Anthropic Signs $45B Vera Rubin Deal as Microsoft Exits Nscale's West Virginia Campus

Anthropic has agreed to pay Nscale roughly $45 billion over six years for 460 megawatts of AI computing capacity at the Monarch Compute Campus in Mason County, West Virginia — a site Microsoft had reserved with a letter of intent in March before walking away earlier this summer. Bloomberg broke the story on August 26, with CNBC and Reuters independently confirming the terms within hours through their own confidential sources. The deal, which will draw on Nvidia's forthcoming Vera Rubin chip platform, is Anthropic's largest single disclosed compute commitment and carries a delivery date that does not arrive until late 2027 — well past both companies' planned public offerings.

The transaction puts two pre-IPO companies in an unusually entangled position. Anthropic's $45 billion in contracted spending is the revenue anchor that makes Nscale's September IPO pitch credible to institutional investors. Nscale's ability to build and operate the Monarch campus on schedule, in turn, determines whether Anthropic actually receives the Vera Rubin capacity it has committed to paying for. Both companies will need to disclose this dependency in their prospectuses — documents that have not yet been filed publicly as of this reporting.

Microsoft's Exit Turned a Letter of Intent Into an Opening

The structural context for the deal begins in March 2026, when Nscale used Nvidia's GTC conference to announce two things simultaneously: a $2 billion Series C round at a $14.6 billion valuation — which it described as the largest Series C in European history — and a letter of intent from Microsoft covering up to 1.35 gigawatts of AI compute capacity at the same West Virginia campus, the full scope of Monarch's planned first phase. The letter of intent was publicly characterized as a framework for a long-term partnership, with deployments phased beginning in late 2027.

Letters of intent are generally not legally binding. They function as reservations — expressions of mutual intent that give both parties time to negotiate binding terms — rather than committed payment obligations. Microsoft never disclosed why it declined to convert the Monarch LOI into a signed contract, and the company has not commented publicly on its decision. According to Bloomberg, sources familiar with Microsoft's recent infrastructure planning noted the company has grown more selective about committing to specific development sites, even as its overall capital expenditure trajectory — more than $80 billion in fiscal 2025 — remains intact. The Monarch exit appears to reflect site-specific or timing-specific considerations rather than a pullback from AI infrastructure spending broadly.

When Microsoft's reservation lapsed, Nscale had an asset to offer: the first building of a three-building first phase at Monarch, already under construction, with approximately 460 megawatts of planned capacity and a power infrastructure designed for Vera Rubin NVL72 hardware. Anthropic, which had been publicly acknowledging capacity strain since early 2026, stepped in with what Bloomberg describes as a binding agreement — a materially different instrument from the LOI Microsoft signed. According to a regulatory filing from Aker ASA, a lead investor in Nscale's Series C, a $45 billion contract had been secured at the Monarch campus, providing independent corroboration of the deal's existence if not its precise terms.

What Vera Rubin NVL72 Actually Delivers — and When

The hardware that Anthropic is paying for does not yet exist at the scale this deal requires. Nvidia's Vera Rubin platform is entering limited partner deployments in the second half of 2026, but commercial-scale availability — the kind a 460-megawatt data center requires — has not yet been established. The Monarch campus's first building is expected to begin coming online in late 2027; the remaining two buildings in the first phase are scheduled for 2028.

That timeline matters for understanding what Anthropic is actually buying. At the chip level, the Vera Rubin NVL72 is a rack-scale system integrating 72 Rubin GPUs alongside 36 Vera CPUs in a fully liquid-cooled chassis connected by NVLink 6 fabric, which Nvidia describes as delivering 260 terabytes per second of all-to-all bandwidth with cable-free modular tray designs. A single rack delivers 3.6 exaflops of NVFP4 inference throughput and approximately 2.5 exaflops of training throughput — figures that represent Nvidia's seventh-generation AI compute platform and that have been published by the company but await third-party corroboration at the rack level. A complete 40-rack Vera Rubin POD scales to 1,152 Rubin GPUs and 60 exaflops of aggregate compute, a configuration whose capacity exceeds any prior publicly known AI-specific supercomputer.

The architecture behind those numbers is not simply faster GPUs. Nvidia has designed Vera Rubin around Attention-FFN Disaggregation, a technique that routes the two computationally distinct operations inside a transformer's decode loop to different processor types. Attention over the KV cache is memory-bandwidth-bound and handled by the Rubin GPU's high-capacity HBM memory; feed-forward network execution — and in mixture-of-experts models, the routing of each token to a subset of specialized expert parameter blocks — is better served by the Vera CPU's on-chip processing. The two processors exchange activation tensors over the NVLink 6 fabric within the chassis, with Nvidia's Dynamo serving framework orchestrating scheduling, KV-cache-aware load balancing, and session tracking. This disaggregated approach is what produces the efficiency gains on agentic workloads — where context windows grow session by session and inference is interrupted by tool-execution gaps — that Nvidia demonstrated at Hot Chips 2026 in late August.

Read more: Nvidia's first Vera Rubin benchmark data

Anthropic's 460 megawatts at Monarch will require approximately 194,000 Vera Rubin GPUs, according to Bloomberg's reporting on Nscale's contracted GPU inventory. To serve a 460-megawatt deployment, all of that hardware must not only exist but be manufactured, shipped, installed, and brought into production at a site that is currently under active construction. Nvidia's production ramp for Vera Rubin in the second half of 2026 will determine whether the supply chain can absorb a deployment of this scale by Nscale's late 2027 target.

The Monarch Campus: Natural Gas, Not Hydroelectric

The Monarch Compute Campus sits on 2,250 acres in Mason County that Nscale acquired in March 2026 when it purchased American Intelligence & Power Corporation, a joint venture between Fidelis New Energy and 8090 Industries. Nscale describes the site as the United States' first state-certified AI microgrid — a designation enabled by West Virginia's House Bill 2014, signed in 2025, which allows large data centers to establish independent power microgrids and preempts local municipal zoning in favor of unified state standards.

Operating off the local grid is a practical necessity given the scale of power involved. Monarch's first phase will rely on Caterpillar G3516 reciprocating natural gas generator sets as primary power sources — a combined on-site generation capacity of 2 gigawatts for the first phase, with deliveries scheduled from late 2026 through mid-2027, according to a purchase agreement Caterpillar announced in January 2026. The full campus, when built to its 8-gigawatt potential, is projected to cost approximately $71 billion in total development costs, with roughly $47 billion of that attributable to chip procurement.

The natural gas approach contrasts with Anthropic's Norway arrangement, where the $10 billion Volta Infra Holdings deal runs through a hydroelectric facility that achieves a Power Usage Effectiveness ratio of approximately 1.1 — well below the US data center average of roughly 1.58. Monarch will not achieve Norway's carbon profile. Environmental advocacy groups, including the West Virginia chapter of the Sierra Club, have raised air quality concerns about the scale of natural gas generation planned for the state's data center buildout; a West Virginia Department of Environmental Protection inspection at Monarch in May 2026 documented dust conditions at the construction site, and an air quality permit application for the campus was under review as of mid-2026. The design trade-off is deliberate: an off-grid natural gas microgrid can be deployed faster and at larger scale than renewable-backed alternatives in a state where hydroelectric capacity does not exist, and it isolates Monarch's power demands from the local utility grid that serves Mason County residential and commercial customers.

West Virginia Governor Patrick Morrisey praised the project publicly following the Bloomberg report, crediting the state's microgrid legislation and singling out State Senator Glenn Jefferies for marketing efforts that brought the development to West Virginia.

Nscale's IPO Arithmetic: $51B Backlog, $400M in Annual Revenue

For Nscale, which was founded in 2024 and recorded roughly $33 million in revenue for all of 2025, the Anthropic deal is not just a commercial milestone — it is the foundation of an IPO story. The company is targeting a US listing as early as September 2026, with Goldman Sachs and JPMorgan serving as advisors, according to Bloomberg. Both banks also placed Nscale's $2 billion Series C in March — an advisory continuity that is standard IPO groundwork. The company has since added a $900 million revolving credit facility and completed the acquisition of distributed AI software startup Anyscale for approximately $1.65 billion, a transaction expected to close in the second half of 2026 and intended to add software and optimization capabilities to complement Nscale's infrastructure offering.

Nscale has disclosed to prospective IPO investors a total contracted revenue backlog of approximately $51 billion. The Anthropic deal, at $45 billion, is the single largest entry in that backlog. The figure requires context that the headline number obscures. Infrastructure companies routinely count multi-year agreements in full at signing — meaning a six-year deal is entered at face value the moment both parties execute — even when the underlying capacity does not come online for more than a year. By that standard, Nscale's $51 billion backlog is a standard disclosure practice, not a misrepresentation. But it is also not revenue that Nscale has earned or will earn soon.

The company's actual recognized revenue tells a different story in scale. First-quarter 2026 revenue reached approximately $37 million; the second quarter exceeded $100 million as operational capacity began delivering, according to Bloomberg's reporting on Nscale's IPO pitch to prospective investors. Annualized at the Q2 pace, Nscale's current run rate is tracking nearer $400 million to $500 million — a figure that is genuinely growing at speed but that sits roughly 100 times below the contracted backlog that leads its investor pitch. The gap is not unusual for an infrastructure company at Nscale's stage; CoreWeave carried a similar ratio at its March 2026 IPO. But for public-market investors evaluating a September listing, the math of when contracted revenue becomes recognized revenue — and what happens if a customer encounters headwinds before Monarch comes online — is the central risk question.

That question applies directly to Anthropic. The company is navigating a $65 billion annualized revenue run rate against an annual compute bill that analysts estimate at approximately $19 billion, active litigation over a Department of Defense supply chain risk designation it has contested in court, and a June 2026 export control action that investors said affected June revenue. Anthropic's IPO, targeting a Nasdaq debut as early as October 2026 at a valuation investors have discussed as approximately $2 trillion, is itself subject to market conditions neither the company nor its bankers can control.

Read more: Anthropic's record IPO preparations

Anthropic's $100B Compute Ledger, Diversified by Design

The Nscale agreement is the most recently disclosed entry in a compute acquisition campaign Anthropic began in earnest in April 2026, when it announced an expanded partnership with Google and Broadcom for 3.5 gigawatts of next-generation TPU capacity beginning in 2027. In May, the company secured access to more than 220,000 Nvidia GPUs and approximately 300 megawatts at SpaceX's Colossus 1 facility in Memphis, Tennessee, under a $1.25 billion monthly arrangement that either party can terminate with 90 days' notice — a flexibility clause Anthropic has acknowledged as a structural risk. In July, AMD and Anthropic announced a strategic partnership in which AMD committed a $5 billion equity investment in Anthropic while Anthropic agreed to deploy up to 2 gigawatts of AMD Instinct accelerators beginning in the first half of 2027. The Volta Infra Holdings Norway deal followed on August 4 — $10 billion over six years for 121 IT megawatts of Vera Rubin capacity backstopped by JPMorgan letters of credit. The Nscale deal, disclosed August 26, caps a sequence that has pushed Anthropic's total contracted compute commitment past $100 billion.

The diversity within that portfolio is deliberate. Anthropic's compute supply chain now spans two continents, multiple chip generations (Nvidia Hopper at Colossus, AMD Instinct, Google TPUs, and Vera Rubin at both Volta and Nscale), and a range of financing structures from direct monthly leases to six-year infrastructure agreements backed by private credit. No single supplier or single site controls more than a fraction of Anthropic's planned capacity — a meaningful departure from the concentration that made OpenAI's early dependence on Microsoft Azure a structural risk in competitive infrastructure discussions.

The tradeoff for that diversification is complexity. Anthropic is simultaneously managing live capacity obligations at Colossus 1 — which requires $1.25 billion in monthly payments regardless of actual utilization — alongside multi-year infrastructure commitments to sites that will not deliver capacity until 2027 and beyond. The company acknowledged earlier in 2026 that surging demand for Claude models had caused what it publicly described as inevitable strain, degrading reliability during peak usage periods. Claude Code, the company's agentic coding tool, reached an estimated $2.5 billion in annualized revenue within nine months of its launch, consuming compute at a pace that infrastructure planned for a smaller company could not absorb.

Two Companies Betting on the Same 2027 Delivery

The Nscale deal is ultimately a wager on the convergence of three timelines that no one party controls: Nvidia's production ramp for Vera Rubin chips, Nscale's construction and commissioning schedule at Monarch, and Anthropic's demand trajectory for compute capacity that does not exist yet.

The chip timeline is the most uncertain. Vera Rubin NVL72 is the most advanced AI compute platform Nvidia has announced, and partner deployments are beginning in the second half of 2026 — but the gap between early partner systems and the production volume required to fill a 460-megawatt campus represents a manufacturing challenge that Nvidia has not publicly characterized in detail. Nscale has contracted for approximately 194,000 Vera Rubin GPUs. Whether those chips are available, installed, and running when Monarch's first building expects to come online in late 2027 depends on execution across Nvidia's supply chain, Nscale's construction timeline, and the regulatory approvals that determine when the natural gas microgrid can begin operation.

Nscale's September IPO, if it proceeds on schedule, will force that uncertainty into public view before the deal's capacity materializes. Prospective investors will be asked to evaluate a company whose largest revenue commitment is $45 billion contracted but not yet earning, from a customer that is itself navigating a pre-IPO regulatory and competitive environment. The investment case rests on the thesis that both companies' trajectories are robust enough to survive the 14-month gap between when this agreement was signed and when the first megawatt of Monarch compute is expected to reach Anthropic's workloads. Whether that thesis holds will be tested not by the deal itself — already signed and corroborated — but by the capacity delivery that begins in late 2027.

For Anthropic, the structural question the Nscale deal sharpens is whether its approach to compute — locking in capacity years in advance across a deliberately fragmented portfolio — is a strategic asset or a compounding liability. The case for the asset view is that forward commitments like Monarch protect against the scenario where demand outstrips available supply and forces rate-limit cuts that damage enterprise customer relationships. The case for the liability view is that Anthropic is now contractually obligated to nearly $100 billion in infrastructure spending regardless of what happens to demand, competitive dynamics, or the regulatory environment between now and the early 2030s when the longest commitments expire. The SpaceX Colossus arrangement's 90-day termination clause was designed to preserve some flexibility on the live-capacity side; the six-year Nscale and Volta agreements carry no equivalent exit provision in publicly available descriptions.

What neither view changes is the immediate operational reality: the Monarch campus will not produce a single inference request for Anthropic before late 2027 at the earliest. In the meantime, capacity runs on the existing portfolio — and Anthropic's engineering and infrastructure teams will spend the next 14 months watching Nscale build what $45 billion of the company's future earnings have already been pledged to buy.