Anthropic's Infrastructure Bet: $19B Leases, a Rival's Compute, and a Trillion-Dollar IPO Countdown
A week in which SpaceXAI commanded 3-5x premium pricing for gigawatt-scale AI compute has clarified both the opportunity and the concentrated risk facing Anthropic as it rushes toward a $965 billion public listing backed by $19 billion in long-term infrastructure contracts.
SpaceXAI's unveiling this week of near-immediate gigawatt-scale AI data center capacity — carrying annualized rental revenue of $2.8 billion at per-megawatt rates that reportedly run three to five times higher than traditional multi-year cloud commitments — sent an unmistakable signal about where the AI infrastructure market is heading. Purpose-built, dedicated compute now commands a structural premium that generic hyperscaler contracts cannot match. For Anthropic, arriving at that conclusion is no longer academic: in the span of two weeks, the company has signed what is reportedly the largest AI infrastructure contract on record, announced hundreds of billions in future commitments, filed a confidential IPO prospectus at a $965 billion valuation, and — according to multiple reports on July 18 — entered negotiations to lease $10 billion in computing power from Meta, a direct competitor. The velocity and the dollar figures together mark a company in a definitive, perhaps irreversible, scale-up moment.
The anchor of Anthropic's buildout is the 20-year, $19 billion lease signed July 7 with TeraWulf for a 401-megawatt campus at a former aluminum smelter in Hawesville, Kentucky, with the facility slated for 2027 commissioning. TeraWulf subsequently announced it would raise $3.5 billion in debt financing, led by Morgan Stanley, to fund construction — a capital raise that itself signals how bond markets are pricing long-duration AI demand. Anthropic separately announced plans for up to $15 billion in Australian data center investment targeting 1.4 gigawatts, and an overarching $50 billion U.S. infrastructure commitment spanning Texas and New York. Taken together, these obligations dwarf Anthropic's earlier reported revenues by a large multiple, a disproportion that reflects both the capital intensity of frontier model training and the forward-looking logic of locking in dedicated capacity before per-megawatt costs inflate further — exactly the dynamic SpaceXAI's premium pricing now illustrates in market-rate terms.
Against this infrastructure push, Anthropic's path to public markets is accelerating in parallel. The company filed a confidential draft S-1 in early July, targeting a Nasdaq listing as early as October, with Fortune reporting annualized revenue of $47 billion — which, if accurate, would represent the first time Anthropic has outpaced OpenAI on that metric. The reported $10 billion compute deal with Meta, if completed, would add a significant recurring revenue line while also securing chip access ahead of the IPO: a structural hedge against supply-chain narratives that tend to unsettle pre-listing due diligence. The arrangement — with Meta simultaneously playing competitor and infrastructure supplier — underlines how industrial logic in AI can override competitive instincts. Investor meetings are reportedly being arranged within weeks, and Polymarket's prediction markets were pricing an 84.5% probability of Anthropic reaching a $1 trillion valuation within 18 months of the TeraWulf deal being announced.
The risks, however, are neither small nor speculative. Analysts noted in mid-July that Anthropic, OpenAI, and DeepSeek are facing accelerating price compression in AI inference as the market commoditizes rapidly. The more pointed challenge comes from Zhipu AI's Kimi K3, which reportedly delivers performance on par with Anthropic's Opus 4.8 while having been developed by a team of around 300 — echoing the disruption pattern established by DeepSeek in January and raising fresh questions about whether Western compute advantages, already squeezed by export controls, translate into durable capability leads. Former White House economists warned in July that the AI infrastructure investment cycle is a bubble "still inflating," linking Anthropic's and OpenAI's capital commitments to systemic overinvestment risk. On the geopolitical side, Alibaba banned Anthropic's Claude Code tool, alleging a hidden China-detection mechanism — a claim Anthropic has not publicly addressed, but one that, if substantiated, would meaningfully constrain the company's addressable market across Asia.
What emerges is a company at the intersection of formidable commercial momentum and concentrated structural risk. A reported $47 billion ARR, apparently ahead of OpenAI, is a genuine milestone if sustained; but the infrastructure obligations that underpin that trajectory carry terms spanning two decades at a time when model architectures, inference economics, and competitive dynamics are evolving on a monthly basis. The Meta compute arrangement is pragmatic de-risking, but it also indicates that Anthropic cannot yet self-supply at the scale its growth requires. Three signals will clarify the picture over the coming months: whether IPO book-building confirms that public markets will price a near-trillion-dollar valuation for a company still deploying capital at this rate; whether the TeraWulf Kentucky facility and the Australian projects break ground on schedule, validating the build thesis; and whether Anthropic's next model generation maintains a measurable capability gap over rivals like Kimi K3 — the entire investment logic of the compute arms race depends on training dollars continuing to widen that gap rather than narrowing it.