London-based 'neocloud' Nscale filed an S-1 registration statement with the SEC on September 18, 2026, for an IPO on the New York Stock Exchange under the ticker NSCL, targeting a valuation of approximately $30 billion. For the half-year ended June 30, 2026, the company's revenue grew 13.5-fold to $140.6 million, but its net loss reached $1.02 billion. This is the first detailed public window into the economics of GPU rental providers since CoreWeave, and it shows that AI infrastructure growth is currently being paid for with debt rather than profit.

image
image

What happened

Nscale was spun out of Arkon Energy, a company engaged in Bitcoin mining, in 2024, and today it rents GPU data center capacity to AI labs. The S-1 filing for the SEC was prepared by underwriters Goldman Sachs, J.P. Morgan, and Morgan Stanley, and the target valuation for the offering significantly exceeds the private valuation of $14.6 billion in March 2026. The document provides comparative financial data: in the first half of 2025, revenue was $10.4 million, the loss for the same half-year a year earlier was $368.9 million, and for the full year 2025, the company generated only $33 million in revenue. The operational section of the S-1 records 25,000 active GPUs out of 461,000 active and contracted, a contracted transaction volume (TCV) of $103.4 billion, and debt of over $8 billion; one unnamed client accounted for more than half of revenue. It is separately noted that under four agreements with Anthropic related to the Monarch campus in West Virginia, the necessary funding obligations had not yet been received as of the filing date. The offering itself had not yet taken place at the time of the filing.

Context

'Neoclouds' build data centers and rent GPU capacity to AI model creators: this is a capital-intensive business where accelerators and construction are paid for today, while revenue comes later through long-term contracts. The sector is already familiar to investors through CoreWeave, and Nscale's debut will be compared with its trajectory. The scale of current spending is evident from the August 2026 deals: $1.85 billion was raised for a Texas site and $1.2 billion for a North Carolina site, with both financings tied to future construction. The key to understanding the reporting is that TCV, or contracted obligations, is not revenue: between signing a contract and the appearance of servers stand construction, GPU procurement, and new financing, so the fate of the promised capacity depends on continuous access to capital. The half-year growth, however, was calculated from an extremely low base in 2025 and in itself does not yet prove a mature unit economics.

Why this matters for the industry

For the industry, the S-1 is a rare detailed portrait of the 'neocloud' economics, and it describes a business based on access to capital rather than margin: contracted mega-deals turn into servers only through new loans, without external financing the declared capacity does not materialize, and the risk is shifted to public investors. The unfunded Anthropic deals are a specific marker of the fragility of the 'contract → construction → revenue' chain: until the money is raised, part of the promised GPU fleet exists on paper. The gap between the active and contracted fleet of more than 18 times shows that only actually deployed capacity matters, and the distribution of computing is determined by mega-deals with the largest customers, not the open market. The fact of the public filing itself has already given computing buyers arguments for negotiations on prices and terms, and startups a reason to design the compute layer as swappable and multi-provider. If the offering takes place, quarterly reports will turn the verification of converting contracted billions into a regular public test.

Why this matters for users

For working engineers and product teams, almost nothing changes in the immediate term: this is a financial event, not a product release. The significance of the event lies in the horizon. Low inference prices at 'neoclouds' today are subsidized by borrowed capital rather than efficiency, so they may be revised in the event of a market cooling or financing problems. The actual availability and price of GPUs in 2027–2028 will depend on whether contracted volumes are converted into deployed servers. A specific action is available now: conduct an audit of dependence on compute providers and build an abstraction over the inference API so that changing suppliers does not turn into rewriting the product. In an optimistic scenario, the conversion of contracts will make large experiments cheaper and make product classes with heavy inference profitable — mass agentic workloads and cheap batch tasks; in a pessimistic scenario, consolidation of 'neoclouds' and a revision of long-term contracts are likely.

What is still unknown / limitations

The IPO has not yet taken place: the price, volume, and dates of the offering are undetermined, and the valuation will remain a target until the book of orders is formed. Under the four agreements with Anthropic, financing is not secured, and it is unknown whether it will be found before the offering. The sample for conclusions about demand is extremely narrow due to the concentration of revenue on one unnamed client, and the half-year figures reflect a very early stage of the business, so conclusions about mature unit economics are premature. The title of the original material — 'Nscale S-1 demonstrates unprofitability of AI' — is the author's interpretation of the discussion: the S-1 data shows the unprofitability of a specific 'build on credit' model, not AI as a technology, and extrapolating one issuer to the entire GPU market without additional sources is incorrect. Finally, the Anscale deal, which, according to available data, is supposed to close simultaneously with the IPO, is described in the available materials without details and requires separate verification.

Sources

Author

Look at AI, editorial team