Anthropic Locks in Massive Compute Capacity
According to a report from The Information, Anthropic has signed compute contracts worth as much as $517 billion over an eleven-month stretch. That figure is striking not only for its scale, but for the speed at which the artificial intelligence company has accumulated commitments. Since October 2025, Anthropic has secured at least 14.8 gigawatts of additional computing power, building on an existing base of roughly one to two gigawatts. The company is also planning to develop its own data center portfolio, a sign that its ambitions extend far beyond renting capacity from third-party cloud providers.
The contractual commitments highlight the extraordinary race currently underway among leading AI developers. However, Anthropic's total planned capacity is still expected to trail OpenAI's stated target of 30 gigawatts by 2030. The Information adds a note of caution about direct comparisons: many of Anthropic's contracts extend well beyond 2030, so the precise capacity trajectory remains difficult to map onto a single point in time.
A Shift That Underscores Competitive Pressure
The deal-making marks a notable reversal for Anthropic's leadership. In early 2026, CEO Dario Amodei warned against the pace of investment he observed among competitors. He argued that those rivals "don't really understand the risks they're taking," raising fears that poorly managed expansion could create systemic problems for the AI ecosystem. Now, just months later, Anthropic is itself moving aggressively to secure compute resources at a historic scale.
This shift illustrates how the dynamics of the AI industry can quickly override even the most cautious public stance. When a competitor is building out enormous infrastructure to support increasingly capable models, slower developers risk being left behind, both in research and product deployment. Anthropic's recent actions suggest the company concluded that the cost of inaction now exceeds the risk of over-committing to infrastructure.
The Financing Reality: Revenues Are Not Enough
None of these commitments can be covered by current revenue, and the gap remains wide. Anthropic's annualized revenue has climbed past $65 billion, according to Bloomberg. OpenAI, for its part, reported more than $40 billion in annualized revenue as of July. Even with these figures, neither company can fund multi-hundred-billion-dollar compute programs from operating cash flow alone. Debt financing, investor partnerships, and special-purpose vehicles are likely to play central roles in bridging the gap.
The sheer size of these numbers raises questions about the long-term economics of frontier AI. Compute contracts are typically multi-year obligations, and once signed, they can be difficult to unwind. If a company overestimates its future needs, or if more efficient hardware and algorithms lower demand, those commitments could turn into significant liabilities. This is a risk that executives like Amodei and Altman have acknowledged, albeit with different rhetorical emphasis.
Amodei's Warning vs. Anthropic's Strategy
Amodei's warning, delivered in early 2026, appears to have been aimed at competitors racing to build out "super clusters" of data centers and rack up unprecedented capital expenditures. He suggested that many of these projects were being undertaken without a clear-eyed assessment of their technical, financial, and operational risks. The implicit message was that a more measured approach would ultimately produce better outcomes, both for individual firms and for the broader AI field.
Yet Anthropic's reported 14.8 gigawatts of new capacity is hardly the signature of a cautious company. In fact, the scale is comparable to major national power grids. To put the number in perspective, 14.8 gigawatts is enough to power millions of homes, or a dense urban region. Redirecting that amount of electricity toward AI training and inference represents a profound change in how energy is consumed by the technology sector. Governments and utility providers will need to expand grid capacity and generation sources to accommodate these facilities, while data center builders will need to tackle heat, water, and interconnection challenges.
Altman's Counter-Message: Questions of Timing
Meanwhile, OpenAI's Sam Altman has begun urging caution in the other direction. He warns of "unsustainable silliness" from so-called neo-cloud providers, a reference to a new wave of companies raising enormous capital allocations to build AI-specific data centers. Altman suggests these buildouts may be based on unrealistic expectations about future demand or about the persistence of today's technical bottlenecks.
Altman's underlying concern is that technological progress could advance so quickly that today's expensive, purpose-built infrastructure becomes outdated or underutilized sooner than expected. For example, if breakthroughs in model efficiency, algorithmic optimization, or specialized chips dramatically reduce the compute needed to achieve frontier performance, some of the on-paper capacity being ordered today may never need to be switched on. Amodei and Altman both recognize this possibility, but their public statements frame it differently: while Amodei worried that competitors were reckless in their speed, Altman now seems worried that the market as a whole, including some infrastructure financiers, is getting overexcited.
Strategic Implications for Anthropic
Why did Anthropic go from warning against rapid investment to joining the race itself? One likely explanation is the necessity of compute scale for model training runs. As frontier labs push toward models measured in the trillions of parameters, training runs require clusters that can operate cohesively, with vast memory and energy budgets. Without contracted capacity far in advance, labs risk waiting years for available fraction of cloud capacity or paying premium rates.
Anthropic's own planning now includes building dedicated data centers, which gives the company more control over hardware design, network topology, cooling, and power supply. This vertical-integration trend is reminiscent of the approach taken by companies such as Google and Microsoft, which have moved from buying generic servers to designing custom chips, racks, and even entire data center campuses. For a firm like Anthropic, controlling the stack could mean the difference between a training run that takes weeks instead of months, and ultimately that determines how quickly a model can be iterated.
Hidden Complexity in a $517 Billion Figure
The $517 billion headline masks significant nuance. Much of this budget is likely not yet locked into firm, all-inclusive contracts; it may reflect options, letters of intent, and framework agreements that anticipate future reservation. Some of the reported capacity may be tied to longer horizons, well past 2030, so the annual cost is not as staggering as it appears at first glance. Still, even a fraction of that committed amount arriving by 2030 would represent one of the largest capital programs in corporate history.
The Information's reporting suggests that Anthropic's capacity plans likely remain below OpenAI's 30-gigawatt target, but that may change as contracts are revised or expanded. Because many of Anthropic's agreements are structured for longer terms, the company may end up with more building capacity available in the mid-2030s depending on how those facilities are staged. The race, then, may not be about 2030 alone; it could be about the trajectory that extends for at least a decade into the future.
Wider Market Fragility
Altman's critique of neo-cloud providers also points to structural fragility in the infrastructure financing market. These new providers -- often started by former cloud executives and energy traders -- are betting that they can construct dedicated data centers, sign anchor tenants such as OpenAI or Anthropic, and then sell capacity to smaller AI developers once the frontier labs move to newer generations. If the anchor contracts don't cover costs, or if demand from smaller tenants doesn't materialize, the financial fallout could be broad.
For established cloud players, the huge level of compute ordering also changes the competitive balance. For example, if Anthropic and OpenAI both secure wave after wave of capacity, other AI labs -- or even the same labs' partners -- could face scarcity and price rises. The national security dimension adds another layer of urgency: governments now treat advanced compute as a strategic resource, and export controls or domestic energy policy can shift the economics of data center location overnight.
Looking Forward: Reckless Risk or Necessary Hedging?
Dario Amodei's original warning reflected a belief that acting too fast can be dangerous. In the volatile world of frontier AI, however, moving too slowly can be just as dangerous. Anthropic's decision to sign these contracts can be seen as a hedge against the downside of missing out on future compute. The tech industry has a long history of companies making expensive bets that look foolish at first, but which prove essential when demand accelerates unexpectedly.
At the same time, Amodei and Altman are both signaling an awareness that something is not entirely healthy about spending fractions of a trillion dollars before the technology has reached its final form. The next several years will test whether giant, long-dated infrastructure bets can deliver enough innovation to justify their enormous costs. For now, Anthropic is a buyer of compute at historic scale; its leaders' previous warnings now serve as background context for a deal spree that rivals the largest energy and defense investments in history.
The story is far from over. As new model presentations arrive, and as more details emerge about the structure of these contracts, the big question will shift from whether Anthropic can sign the deals to whether its multi-billion-dollar commitments turn out to be strategically brilliant, or simply the same "reckless risk" that Amodei perceived in others.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com
