Everything About AI
NextFin News - Anthropic is in talks with Nscale for a second cloud-computing agreement worth about $45 billion, people familiar with the negotiations said on October 19, as the Nvidia-backed data-center developer races to lock in demand before its planned New York listing. The talks would add a second facility at Nscale's Monarch campus in West Virginia to an existing Anthropic commitment of up to $44.6 billion already disclosed in Nscale's IPO filing — effectively doubling the Claude maker's exposure to a company whose public-market case rests on a handful of hyperscale customers.
The Backlog That Is Not Yet Revenue
The numbers explain both the opportunity and the risk. Nscale filed its registration statement with the Securities and Exchange Commission on September 18, 2026, reporting $103.4 billion in active and contracted total contract value as of August 31 — up from $38.0 billion at the end of 2025. Only $2.6 billion of that was generating revenue. The remaining $100.8 billion is a promise: capacity not yet built, GPUs not yet delivered, power not yet connected. By the company's own accounting, roughly 97.5% of its contracted value had not yet converted into active revenue.
Into that gap steps Anthropic — twice. On August 25, 2026, the two companies signed a series of GPU services agreements under which Nscale would provide dedicated GPU infrastructure at the Monarch Compute Campus using Nvidia's next-generation Vera Rubin NVL72 systems, for aggregate payments of up to approximately $44.6 billion. The agreements are structured as four separate contracts, each between Anthropic and a wholly owned Nscale subsidiary, and are contingent on "specified delivery and service availability requirements." The new talks, reported in October, would add a second facility worth roughly the same amount at the same campus.
The Monarch campus is the scale of the ambition. Nscale's March 2026 announcement described a 2,250-acre site in Mason County, West Virginia, with a first phase designed for 1.35 gigawatts of AI-compute capacity and a power runway scalable beyond eight gigawatts, including what the company calls the United States' first state-certified AI microgrid. Microsoft had signed a letter of intent for 1.35 gigawatts at the same site in March; Anthropic's commitment later anchored the campus. Reporting ahead of the listing placed Nscale's target valuation as high as $35 billion, with Goldman Sachs, J.P. Morgan and Morgan Stanley as lead underwriters and a proposed NYSE ticker of "NSCL."
The timing is not incidental. Anthropic confidentially submitted its own draft S-1 to the SEC on June 1, 2026, and has held banker meetings with a possible listing in view. A second $45 billion anchor customer, secured before Nscale prices its shares, converts a story about speculative capacity into a story about contracted revenue. But the arithmetic also exposes the fragility: Nscale posted a $1.02 billion net loss on $140.6 million in revenue during the first half of 2026 — a 1,252% jump in revenue from $10.4 million a year earlier, but a loss nonetheless. The company's weighted average contract life is about 5.7 years. Its entire balance sheet is a bet that the GPUs arrive on time, the power gets connected, and Anthropic keeps paying.
Why a Second Deal Matters More Than the First
The first-order read of the Nscale-Anthropic talks is simple: Anthropic needs GPUs, Nscale has a campus. That is true, and it is incomplete. The deeper function of a $45 billion take-or-pay contract signed months before an IPO is that it serves as collateral for the public offering itself.
Nscale's filing makes the structure explicit: "Our commercial arrangements are primarily structured as long-term, take-or-pay contracts. These take-or-pay contracts, totaling $2.6 billion of active and $103.4 billion of active and contracted TCV as of August 31, 2026, provide fixed pricing and contracted revenue over the term." A take-or-pay contract means the customer pays whether or not it uses the capacity. For an investor reading a prospectus, that converts a capital-intensive infrastructure gamble into something that resembles an annuity. The filing also disclosed $56.4 billion in remaining performance obligations.
The mechanism runs through the cost of capital. A data-center developer builds ahead of demand because GPUs and power take years to permit and construct. That upfront spend must be financed — with debt, with equity, or with customer pre-commitments. Customer pre-commitments are the cheapest capital of all: they carry no interest, no dilution, and they de-risk the debt tranche that follows. Every $45 billion agreement signed before the IPO lowers the rate Nscale can borrow at, which lowers the cost per megawatt, which widens the margin between what Anthropic pays and what Nscale spends to deliver.
This is why the second deal matters more than the first. One $45 billion contract is a flagship. Two is a pattern — evidence that the first was not a one-off negotiation advantage but a repeatable sales motion. For a company targeting a valuation as high as $35 billion against $140.6 million of half-year revenue, pattern is the only thing the market can underwrite.
The counter-reading is that these contracts are not collateral so much as camouflage. Only $2.6 billion of the $103.4 billion was active as of the end of August, and reporting on the filing noted that Nscale had not yet secured financing for the Anthropic deal. The revenue is contingent on "specified delivery and service availability requirements" — the filing's own language — and failure to meet defined milestones and sustain reliable computing performance gives Anthropic grounds to walk away. If the Vera Rubin systems slip, if the microgrid permitting lags, if the power does not connect, the revenue does not materialize and the take-or-pay protections may not either. A $103 billion backlog that cannot be recognized is not revenue; it is a marketing document.
Nvidia's own exposure runs deeper than chip sales. The chipmaker signed computing lease agreements worth approximately $1.2 billion and guaranteed about $860 million in lease obligations assumed by Nscale for a Texas facility. In September, Nscale raised $3.36 billion in pre-IPO convertible financing led by Third Point, with Nvidia committing $1 billion of the round, convertible into ordinary shares or non-voting shares upon completion of the IPO. The relationship helps Nscale obtain GPUs, capital and long-term orders — and it means the company carries a high degree of reliance on the Nvidia ecosystem, a dependency the filing itself flags alongside a warning that the partnership "does not eliminate supply chain risks."
Anthropic Is Buying Insurance Against a Supply Crunch
From Anthropic's side, the logic is defensive. In its own announcement, the company said its annualized revenue run rate had surpassed $30 billion, up from approximately $9 billion at the end of 2025, and that the number of business customers spending more than $1 million a year had doubled from 500 to more than 1,000 in less than two months. "Demand from Claude customers has accelerated in 2026," it said. Growth at that pace is useless without the inference capacity to serve it — every refused query is revenue left on the table, and every latency spike is a customer considering a competitor.
Anthropic has responded by diversifying across every supplier that can move metal. It holds agreements with Nscale, a $35 billion deal with Lambda for a Texas data center being developed by Hut 8, a $50 billion agreement with neocloud Fluidstack, a $45 billion arrangement with SpaceX for the Colossus and Colossus II clusters, a multibillion-dollar CoreWeave contract announced in April 2026, and now an expanded agreement with Google and Broadcom for "multiple gigawatts" of next-generation TPU capacity coming online starting in 2027. "We train and run Claude on a range of AI hardware — AWS Trainium, Google TPUs, and NVIDIA GPUs — which means we can match workloads to the chips best suited for them," the company said.
"This groundbreaking partnership with Google and Broadcom is a continuation of our disciplined approach to scaling infrastructure: we are building the capacity necessary to serve the exponential growth we have seen in our customer base while also enabling Claude to define the frontier of AI development," said Krishna Rao, Anthropic's chief financial officer. "We are making our most significant compute commitment to date to keep pace with our unprecedented growth."
The second-order effect is what this does to Nvidia's pricing power. Anthropic's portfolio is a deliberate hedge: by committing to TPUs, to SpaceX's in-house clusters and to at least two neoclouds, Anthropic reduces its dependence on any single GPU supplier at the moment Nvidia is preparing to ship Vera Rubin — the successor to Blackwell, which Nvidia says delivers up to 5x the inference performance of the current generation. If Anthropic can credibly threaten to shift workloads, every supplier must price more competitively. The $45 billion Nscale deal is not just capacity procurement; it is leverage in a negotiation with the world's most valuable company.
But insurance has a premium, and Anthropic is paying it in concentration. The company has projected 2028 revenue of roughly $190 billion to $200 billion. To hit that, it must keep spending — capital expenditure is the price of admission. The risk is that the industry is simultaneously over-ordering. Every frontier lab has signed gigawatt-scale deals; every neocloud has raised debt against similar take-or-pay contracts. If demand growth slows even slightly, the market discovers that "multiple gigawatts" of contracted capacity is not the same as "multiple gigawatts" of utilized capacity.
Cyclical Capex Wave, Structural Change in Financing — Keep Them Separate
The central question for investors is whether this is a cycle or a regime. The answer is both, and confusing them is how money gets lost.
The capex wave is cyclical. It is mean-reverting by construction. Data centers are being built ahead of demand because every participant believes demand will compound. History is littered with infrastructure cycles that ended in idle capacity: the fiber-optic buildout of the late 1990s, when companies laid enough dark fiber to carry traffic that never arrived; the shale boom, when drilling outpaced pipeline takeaway capacity. The tell in those cycles was the same: financing available, demand assumed, utilization never caught up. Nscale's own numbers carry the signature — $103.4 billion contracted against $2.6 billion active, $1.02 billion of losses on $140.6 million of revenue. If Anthropic's growth decelerates from triple digits to something ordinary, the contracted capacity becomes a fixed cost that cannot be turned off.
The financing structure, however, is structural. The take-or-pay, multi-year contract with a 5.7-year weighted average life is a genuine regime change in how AI infrastructure is funded. In the 2010s cloud buildout, hyperscalers owned the assets on their own balance sheets — Amazon, Microsoft and Google absorbed the capex and the risk. In this cycle, the risk has been pushed out to specialized developers and neoclouds, who then securitize it through customer contracts. That is not going back. The capital markets now have a template for pricing AI infrastructure as contracted cash flow rather than as speculative growth, and templates, once created, persist.
So the correct read is layered: short term, this is a capex cycle with all the usual overshoot risk; long term, the contractual architecture that finances it is durable. The investment question is not "is AI growing" — it is "who owns the asset when utilization disappoints." Nscale's answer is: the customer pays anyway. The market's job is to decide whether that promise survives a downturn.
The strongest counter-thesis attacks exactly this. A skeptic would argue that take-or-pay protections are only as strong as the counterparty's willingness to walk. Anthropic is private, well-funded and has options — it can restructure, delay or renegotiate if the capacity arrives late or if demand cools. The filing itself conditions the $44.6 billion on "specified delivery and service availability requirements," which means Nscale must perform before it gets paid. In a downturn, the party with the most alternatives — Anthropic — holds the leverage, not the party with the stranded asset. That is the opposite of the annuity story, and it is backed by the filing's own contingency language.
The falsifying signal is concrete: watch Nscale's first quarterly report after the IPO for the conversion rate of contracted TCV into active revenue. If less than 15% of the $103.4 billion converts to active within 12 months of the listed facilities coming online, the annuity thesis is wrong and the backlog is inventory. Conversely, if conversion tracks the 5.7-year contract life smoothly and gross margins on active capacity exceed 40%, the regime-change thesis holds.
Who Benefits, Who Is Exposed, and What to Watch
In the short term, the beneficiaries are clear: Nscale gains a second anchor customer before its IPO, Nvidia gains a buyer for Vera Rubin systems that do not ship until late 2027, and Anthropic gains capacity insurance. The exposed party is the public-market investor who buys Nscale stock on the strength of a $103.4 billion backlog that is 97.5% not yet generating revenue.
By time horizon:
- Short term (sentiment and liquidity): the talks support the AI trade. Nvidia shares rose 1.45% to $217.23 on October 19, and the stock carries a consensus price target of $339.17 across 79 analysts — a 56% implied upside. Any confirmation of the second deal before the IPO prices is likely to be read as validation.
- Medium term (fundamentals): the test is execution. Monarch's first phase is due online in late 2027, with reporting on the filing pointing to initial capacity activation in 2028. Between now and then, every permitting delay, every GPU allocation slip and every Anthropic renegotiation headline will move the stock. The campus's multi-gigawatt buildout cannot be unwound cheaply.
- Long term (structural): if the take-or-pay model proves durable through a downturn, Nscale and its neocloud peers become the utilities of the AI age — lower-growth, contracted, bond-like. If it does not, the cycle ends in stranded capacity and equity wipeouts.
Scenarios:
- Base case: the second deal closes near $45 billion, Nscale lists near its targeted valuation, and capacity comes online in 2027–2028 as scheduled. Revenue converts at a pace consistent with the 5.7-year contract life.
- Upside case: Anthropic's own IPO prices at a premium, demand for Claude continues to accelerate, and Nscale signs a third anchor customer at Monarch — the campus fills beyond 1.35 gigawatts toward its eight-gigawatt runway.
- Downside case: Vera Rubin shipments slip, the Microsoft commitment for the site does not convert into binding capacity as scheduled, and Anthropic renegotiates one or both contracts as its own growth normalizes. The backlog shrinks rather than converts.
What to watch: Nscale's IPO pricing and the first post-listing quarterly report on TCV conversion; Anthropic's own IPO timing and its 2028 revenue guidance; and Nvidia's Vera Rubin shipment schedule. The single falsifying signal for the bullish read is a contracted-to-active conversion rate below 15% within a year of facilities going live.
This is not a bet on AI demand. It is a bet that a contract signed today will still be honored when the capacity arrives in 2027 — and the history of infrastructure cycles says the counterparty, not the builder, gets to decide that.
Explore more exclusive insights at nextfin.ai.
