Anthropic’s Nearly $900 Billion Compute Bill: A Bet Seven Times Its Cash and Three Hidden Risks

On November 2, 2026, Anthropic submitted an amended S-1 filing for its IPO to the U.S. Securities and Exchange Commission. One set of figures set off a chain reaction in the industry: the company’s cumulative compute commitments had reached approximately $895.4 billion, equivalent to 6.8 times its cash and short-term investment reserves (about $131.9 billion). This is not a single-quarter bill but a matrix of purchase contracts spanning nearly two decades.

According to the filing, of the $895.4 billion, about $447.5 billion will come due by December 31, 2030; the remainder is spread between 2031 and 2045. About $69.8 billion matures within the next 12 months—nearly half of Anthropic’s current total liquid assets (about $146.6 billion, including a $14.6 billion credit facility).

One Quarter’s Profit, Drowned by Short-Term Compute Costs

The best entry point for understanding this bill is the profit reversal at Anthropic from Q2 to Q3. According to the amended S-1, the company posted operating profit of $974 million in Q2—its first profitable quarter in history; however, Q3 immediately recorded an operating loss of $567 million, a single-quarter net swing of more than $1.5 billion.

Anthropic directly identified the cause in the filing: the Q3 loss was “driven primarily by $5 billion in short-term compute commitments.” This $5 billion cost is far higher than an equivalent multi-year compute agreement. In other words, buying compute on the spot market temporarily is much more expensive than signing long-term contracts.

This detail reveals the financial logic behind the $895.4 billion commitment: locking in long-term agreements is a hedge. The company is betting that compute demand will persist over the next two decades and that locking in prices in advance will be more economical than purchasing quarter by quarter. If this judgment holds, the seemingly terrifying bill is in fact rational hedging; if it fails—whether because compute demand shrinks or new technologies disrupt the existing algorithmic paradigm—the cost of this commitment will be extremely heavy.

Supplier Matrix: Who Received This Bill

The $895.4 billion is spread among seven major suppliers, and the distribution itself is a power map of the AI compute ecosystem.

According to the S-1, Broadcom ranks first with $161.2 billion, all in the form of equipment leases. This arrangement points to Anthropic’s long-term strategy of co-developing custom AI accelerator chips (ASICs) with Broadcom—custom chips can be significantly superior to general-purpose GPUs in cost per unit of compute and energy efficiency, but they require deep binding at the hardware level. Second is Google, with commitments of at least $121.3 billion (covering the period from April 2026 to July 2033), but this figure excludes an additional $125 billion to $150 billion TPU compute plan, which was not included in the $895.4 billion because it has not yet formed a fixed-amount commitment. That is, the actual depth of Anthropic’s binding with Google may far exceed the figures directly presented in the filing.

SpaceX’s AI division, owned by Musk, ranks third with up to $120.5 billion, a large previously undisclosed cooperation. Amazon has committed at least $110 billion (through 2036), Microsoft at least $44.9 billion, Nscale up to $44.6 billion, and Lambda about $41.3 billion.

Beyond the above, there is an “off-balance-sheet account”: additional agreements, including further purchases of Google TPUs and a proposed transaction with AMD, amount to between $150 billion and $175 billion and are not included in the $895.4 billion statistics because no fixed amount has been set. Adding them up, Anthropic’s potential upper limit for compute purchase commitments could approach $1.1 trillion.

Revenue Side: Quarterly Growth and a 7% Compliance Risk

Whether the commitments can be honored ultimately depends on whether revenue can match them. The amended S-1 shows that Anthropic’s cumulative revenue in the first three quarters was $33.242 billion, of which Q3 alone brought in nearly $17 billion—meaning Q3 revenue was close to the sum of Q1 and Q2, showing strong acceleration. According to Fortune, despite this, Q3 growth slowed sequentially, while costs rose more steeply.

At the same time, the filing disclosed a structural revenue risk: Anthropic estimates that as much as 7% of revenue currently comes from users who violate platform usage rules. The company cautioned that ongoing compliance enforcement could have a material impact on future revenue growth. Seven percent is not a small number—based on first-three-quarter revenue, this means about $2.3 billion in potential revenue is in an uncertain state.

Industry Benchmark: Where $895.4 Billion Has to Be Placed to Be Perceptible

Viewed in isolation, $895.4 billion lacks a reference frame. According to TMT Finance, the combined capital expenditure guidance of the four major hyperscale cloud providers (Amazon, Google, Meta, and Microsoft) for 2026 is about $720 billion to $745 billion—a single-year figure. Anthropic’s $895.4 billion is a cumulative commitment spread from 2026 to 2045, averaging about $45 billion per year, roughly equivalent to the annual capital expenditure scale of a single mid-sized hyperscale cloud provider.

But there is a key difference: hyperscale cloud providers are both consumers of compute and builders and sellers of compute—the money they spend flows back in the form of cloud service revenue. Anthropic’s compute spending, by contrast, is almost entirely on the cost side. This makes Anthropic far more sensitive to the compute price curve than competitors that build their own infrastructure.

Independent Judgment

$895.4 billion is not cash that Anthropic will burn through in the near term; it is a portfolio of options that locks in future supply capacity. This strategy holds in two scenarios: first, AI compute demand continues to grow rapidly, and spot prices remain high or even rise; second, the unit-cost advantage of the custom chip route (Broadcom ASIC) is gradually realized.

However, three risks need to be tracked. First, the lesson of Q2’s fleeting profitability shows that the spread between short-term and long-term compute costs is the Achilles’ heel of Anthropic’s profitability—as long as this spread exists, the company is structurally forced to keep pushing commitments forward and cannot retreat. Second, hyperscale cloud providers are both Anthropic’s suppliers and its competitors (Amazon has Titan, Google has Gemini); these $110 billion and $121.3 billion purchase commitments are simultaneously a strategic lock on competitors’ future compute resources, but they also deepen the other side’s knowledge of its own roadmap. Third, the 7% compliance revenue risk has not yet been quantified into a specific enforcement timetable; if regulatory pressure accelerates, this gap will directly affect short-term solvency.

From the logic of capital markets, the essence of this bill is: Anthropic is using commitments close to seven times its entire existing liquid assets to declare to the whole market that “AI will not be a bubble”—and the judge of this bet is the compute economics of the next two decades.