Anthropic’s Quarter: $17 Billion In, $9.9 Billion Out to Compute

The amended S-1 puts third quarter compute at 58% of revenue, with $895 billion of purchase commitments against an $81 billion run rate

Anthropic’s amended prospectus, filed on 2 November 2026, added one quarter to the numbers investors had already seen. The quarter arrived with $16.96 billion of revenue, up from $11.56 billion in the previous three months and up 1,127% from a year earlier. It also arrived with an operating loss of $567 million, which reversed the $974 million operating profit the company had reported for the second quarter.

The swing is small next to the revenue. It is not small next to the story. Anthropic went into its listing as the AI lab with the cleanest business model, the one selling to enterprises rather than chasing consumers. A quarter that flips from profit to loss because of how a company bought its electricity is a different kind of fact about that business.

The filing is unusually specific about why. Of the compute delivered in the three months to September 2026, $5.0 billion came through short-term commitments, which the company says cost significantly more than comparable capacity bought under multi-year contracts. Add $2.0 billion of non-cash charges tied to its employee donation matching programme, and the operating line goes negative.

The quarter flipped on $5 billion that arrived late

Anthropic explains the short-term buying as a timing failure, not a strategy. Capacity under long-term contracts is typically signed 12 to 18 months before it shows up. When customer demand runs ahead of that pipeline, the company either waits or pays the spot price. In the third quarter it paid.

The premium it paid is not disclosed in dollars. What is disclosed is the effect on the ratio that matters: revenue generated per dollar of compute fell to 1.72x, after peaking above 2x in the previous quarter. If the short-term capacity had cost the same as contracted capacity, the operating line would have been positive.

Two details suggest this is manageable rather than terminal. Adjusted operating income, which strips out share compensation and the donation charge, rose to $2.46 billion from $2.23 billion, though the margin fell to 15% from 19%. And the customer base is broadening: 79% of third-quarter revenue came from outside the five largest customers, up from about 64% in the first quarter. The company also reported more than 11,900 customers spending over $100,000 a year, with a net revenue retention rate near 900%.

A business with that retention can absorb a bad quarter of procurement. It cannot absorb a structural rise in the price of the input that dominates its cost base.

Compute went from 251% to 48% to 58%

The filing puts the whole trajectory in one series. Compute and infrastructure spending was 251% of revenue in the first quarter of 2025, when the company was small and building ahead of demand. It fell to 48% by the second quarter of 2026, as revenue grew faster than committed capacity. It rose again to 58% in the third quarter.

That last move is the one investors will read twice. At the exact moment a company normally shows improving unit economics before a listing, Anthropic showed the opposite. The filing does not hide it. It says the company expects compute and infrastructure to remain its largest operating expense for the foreseeable future, in absolute dollars. It adds: “We expect our use of short-term compute to continue to increase in the near term and to in turn compress our operating margin.”

Table 1: The quarterly ladder disclosed in the amended filing

Measure Q3 2025 Q2 2026 Q3 2026
Revenue ($bn) 1.38 11.56 16.96
Compute and infrastructure ($bn) not disclosed 5.55 9.86
Compute as share of revenue not disclosed 48% 58%
Operating result ($m) not disclosed +974 -567
Adjusted operating margin not disclosed 19% 15%

Source: Anthropic Form S-1/A, 2 November 2026, as reported by WebProNews, with second quarter compute dollars derived from the disclosed ratio

The 58% is the number to hold on to. It means that for every dollar a customer pays Anthropic, 58 cents goes out the door to compute and infrastructure before a single salary, sales commission or research project is funded. The company still earns an adjusted margin, because the compute number includes depreciation and prepaid capacity timing. But the gross spread is thin for a software business, and it is the spread that decides what the company is worth.

Eleven years of revenue sits in the commitments note

The commitments section is where the filing stops describing a software company. Anthropic had signed purchase and lease agreements worth about $895.4 billion by 30 September 2026, substantially all of it for compute and infrastructure. Of that total, roughly $69.8 billion falls due within 12 months of 30 September 2026, and about $447.5 billion falls due before the end of 2030.

Put those against the run rate. The November filing reports annualised revenue of about $81 billion. That means the total commitment equals roughly 11 years of revenue at the current pace. The near-term slice alone, $69.8 billion, is about 86% of a full year of revenue at that run rate.

The company is not insolvent. It ended September 2026 with $146.6 billion of liquidity, including $95.6 billion in cash and $36.3 billion in short-term investments, and it has up to $77 billion of note financing capacity it has not drawn. Nine months of operating cash flow came in at $6.6 billion, which is positive and growing.

But the commitments have a shape that matters. For every dollar of revenue the business generates this year, it has already promised more than a dollar of future compute spending. If the 58% intensity holds, absorbing the $447.5 billion due by 2030 requires roughly $772 billion of cumulative revenue over the same period. That is nearly ten times the current annual run rate, compounded over four years.

That is not a prediction of failure. It is the arithmetic that the IPO price has to justify.

The nine month view behind the quarter

One quarter can mislead, so the nine month column is worth reading first. Revenue for January to September 2026 reached $33.2 billion, against $2.5 billion in the same period of 2025. The operating loss for those nine months was $1.26 billion, which means the business was close to break even across three quarters and lost money in only one of them.

The net loss tells a different story, and almost none of it is operational. Anthropic reported a net loss of $104.04 billion for the nine months, including a $23.29 billion net loss in the third quarter alone, of which $21.9 billion came from marking convertible notes and similar instruments to a higher valuation. That is an accounting consequence of the company becoming more valuable, and it will not repeat once the notes convert.

The cash flow statement is the most useful part for anyone outside the company. Operating activities generated $6.6 billion over the nine months, up from a $2.0 billion outflow a year earlier. Investing activities consumed $36.5 billion, mostly short-term investments, lease prepayments and capital expenditure. Financing brought in $110.7 billion, overwhelmingly preferred stock. In plain terms, the business now funds its own operations and still needs investors to fund its buildings.

That split is why the listing matters more than the loss. Anthropic’s operating cash flow covers the lights. It does not cover $447.5 billion of commitments due by 2030, and equity is the only instrument large enough to close that gap.

Renting at the top of the market

Where the compute comes from explains the price. Anthropic has been paying SpaceX $1.25 billion a month for capacity through May 2029, alongside multi-billion dollar deals with Nscale, Akamai and Volta. These are neoclouds rather than hyperscalers: smaller operators that sell flexible capacity at higher rates, sometimes in buildings they lease themselves.

The trade is speed for margin. A hyperscaler contract priced 18 months ahead is cheap and inflexible. A neocloud contract is expensive and can be switched on in weeks. When demand doubles in a quarter, the second option is the only one that works.

The awkward part is that the largest cloud providers are also becoming competitors. Reporting on the S-1 said Anthropic had struggled to secure enough capacity from AWS, Google and Microsoft, which increasingly view the Claude maker as a rival rather than a customer. The company’s own filing warns that the third-party data centre operators it leases from have a limited operating history.

Turbines are the same decision in physical form. In November 2026 Anthropic agreed to buy about $3.4 billion of Caterpillar gas generators, capable of producing up to 1.7 GW, from equipment that the supplier OnePWR had ordered years earlier and could not finance. Paying for someone else’s abandoned order is how you skip the queue for grid connection.

Anthropic Direct is the bet that fixes the margin

The filing names the mechanism the company intends to use to repair its compute margins: Anthropic Direct, a programme that leases chips and hosts them in dedicated facilities. The logic is straightforward. Instead of buying capacity from a neocloud at spot rates, the company takes the equipment and the building on long leases and operates them itself, which converts an expensive variable cost into a cheaper fixed one. The filing says the programme will let it backfill expensive short-term capacity as its own infrastructure comes online.

That is the same bet every large lab is making, and it carries the same risk. A fixed lease is cheaper per GPU hour only if utilisation stays high for the length of the lease. If demand softens in 2028, the company is left paying for capacity it does not need, which is exactly the position the short-term purchases were meant to avoid. The difference between a good and a bad outcome is a utilisation forecast that nobody has published.

The product mix pushes in the opposite direction. In November 2026 alone, Anthropic shipped Claude Meetings, which turns a live discussion into a plan and a task list, and a personal finance feature that connects Claude to bank accounts through Plaid. Both are agentic: they run continuously, call tools and keep context. Work of that shape consumes far more compute per user than a chat window, which means the intensity number is more likely to rise than fall.

The turbines cost $2,000 per kilowatt

That deal is worth pricing per unit, because it shows what urgency costs. For delivery from 2027, $3.4 billion for 1.7 GW works out to about $2,000 per kilowatt of generating capacity, for equipment only, before installation, fuel and grid interconnection.

New combined cycle gas plants in the Gulf are being built for roughly $1,150 per kilowatt of installed capacity, based on the $3 billion budget for the 2.6 GW Taweelah C project in Abu Dhabi. The comparison is not like for like, since Anthropic is buying simple cycle generators rather than a combined cycle plant, and simple cycle costs less per kilowatt to build but burns more fuel per megawatt hour.

Even allowing for that, the premium is for time. The global backlog for gas turbines runs into the early 2030s, and grid connection queues are just as long. A company that expects to be serving $81 billion of annual demand cannot wait five years for a slot, so it pays for somebody else’s.

The fuel bill is the part that stays. Simple cycle generators run at roughly 35% thermal efficiency, against 60% for a combined cycle plant. Every megawatt hour from those turbines costs more in gas than the same hour from a modern plant, and gas prices in some US markets are forecast to more than triple by the end of the decade. The $3.4 billion buys the capacity. It does not fix the operating cost.

What has to be financed before 2030

Bring the pieces together and the shape of the company becomes clear. In the twelve months to September 2026, Anthropic turned a research lab into a business with an $81 billion run rate, 11,900 customers above $100,000 a year and a retention rate near 900%. In the same period it signed commitments equal to 11 years of that revenue, borrowed nothing, and funded itself with preferred stock.

That combination works while equity markets are willing to price growth. It becomes uncomfortable the moment a quarter of margin compression coincides with a credit rating that limits borrowing. The $567 million loss in the third quarter is a detail. The 58% compute intensity and the $69.8 billion due within a year are the terms of the business.

What the listing prices

The IPO terms put the company at $70 to $78 a share, or roughly $1.27 trillion to $1.45 trillion on a fully diluted basis, with 825 million shares on offer and a separate $26 billion private placement involving Google, Microsoft, Nvidia, AMD, Broadcom and Micron. The shares are expected to start trading on Nasdaq on 19 November under the symbol ANT.

At the November 2026 run rate of $81 billion, a $1.4 trillion valuation is about 17 times sales. That is a growth multiple, not a margin multiple, which tells you what buyers are being asked to believe: that revenue keeps compounding at a rate that lets a company with a 15% adjusted margin and 58% compute intensity grow into its commitments.

Credit agencies are less romantic. The Financial Times reported that Anthropic is expected to receive a high junk rating after listing, which would make future borrowing more expensive. A company that has promised $895 billion of compute spending and carries a junk rating has one source of funding left at scale: its own equity.

The filing also shows the price of competing on price. Reductions introduced with new model versions and contractual discounts cut about $600 million from third-quarter revenue in 2026, roughly 3.5% of the total. Discounting to hold share against OpenAI and Google is a choice the company can make while it is raising money. It is harder to sustain once the money is raised and the commitments come due.

Three blanks left in the filing

Three things would change the picture, and none of them is disclosed yet.

  • The premium Anthropic pays for short-term compute, in dollars per GPU hour. Without it, nobody outside the company can model how much of the margin compression is temporary.
  • The mix between owned capacity and rented capacity by 2028. Owned capacity carries depreciation and capital cost; rented capacity carries the neocloud margin. The two produce very different margins on the same revenue.
  • The run rate as of 31 October, which the amendment left blank. Every valuation comparison in the market depends on that number and the filing stops at 30 September.

Until those arrive, the defensible reading is narrow. Anthropic has built a business that grew revenue 1,127% in a year, retains customers at a rate software companies rarely see, and just discovered that its largest cost is also its least predictable. The $567 million loss is small. The $895 billion commitment is not.

Computation

Data as of November 2026. All figures come from Anthropic’s amended Form S-1 filed with the SEC on 2 November 2026, as reported by WebProNews, FourWeekMBA and Caixin, or from the original S-1 filed on 26 October. Second quarter compute spending of $5.55 billion is derived by applying the disclosed 48% ratio to disclosed revenue, because the filing gives the ratio rather than the dollar figure. The $772 billion cumulative revenue estimate divides the $447.5 billion commitment due by 2030 by the disclosed 58% compute intensity and holds that intensity constant, which is an assumption, not a forecast. The $2,000 per kilowatt figure divides the $3.4 billion turbine purchase by 1.7 GW of capacity and excludes installation and fuel. Taweelah C’s $1,150 per kilowatt uses its $3 billion budget and 2.6 GW nameplate.

Sources

  • WebProNews 🔗, Anthropic brought in nearly $17 billion and slipped back to an operating loss, 2 November 2026
  • FourWeekMBA 🔗, twelve new numbers in Anthropic’s S-1/A, 2 November 2026
  • TechCrunch 🔗, Anthropic is still grappling with the economics of compute, 3 November 2026
  • TechCrunch 🔗, Anthropic buys $3.4B worth of turbines, 9 November 2026
  • TechCrunch 🔗, Anthropic expected to receive a high junk rating, 30 October 2026

About Global Infinity Labs

Global Infinity Labs is an AI innovation company specializing in MCP data services. This article was prepared by our expert team using insights and data provided through our MCP platforms.

Contact: support@globalinfinitylabs.com