Nscale Heads for the NYSE: $1.02 Billion Loss in Half a Year, Revenue Up 12x

The new IPO playbook for AI infrastructure companies: show everyone your loss figure first, then ask the market if it still wants in.

Three key facts

On September 18, AI cloud provider Nscale filed its S-1 with the SEC, applying to list on the NYSE under NSCL, with Goldman Sachs, JPMorgan and Morgan Stanley as lead bookrunners. Offering size, valuation and price range are all undetermined and the registration statement isn't effective yet — but the financial cards are face up: H1 2026 revenue of $140.6 million, up 1,252% year over year (from just $10.4 million); net loss over the same period of $1.02 billion, nearly triple last year's $368.9 million loss.

The loss structure deserves a close look: of the $1.02 billion, $457 million comes from fair-value adjustments (on paper) and $492 million from operating activities (real cash). As of August 31, the company runs 12 data-center locations across the U.S. and Europe with more than 1,000 full-time employees. The model: rent Nvidia GPUs to customers for training and inference — self-described as a "full-stack AI cloud" combining software, compute, data centers and power.

The most conspicuous name on the shareholder list is Nvidia: a stake above 5% via non-voting warrants, and simultaneously Nscale's core supplier. The customer list is equally starry — signed deals include OpenAI and Anthropic. The shovel seller is a shareholder in one of its biggest customers, who then use their raised capital to buy shovels. This loop is printed in black and white in the S-1.

Wang Dou's Take

A 1,252% growth rate and a $1.02 billion loss printed on the same page — that's what the AI infrastructure business actually looks like right now: revenue is the flow from renting GPUs, the loss is the bill for buying them. The essence of the business is a capital-heavy middleman — get cards cheap (at a shareholder's price, no less), rent them out dear, and bet the GPU rent spread outruns depreciation. GPUs turn over every three years; the depreciation schedule is far more reliable than customer loyalty.

Nvidia's 5% is the part that lingers. A supplier taking equity in a big customer looks like strategic synergy on paper; in practice it securitizes the sales risk of its own chips: much of every dollar Nscale raises becomes Nvidia revenue, and Nvidia then shares in Nscale's IPO premium via its stake. Left hand to right hand, both sides book income. Wall Street calls this an ecosystem; accountants call it related-party transactions; retail investors only need to remember one thing — when reading an AI infrastructure company's filings, first check whether Nvidia is on the shareholder list.

For public markets, this S-1 is a litmus test. The reference case sits right there: peer CoreWeave IPO'd at $40 in March 2025, surged 250% within three months, cratered 57% in this year's first quarter, and this very week logged a five-day, 19% slide that erased roughly $10 billion in market value — AI infrastructure stocks are a rollercoaster now, not a staircase. The era of paying for "growth" is over; the market now pays for "unit economics." Nscale hasn't given a valuation range, which tells you the underwriters themselves aren't sure. When the price lands, the number will honestly answer one question: is AI infrastructure the railroad of a new era, or the fiber optic cable of 2000?

Source: PR Newswire · CNBC · SuperpowerDaily · Trefis (CoreWeave data)