DataNSCLSEP 21, 2026

Nscale files $30B NYSE IPO with $103.4B backlog, exposing AI compute cost floor

The UK neocloud's Sept. 18 S-1 details a $44.6B Anthropic contract and $1B Nvidia convertible note, giving small-business AI buyers the clearest read yet on where infrastructure pricing settles.

BY LINH VINH · DATA INFRASTRUCTURE

Nscale, the London-based AI cloud builder, filed its Form S-1 with the SEC on September 18, disclosing a $103.4 billion contracted backlog against $140.6 million in first-half 2026 revenue. The filing, which lists NSCL as the intended NYSE ticker and targets roughly $3 billion in proceeds, is the most transparent look yet at what durable AI infrastructure pricing looks like when a hyperscaler-adjacent builder has to show its contracts to public markets.

The headline commitments do the analytical work. Nscale signed GPU Services Agreements with Anthropic on August 25 worth up to approximately $44.6 billion, and disclosed Microsoft agreements totaling up to $43.8 billion running through December 2033. Together, per The Energy Mag’s read of the filing, those two customers represent about 85% of the disclosed contract total. The weighted-average contract life is 5.7 years. This is take-or-pay compute, sold in blocks that stretch past the current administration and most VC fund lives.

The capital stack matches the contract structure. On September 15, Nscale subscribed a minimum $3.1 billion in convertible notes, $2.1 billion in general notes plus $1 billion earmarked for Nvidia, whose investment is expected to close around November 16, with up to $10 billion in additional notes or shares authorized over time. Goldman Sachs, J.P. Morgan, and Morgan Stanley are lead bookrunners. The March private round valued the company at $14.6 billion.

The gap between promise and delivery is the part small-business AI buyers should read carefully. Nscale has 25,000 GPUs live against 461,000 active and contracted, an 18:1 ratio, and only $2.6 billion of the $103.4 billion backlog is classified as active. First-half revenue grew 1,252% year-over-year from $10.4 million, while net loss tripled to $1.02 billion from $368.9 million, per Bloomberg. One customer produced 52% of that revenue.

What the filing settles is that the Nvidia Vera Rubin NVL72 systems anchoring sites like the 2,250-acre Monarch Compute Campus in Mason County, West Virginia are being financed on multi-year, dollar-committed terms. Inference costs at the infrastructure layer aren’t collapsing. They’re being locked in. Any Q4 pitch that assumes per-seat AI pricing will drift lower next year is pricing against contracts that already exist.

Sources

About the author

Linh Vinh

DATA INFRASTRUCTURE

Linh Vinh covers vector databases, retrieval, feature stores, and the plumbing layer of enterprise AI. She has been writing about data infrastructure for several years and files on infra vendor consolidation, pricing pressure, and technical architecture shifts.