An AI Cloud Provider Just Filed to Go Public With a $1 Billion Loss on $140 Million of Revenue
Nscale filed its S-1 for a US IPO on September 18, and the numbers inside it are worth sitting with. In the first half of 2026, the company posted a net loss of $1.02 billion against revenue of $140.6 million. A year earlier, the same period showed a $368.9 million loss on $10.4 million of revenue. Revenue grew more than 1,200% year over year, and the loss grew right alongside it. This is a real GPU cloud provider, backed by Nvidia and Microsoft, with a $44.6 billion services agreement with Anthropic, telling public investors exactly how much it costs to build the compute capacity that makes today's AI pricing possible.
The numbers behind the filing
Nscale plans to list on the NYSE under the ticker NSCL, with Goldman Sachs, JPMorgan, and Morgan Stanley as lead underwriters. On August 25, it signed a GPU Services Agreement with Anthropic worth up to approximately $44.6 billion in aggregate payments over the life of the deal. On September 15, just three days before filing, it closed a $3.1 billion subscription agreement with Nvidia, split between $2.1 billion in unsecured convertible loan notes and $1 billion in either convertible notes or non-voting shares. Nscale generates revenue primarily on a per-GPU-hour basis through its cloud platform, the same pricing model every solo builder renting compute is already familiar with.
Put the pieces together and you get a company whose entire business is buying, building, and operating GPU infrastructure, funded by enormous vendor and customer commitments, running at a loss that scales with its growth rather than shrinking as it gets bigger. That's not necessarily a red flag on its own, infrastructure buildouts often look like this in their early years. What it is, is a rare, audited look at the actual unit economics behind the compute layer that everyone building AI products depends on.
Why a loss this size, this early, matters to you
If you rent GPUs, whether directly from a provider like Nscale or indirectly through a model API that's built on top of similar infrastructure economics, you're currently paying a price that doesn't fully reflect what it costs to provide that compute. A $1.02 billion loss on $140.6 million of revenue means Nscale is spending roughly $8 for every dollar of revenue it books in this specific period, an extreme ratio even for a capital-intensive buildout phase. Some of that is upfront infrastructure spend that will pay off over years, not a permanent cost structure. But some of it reflects genuinely below-cost pricing to win customers and lock in long-term commitments like the Anthropic deal while the land grab for AI infrastructure market share is still underway.
That land grab doesn't last forever. At some point, either the loss narrows because efficiency improves and pricing holds, or it narrows because pricing rises to meet the actual cost of running the infrastructure. Both are plausible outcomes. Neither is something a solo builder gets advance warning about before it hits their bill.
The part of the filing that's actually reassuring
I don't want to overstate the doom here. The revenue growth is genuinely enormous, over 1,200% year over year is not a rounding error, and the fact that Nvidia and Anthropic are both willing to sign multi-billion-dollar commitments with Nscale suggests the largest, most sophisticated players in AI infrastructure believe this capacity is worth locking in at scale. Companies with that level of due diligence access don't sign $44.6 billion agreements on a whim. If Nscale's model is fundamentally sound and the losses are the normal shape of an infrastructure buildout rather than a sign of mispriced unit economics, the current pricing environment could hold longer than the raw loss number makes it feel like it will.
The honest read is that this filing doesn't tell you which outcome you're getting. It tells you the range is wide, and that "GPU rental is a stable, permanent line item in my cost model" is not a safe assumption to build a five-year plan around, regardless of which direction it eventually resolves.
What I'd actually do
I treat current GPU and inference pricing the way I'd treat an introductory rate on anything else: real, usable, and not something to build a permanent cost assumption around. For anything I'm building that depends on compute costs staying roughly where they are today, I model at least one scenario where per-unit compute costs rise 30 to 50% over the next two to three years, and I check whether the product still works at that price point before I commit to an architecture that assumes today's pricing indefinitely. If a product only pencils out at today's subsidized rate, that's useful information to have now, not after the rate changes.
I also pay more attention than I used to to who's actually absorbing the loss in any "affordable AI infrastructure" story. When a provider's own S-1 tells you it's losing eight dollars for every dollar of revenue, that's not marketing copy, it's a legally required disclosure, and it's about as close to ground truth as you'll get on what compute actually costs to provide right now.
Where I could be wrong: infrastructure buildouts genuinely do follow a J-curve, and Nscale's losses could narrow sharply as utilization catches up to the capacity it's building, especially with committed revenue like the Anthropic deal already on the books to fill it. If that plays out, today's pricing might prove more durable than this filing makes it look, and I'll have modeled a cost increase that never showed up. I'd rather plan for a price move that doesn't happen than get caught by one that does.
Author
Lukas
@lukcombinator