazon between May 2026 and April 2036, at least approximately $114.5 billion with Google between April 2026 and July 2033, and at least $44.9 billion with Microsoft between November 2026 and May 2033, each “regardless of actual usage.”

CounterpartyAmount as the S-1 states itTerm or nature
Broadcom (equipment leases, TPU capacity)Rent of approximately $161.2 billion, excluding interestFive-year master equipment leases, accounted for as finance leases
SpaceXAI (compute powered by NVIDIA GPUs)Up to $120.5 billionTerms extending into 2029; about $3.1 billion a month once all capacity is delivered
GoogleAt least approximately $114.5 billionApril 2026 to July 2033
AmazonAt least $110.0 billionMay 2026 to April 2036
MicrosoftAt least $44.9 billionNovember 2026 to May 2033
Nscale (NVIDIA GPUs)Up to approximately $44.6 billionTerms extending into 2033
Lambda (NVIDIA GPUs)Approximately $41.3 billionTerms extending into 2034
AMD (Helios rack-scale systems)No fixed dollar commitmentExpected to exceed $20.0 billion if the full initial capacity is procured

Source: Anthropic, PBC Form S-1, MD&A, Infrastructure Service Arrangements and Infrastructure Equipment Leases, filed 26 October 2026. Minimums, ‘up to’ amounts and lease rent are different kinds of figures, so we do not add them up.

The SpaceXAI arrangements are with SpaceX.AI LLC, which the S-1 describes as “a wholly owned subsidiary of Space Exploration Technologies Corp.,” for compute powered by NVIDIA GPUs. After initial non-cancelable periods ending in January 2027 and March 2027, each of those agreements may be cancelled by either party on 90 days’ written notice, the filing says.

Beyond the Google commitment, the S-1 estimates further TPU purchases, to be delivered between 2028 and 2031, at approximately $125.0 billion to $150.0 billion, and says those payments are not included in its contractual commitments table.

Table of Anthropic compute arrangements as stated in its S-1: Broadcom lease rent about $161.2 billion, SpaceXAI up to $120.5 billion, Google at least about $114.5 billion, Amazon at least $110.0 billion, Microsoft at least $44.9 billion, Nscale up to about $44.6 billion, Lambda about $41.3 billion, AMD no fixed amount
Compute arrangements as described in the MD&A of Anthropic, PBC’s Form S-1 filed 26 October 2026. The figures are minimums, caps or lease rent and are not added up.

The Mismatch the S-1 Flags

The risk factors set the cost base against how revenue arrives: “our revenue is primarily consumption-based, and much of it is not subject to long-term contractual commitments, while our compute and infrastructure commitments generally extend over multiple years.”

Connecting the Coverage

Our S-1 piece reported the headline figures: revenue of $16.29 billion for the first half of 2026 and a net loss of $80.75 billion. Our 22 October commitments piece reported The Wall Street Journal’s figure of about $450 billion from a draft prospectus; the filed S-1 leaves its total blank but itemises the arrangements above. Our leases piece covered $1.2 trillion of leases that big tech companies have signed but not started.

Our SpaceX lease piece covered SpaceX’s 8-K on a $40 billion GPU finance lease with an Apollo-managed entity, filed on the same day. Our SpaceXAI piece covers the largest new arrangement in the filing, up to $120.5 billion for NVIDIA-powered compute, in detail (Anthropic to pay SpaceXAI up to $120.5B). What is new here is the quarterly trend and the itemised commitments, from Anthropic’s own filing.

The Structural Read

As we read it, the ratio moved because revenue outran a cost line the S-1 calls generally fixed in the near term: compute is procured 12 to 18 months ahead, so the expense arrives on the supplier’s schedule while revenue arrives with usage.

The S-1 adds a second efficiency figure: “with approximately 5,300 employees, our TTM revenue per employee was over $6.6 million,” as of 30 September 2026. Elsewhere the filing compares it with a year earlier: “TTM revenue per employee of over $6.6 million as of September 30, 2026, compared to $1.4 million as of September 30, 2025.” As we read it, the filing presents compute and headcount as the two denominators it manages against.

As we read it, the commitments table is what turns the margin trend into a bet: the S-1 lists minimums with Amazon, Google and Microsoft that apply “regardless of actual usage,” while it describes revenue as primarily consumption-based.

Anthropic, PBC Form S-1, 26 October 2026

“We define compute efficiency as maximizing the intelligence produced and revenue generated per dollar of compute invested.”

Three Implications

FOR INVESTORS The S-1 reports operating income of $974 million for the quarter to 30 June 2026 and a net loss of $66.31 billion in the same quarter, driven by a $59.60 billion change in the fair value of convertible notes and related instruments.

FOR COMPUTE SUPPLIERS The filing itemises arrangements including at least $110.0 billion with Amazon, at least approximately $114.5 billion with Google, up to $120.5 billion with SpaceXAI and Broadcom lease rent of approximately $161.2 billion.

FOR RIVAL LABS The S-1 says compute is typically procured 12 to 18 months in advance of availability and reports compute and infrastructure at 48% of revenue in the quarter to 30 June 2026.

The Business Engineer Lens

This story maps onto the Business Engineer framework Tokenomics: The Economics of AI.

The framework starts from one unit: every conversation about AI in 2026, from capex to margins, “eventually collapses to the same atomic unit. The token.”

As we read it, the S-1 offers a company-level version of that ledger: revenue on one side and compute and infrastructure expense on the other, reported quarter by quarter. The framework was published in May 2026 and does not discuss this filing.

What Is Not Established

The quarterly figures are unaudited. Revenue per dollar of compute is our arithmetic, not a measure the S-1 reports, and compute and infrastructure expense is not the same as cash paid for compute. The S-1 does not report a gross margin. It leaves the total of contractual obligations blank, and the amounts in the commitments table are minimums, caps or lease rent, not forecasts of spending. Operating income excludes the change in fair value of convertible notes, which drove the net loss.

This Week In The AI Supercycle

Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls

This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.

Read the full issue →

Business Engineer Framework

Tokenomics: The Economics of AI

A Business Engineer framework on the token as the unit that ties AI capex, capacity and margins together.

Read the Map of AI →

The Bottom Line

Anthropic’s S-1 shows compute and infrastructure expense falling from 251% of quarterly revenue in early 2025 to 48% in the second quarter of 2026, when the company recorded operating income of $974 million. The same filing lists multi-year compute arrangements that include at least $110.0 billion with Amazon, at least approximately $114.5 billion with Google, up to $120.5 billion with SpaceXAI and Broadcom lease rent of approximately $161.2 billion.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. All figures are from Anthropic, PBC’s Form S-1 filed with the SEC on 26 October 2026; quarterly figures are unaudited, revenue per dollar of compute is our arithmetic, and blank amounts in the filing are reported as blank. Earlier reports are as covered in our pieces linked above. This is not investment advice. We did not contact Anthropic.

Sources: Anthropic, PBC Form S-1, filed 26 Oct 2026 (SEC EDGAR)

Scroll to Top

Discover more from FourWeekMBA

Subscribe now to keep reading and get access to the full archive.

Continue reading

FourWeekMBA