Anthropic's $45 Billion Compute Deal With Nscale Marks Infrastructure-Led Path to IPO
A $45 billion computing agreement with Nscale, deploying Nvidia Vera Rubin chips in West Virginia, is the largest single expression yet of Anthropic's strategy to lock in compute capacity ahead of a reported October Nasdaq listing.
On August 27, reports emerged that Anthropic had reached a $45 billion AI computing agreement with Nscale, a cloud infrastructure provider that will deploy Nvidia's Vera Rubin accelerators at a new campus in West Virginia. The scale is arresting: for an AI lab, it is one of the largest single compute procurement commitments ever disclosed publicly. The timing is deliberate. Anthropic confidentially submitted a draft S-1 to the SEC in mid-August, targeting a Nasdaq listing in October 2026 at a reported valuation of approximately $965 billion — with some secondary-market projections and analyst commentary placing the figure closer to $2 trillion. Multiple outlets framed the Nscale agreement as part of a broader computing push preceding the offering, underscoring the company's posture of treating infrastructure at scale simultaneously as a competitive moat and a public-market narrative.
The Nscale deal is not an isolated move. Over the preceding two weeks, Anthropic anchored or was linked to a cascade of infrastructure commitments across the eastern United States. Riot Platforms announced a $9.1 billion compute lease covering 191 megawatts at its Rockdale, Texas campus — though CryptoSlate subsequently reported that Riot's bridge financing matures before rent payments are scheduled to begin, a financial structure that will draw scrutiny from analysts assessing counterparty risk. The Kentucky Public Service Commission approved a 482-megawatt power agreement tied to TeraWulf's Anthropic-designated campus, and Bitdeer disclosed a $4.7 billion data-center lease connected to the lab. Singapore's GIC and Macquarie formed a vehicle called Theseus Infrastructure specifically to develop U.S. data centers for Anthropic's account. Taken together, these announcements describe a hub-and-spoke model of distributed compute capacity rather than concentration in a single facility — an approach that diversifies supply-chain exposure and simultaneously provides regulatory and political surface area across multiple states ahead of what would be a high-profile public offering.
What gives Anthropic's infrastructure ambition credibility with institutional investors is a substantive commercial inflection in the underlying business. The company disclosed in late August that its second-quarter 2026 revenue surpassed OpenAI's for the first time, and that it achieved positive operating income in the same period — a milestone for a lab defined historically by its training expenditure. The revenue base was built from a standing start: Anthropic launched Claude commercially in early 2023, secured an initial $300 million investment from Google the same year, and raised successive rounds through 2025 before completing an H-round at a reported $65 billion on a $965 billion valuation ahead of the S-1 filing. The profitability turn, arriving as the IPO clock began, converts the offering from a pure growth narrative into one with at least a quarter of demonstrated operating leverage — a distinction that meaningfully affects how public-market investors are likely to price the deal.
Running in parallel to the external buildout is a structural bet on reducing purchased GPU dependence. Annual compute spending has reached approximately $19 billion, according to Crypto Briefing, creating a powerful internal incentive for proprietary accelerators. In August, the lab recruited a pioneer from Google's TPU program to lead an in-house silicon initiative. Concurrently, Broadcom has reportedly been exploring an $80 billion debt facility to finance custom chip development and networking solutions for Anthropic, and the lab placed a $250 million pre-revenue bet on Fractile, a startup targeting efficient inference silicon. If Anthropic's custom chip timeline succeeds, it could structurally shift the compute cost curve over a multi-year horizon. The risk is well-established: chip development is capital-intensive, routinely delayed, and difficult to de-risk until tape-out. Google's own TPU effort required years to reach production maturity and has never eliminated large-scale Nvidia procurement — a precedent that contextualizes the ambition without invalidating it.
The bullish case rests on three pillars: demonstrated profitability at scale, a durable enterprise adoption position, and a compute strategy designed to insulate margins from GPU price volatility over a five-to-ten year horizon. Semiconductor analyst Dylan Patel has projected, in coverage dated August 26, that OpenAI and Anthropic together will control the majority of the world's AI compute capacity by 2028 — a concentration that, if realized, would represent formidable structural barriers to new entrants. The bearish case is equally specific. Commentators including analysts at 24/7 Wall St. have questioned whether a valuation approaching $2 trillion can be sustained against simultaneous cash demands — $19 billion in annual compute spend, a $45 billion external procurement commitment, and an active custom-silicon program — of a scale rarely seen outside sovereign entities. The Riot bridge-loan structure is a small but illustrative reminder that counterparty financial engineering at this scale carries its own category of risk. Reports in mid-August also noted some high-end subscribers migrating toward competing services in response to perceived availability gaps, a signal that model quality leadership is not a static asset in a market where rivals are spending comparably. Three signals are worth tracking in the months ahead: whether Anthropic's S-1, once public, confirms the Q2 profitability figures and provides forward guidance on compute unit economics; how Nscale's financing and West Virginia construction timeline holds under the pressure of an October listing window; and whether the in-house silicon program reaches a tape-out milestone before existing GPU procurement contracts lock in rates at current levels.