Anthropic Bets USD 750 Billion on Future AI Demand, With Financial Markets at Risk if It Fails

Securing vast computing resources in anticipation of long-term growth
Expanding long-term contracts requiring payment even if usage falls
Slower growth could increase purchase burdens and hinder investment recovery

Anthropic has placed a huge bet on the AI market’s long-term growth by securing vast computing resources in advance. Buoyed by surging revenue, it has accumulated long-term purchase commitments, but many contracts require payment even if usage declines, creating a structure in which a growth slowdown would increase its financial burden. Suppliers are even lending customers money to fund purchases, raising the risk that failed investments could send shock waves through business partners and creditors. With aging equipment shortening the window for recouping investment and China’s DeepSeek intensifying its low-cost push, uncertainty is growing over whether Anthropic can generate enough earnings to support its enormous spending.

Long-Term Purchase Commitments Weigh Heavily Despite a 15-Fold Revenue Surge

According to The Wall Street Journal (WSJ) on the 22nd (all dates hereafter are local time), Anthropic is pursuing plans to secure up to USD 750 billion (approximately USD 773 billion) in AI computing resources over the coming years. A draft registration statement for Anthropic’s confidential initial public offering (IPO), reviewed by the WSJ, outlines plans to secure vast computing capacity through cloud services and semiconductor leasing. Infrastructure procurement obligations under contracts already signed exceed USD 600 billion (approximately USD 619 billion). Deals currently under discussion could add another USD 150 billion (approximately USD 155 billion) if completed. The plan involves taking on enormous payment obligations to secure computing resources that customers will use over the coming years. Given the long investment recovery period, the company’s financial capacity appears to hinge on the accuracy of its assessment of future demand.

The scale of the investment in future demand is substantial even considering the recent surge in revenue. Revenue in the first half of this year reached USD 16.3 billion (approximately USD 16.8 billion), 15 times the USD 1.1 billion (approximately USD 1.13 billion) recorded a year earlier. Third-quarter revenue disclosed to investors last week was USD 16.8 billion (approximately USD 17.3 billion), up 45% from the previous quarter. In just three months, the company generated roughly 3.7 times last year’s full-year revenue of USD 4.6 billion (approximately USD 4.74 billion). Against this backdrop of rapidly rising revenue, Anthropic signed large purchase commitments to be fulfilled over the next 7–10 years. It is expanding its business by locking in future spending on the assumption that its current growth will continue for a long time.

However, many commitments make it difficult to cut spending even if demand falls short of expectations. Anthropic’s draft registration statement outlines plans to spend at least USD 518 billion (approximately USD 534 billion) with six partners over the next 10 years. About 80% of those commitments were noncancelable or required payment regardless of usage. Minimum spending commitments to major companies amount to USD 111.1 billion (approximately USD 115 billion) for Google, USD 110 billion (approximately USD 113 billion) for Amazon, and USD 31.4 billion (approximately USD 32.4 billion) for Microsoft. Under its agreements with Google and Amazon, Anthropic must pay the difference if actual usage falls below the committed amount. Lease obligations for Broadcom-related equipment also total USD 161.2 billion (approximately USD 166 billion) and cannot be canceled unless contractual exceptions, such as default, apply. Because Anthropic must pay a substantial portion of the promised amounts even if it cannot fully utilize the capacity it has secured, it will bear the burden if its demand forecasts prove wrong.

Table 1. Anthropic’s AI Computing Resource Procurement Plans and Purchase Commitments

Computing Resource Costs Consume Half of First-Half Revenue

As long-term payment obligations accumulate, immediate infrastructure spending is also rising rapidly. Computing and infrastructure spending in the first half of this year totaled USD 9 billion (approximately USD 9.28 billion), four times the level a year earlier. In just six months, spending surpassed last year’s full-year total of USD 7.3 billion (approximately USD 7.53 billion) and amounted to more than half of revenue of USD 16.3 billion (approximately USD 16.8 billion) over the same period. The first-half net loss was USD 80.7 billion (approximately USD 83.2 billion), but USD 72 billion (approximately USD 74.3 billion) of that consisted of noncash accounting charges reflecting changes in the value of financial instruments such as convertible bonds. Valuation charges stemming from the company’s rising valuation sharply increased its reported loss. Separately from those valuation charges, which involve no cash outflow, actual spending to secure and operate computing resources is climbing steeply.

As infrastructure spending surged, Anthropic also entered into an agreement to borrow a substantial sum from a computing resource supplier. Earlier this month, Broadcom agreed to lend up to USD 42 billion (approximately USD 43.3 billion) to support Anthropic’s infrastructure spending. Broadcom can designate a separate financing partner, and the debt securities used to raise the funds are structured to be convertible into Anthropic shares in the future. If fully drawn, the facility could cover roughly one-third of the USD 125.2 billion (approximately USD 129 billion) committed to leasing tensor processing unit (TPU) computing capacity over five years. Semiconductor supply, equipment leasing, and financing are thus bundled together within the Broadcom deal. The registration statement stipulates that breaches of certain payment or contractual performance obligations could require immediate payment of a substantial amount of lease charges and could also restrict the use of the loan facility to cover them.

U.S. Treasury Secretary Remains Optimistic Despite Warnings of Cascading Losses

The transaction structure, in which financial difficulties at Anthropic could lead to losses for suppliers and creditors, also intersects with the debate over risks across AI investment. In its Global Financial Stability Report this month, the International Monetary Fund (IMF) analyzed 14 major AI-related deals since 2023 involving USD 300 billion (approximately USD 309 billion) in equity investment and USD 650 billion (approximately USD 670 billion) in debt. The companies receiving the funding committed to USD 1.4 trillion (approximately USD 1.44e+3 billion) in future service and semiconductor purchases. Because investors also count the companies they finance as customers, a deterioration in those customers’ businesses could simultaneously undermine investment recovery and revenue. The IMF warned that if a key investee company falters, the value of equity holdings and loan claims could decline while expected revenue from that company could also be impaired. This means that if Anthropic cannot make its promised payments, both the companies supplying its computing resources and the creditors lending it money could be hit.

U.S. Treasury Secretary Scott Bessent also weighed in after the IMF warned of cascading repercussions from troubled AI investments, but his assessment of the risks to the financial system differed. At a private investor event in New York on the 21st, Bessent said of the AI investment boom, “I don’t see a systemic issue.” Referring to Pets.com, which went public just before the dot-com bubble peaked in March 2000, he also said, “This is not Pets.com.” His point was that the AI boom differs from the dot-com bubble, when money poured into companies with little substance. He cited the investment capabilities and business track records of executives leading major technology companies as grounds for his assessment. He explained that executives who built some of the world’s most profitable companies believe in AI’s potential while keeping most of their wealth invested in their own companies. He also expressed expectations that AI would boost productivity and contribute to U.S. economic growth.

Halving Equipment Lifespans Would Require USD 6 Trillion in Annual Revenue

However, for the technology companies’ record of success highlighted by Bessent to be repeated in AI investment, earnings must rise enough to cover equipment spending already committed. A Columbia University study submitted to a Brookings Institution academic conference last month estimated U.S. AI infrastructure investment from 2025 to 2032 at USD 10.3 trillion (approximately USD 13,623 billion). Assuming operating cash flow equals 50% of revenue, the study calculated that annual revenue would need to reach USD 3.7 trillion (approximately USD 3.82e+3 billion) in 2032 to achieve a 10% unlevered return on investment. That scale would require growth in both demand sufficient to fill the expanded capacity and revenue from usage fees. The study noted that investment is exceeding companies’ ability to finance it internally, adding debt burdens to businesses whose profitability has not yet been sufficiently proven.

There is also a risk that equipment could lose its competitiveness before the investment is recouped. When the same study shortened the economic lifespan of information technology equipment from six years to three, the annual revenue required to achieve the same return rose to USD 6 trillion (approximately USD 6.18e+3 billion). Additional costs would arise if new semiconductors and cooling technologies required existing facilities to be replaced or retrofitted. The shorter the equipment replacement cycle, the less time there is to recover the investment. The study noted that principal and interest repayment obligations on debt used to finance facilities are fixed, meaning that a combination of slowing demand, operational delays, and falling collateral values could amplify investor losses and hinder debt recovery.

Price Competition From Chinese AI Companies Poses Another Risk

The low-cost push by Chinese AI companies is another major variable. On the 20th, DeepSeek open-sourced DeepInfer, the inference engine used to generate responses from its models. It is designed to reduce unnecessary computation and chip idle time, and supports the latest AI chips from both Nvidia and Huawei. Its focus is on lowering response-generation costs by processing more requests on the same equipment. DeepSeek also adopted the Apache 2.0 license, which permits commercial use, allowing outside companies to modify the code and incorporate it into their own services. If technologies that reduce operating costs spread and lead to lower service prices, U.S. AI companies seeking to recoup massive infrastructure investments through usage fees could face greater pressure on their earnings.

The financial resources supporting DeepSeek’s low-cost push are also growing. DeepSeek is reportedly finalizing a USD 15 billion (approximately USD 15.5 billion) funding round, with Chinese cloud service provider Tencent and battery maker CATL among its major investors. With more than 15 companies participating, including Meituan and Geely Auto, the amount raised has grown to twice the original target. Expansion of the computing infrastructure on which the funds will be spent is also underway. DeepSeek is increasing investment in computing infrastructure within China and working with Huawei to expand model training using Chinese-made chips. It is pursuing U.S. AI companies by combining the release of technology that lowers inference costs with large-scale fundraising and expanded training infrastructure.