Nscale’s $103B Backlog Meets the Substation

In its 2026 S-1, AI-infrastructure builder Nscale reported $103 billion in total contract value, but just $2.6 billion was active as of late August 2026. Between those figures sit substations, GPUs, customer conditions and billions in financing.

The AI infrastructure market has entered a public-market underwriting phase. Investors can no longer treat headline contract value as a sufficient proxy for deliverable revenue. They must price when capacity will operate, which obligations are unconditional, which customers bear them and whether financing lasts until utilization.

A backlog spans five economic states

Neoclouds first needed to answer a simple question: could they find customers for expensive accelerators before those accelerators aged? In a shortage market, a signed commitment offered a persuasive answer. The signed commitment linked demand today to capacity delivered later.

Providers then turned commitments into products. OpenAI, for example, introduced Guaranteed Capacity agreements that exchange one- to three-year spending commitments for discounted access. Customers now reserve future compute at multiple layers of the stack.

An underwriter can arrange those commitments on a ladder. A booking records customer intent. A capacity contract assigns obligations. Active capacity requires the provider to deliver service. Recognized revenue requires the customer to consume or otherwise pay under the applicable terms. Collected cash must then arrive. The contract can remain real at every rung while a different party carries the risk of construction, delivery, acceptance or payment.

Nscale’s operating base shows the distance between those rungs. The company reported $140.6 million in first-half revenue and a $1.02 billion net loss, even as revenue rose 1,252% year over year. Nscale must fund current payroll, equipment and development while much of its disclosed contract value remains attached to future capacity.

Public scrutiny followed that transition. Nscale coverage rose from four articles in the second quarter of 2026 to 17 during the first 83 days of the third. Coverage shifted from securing chips and sites to examining what each agreement requires before it turns into service.

Microsoft and Anthropic make the book stronger and narrower

Microsoft and Anthropic account for 85% of Nscale’s disclosed total contract value. Those customers bring substantial demand and financing capacity, making a concentrated book more credible than a diversified collection of weak buyers. They also tie Nscale’s buildout to two procurement teams that can enforce delivery conditions and allocate future spending among competing providers.

Anthropic’s reported commitment makes the scale visible. The company reportedly agreed to pay Nscale $45 billion over six years for roughly 460 megawatts at a West Virginia data center using Nvidia Vera Rubin chips. A single frontier-lab agreement can therefore dominate the economics of a site, the financing raised against it and the provider’s reported backlog.

Anthropic has spread its planned workloads across providers. Since October, the lab has entered agreements for at least 14.8 gigawatts of compute capacity. That portfolio reinforces Anthropic’s demand while giving it procurement options unavailable to any single provider.

Nscale’s available disclosures do not establish whether the reported Anthropic agreement is take-or-pay, what cancellation rights or credit support apply, or which delivery conditions Nscale must satisfy. Those terms separate a six-year payment obligation from a reservation whose economics change if the site, power or hardware arrives late. Without them, investors cannot know how much of the reported $45 billion can support project financing before delivery.

Nscale must buy the asset before its customers use it

Nscale must secure land, power, construction and chips before a contracted workload can run. Every step pushes cash out before service begins. Whether Nscale can bill through delays depends on contract terms it has not disclosed.

Major lenders are already confronting underwriting, insurance and exposure-management problems as they finance the U.S. data-center buildout. Banks cannot treat a GPU cluster like an ordinary office building: the tenant may depend on a fast-changing model market, the equipment ages quickly, and the site’s value depends on power and network connections that cannot be moved with the racks.

Providers have responded with GPU-backed debt and special-purpose vehicles. Lambda, for example, was reported to be selling a $917 million leveraged loan for GPU purchases connected to an Nvidia contract. An SPV can relocate debt on an organizational chart, but it cannot remove the interval between buying hardware and collecting customer cash.

For Nscale, lenders must align customer commitments with the emerging AI data-center financing stack of construction debt, equipment funding, guarantees and insurance. If financing runs out before capacity activates, the provider owns pieces of a system rather than a revenue-producing asset. If the customer’s obligation begins only after delivery, the provider also carries the cost of delay.

Microsoft and AWS validate demand—and tighten supply

Microsoft and AWS make the broad demand case difficult to dismiss. Microsoft reportedly plans to expand from roughly 12 gigawatts of data-center capacity to more than 38 gigawatts by 2032, with about one-third devoted to AI-specific chips. AWS plans to add two million Nvidia Blackwell Ultra, Rubin and Rubin Ultra GPUs during 2027 and 2028, on top of one million previously announced.

Those plans intensify scarcity. Microsoft and AWS will also procure chips, electrical equipment, power and construction capacity, reserving scarce supply while raising the cost and complexity faced by smaller builders trying to energize their own sites.

U.S. data-center projects exceeded 80 gigawatts in 2025 when built, underway, planned and stalled capacity were counted together. Some sites carried workloads; others still awaited permits, substations, equipment, construction or customers. Counting both the same way hid the distance to revenue.

A powered building without delivered GPUs cannot serve the workload. Delivered GPUs without sufficient power cannot run it. An operating cluster without utilization cannot cover its capital base. This capacity lag is why the contracted megawatt has become a financial unit: lenders need the customer obligation to survive every physical handoff between the signed agreement and the paid invoice.

Crusoe and Applied Digital expose different failure points

Crusoe raised $3.9 billion at a roughly $30.9 billion post-money valuation while pursuing factory-built data centers. But Crusoe also paused a Wyoming project after reportedly failing to secure customers including Google amid concerns about cost and timetable. Investors funded the company without validating every proposed site.

Applied Digital reached a later failure point. The operator reported $53 million in quarterly revenue against a $63 million estimate and attributed part of the shortfall to customers delaying lease renewals; its shares fell more than 13%. Applied Digital had moved beyond announcing capacity, but customer timing still reached the income statement.

Crusoe’s paused site makes customer acquisition visible before construction. Applied Digital’s miss makes renewal timing visible after service begins. Nscale’s loss makes the funding interval visible while capacity is being assembled. Each company exposes a different joint in the same physical sequence, and public investors can inspect each joint with corresponding evidence: commissioned megawatts, delivered equipment, utilization, renewals and cash collection.

Nscale’s filing places $103 billion at the entrance and $2.6 billion on the active side. Public investors must inspect everything between them: the energized substation, installed Rubin racks, binding customer obligation, funded debt and paid invoice. The backlog remains part of the blueprint; underwriting has walked onto the site.