Only 25,000 of Nscale's 461,000 active and contracted GPUs were running when it filed to go public. That gap is the risk every GPU buyer is signing up for.
Nscale's IPO prospectus contains two numbers every GPU buyer should read side by side. As of August 31, 2026, the London-based neocloud had 461,000 GPUs active or under contract. Only 25,000 of them were actually running. That is about 5.4 percent of the fleet it is presenting to public investors, according to the Form S-1 it filed with the U.S. Securities and Exchange Commission on September 18.
The other 94.6 percent is promised capacity, much of it scheduled for 2027 and 2028, and paid for with debt and a pre-IPO round that was widely reported at $3.5 billion. If you are signing a multi-year GPU contract with any neocloud right now, the gap between contracted and delivered capacity is the most important risk in the deal. For the first time since CoreWeave listed in 2025, a filing lets you measure it instead of guessing.
On September 4, Bloomberg and Reuters reported that Nscale was in talks to raise about $3.5 billion ahead of its listing: up to $1.5 billion in convertible notes led by Third Point, plus roughly $2 billion from NVIDIA, with Goldman Sachs running the process. The notes convert at the IPO price minus a double-digit discount, and that discount adjusts only up to a $30 billion valuation. Above $30 billion, the conversion price stops moving.
The S-1 tells a slightly different story. On September 15, Nscale signed subscriptions for a minimum of $3.1 billion: $2.1 billion of unsecured convertible loan notes, plus $1.0 billion of notes or non-voting shares to NVIDIA, with the NVIDIA tranche expected to close around November 16, 2026. Bloomberg later reported the round closed at about $3.36 billion, with Apollo, Citadel, Hudson Bay and the Abu Dhabi Investment Council investing alongside Third Point and NVIDIA. NVIDIA's committed amount ended up about half of the figure first floated.
The $30 billion cap is the real signal. It roughly doubles the $14.6 billion valuation Nscale set with its $2 billion Series C in March 2026, and it is the level the note holders are effectively underwriting. Nscale has applied to list on the New York Stock Exchange under the ticker NSCL, with Goldman Sachs, J.P. Morgan and Morgan Stanley as lead bookrunners. As of early October, it had not disclosed a share count or price range.
Strip out the valuation chatter and the S-1 reads like a construction company with a cloud attached. Four disclosures matter most.
Revenue for the six months ended June 30, 2026 was $140.6 million, up 1,252 percent from $10.4 million a year earlier. Full-year 2025 revenue was $33.0 million. The net loss for the half was $1.02 billion. Cost of revenue, excluding depreciation, was $189.6 million, which is $49 million more than revenue. Depreciation added $174.0 million, and net interest expense jumped to $95.1 million from $2.2 million. About $457 million of the loss came from non-cash fair value adjustments, so the operating picture is better than the headline loss, but adjusted EBITDA was still negative $199.2 million.
Nscale reports $103.4 billion of active and contracted total contract value (TCV) under take-or-pay agreements, up from $38.0 billion at the end of 2025. Only $2.6 billion of that, about 2.5 percent, is active. The largest addition is a set of agreements with Anthropic, signed August 25, worth up to about $44.6 billion for NVIDIA Vera Rubin NVL72 capacity at the Monarch Compute Campus in West Virginia. Monarch's first 2 GW of gross power is not expected online until the first half of 2028.
Nscale's largest customer accounted for 52 percent of revenue in the first half of 2026 and 73 percent in 2025. The S-1 does not name that customer; the Financial Times has reported that it is linked to ByteDance. Nscale says it expects Microsoft and Anthropic to become significant customers as their contracts go live, which would dilute that concentration, but only once the capacity is built.
Nscale owns and operates its GPU fleet, which puts it on the right side of the first question buyers should ask. But the filing says its active capacity today is primarily in colocation and leased facilities, and that in some cases it does not control the underlying third-party building. The wholly owned campuses are mostly in the contracted column, not the active one. Nscale also disclosed material weaknesses in its internal control over financial reporting, a routine but real flag for a company this young.
Neocloud economics are simple to describe and brutal to finance. A provider buys GPUs years before the revenue arrives, borrows against the hardware and the customer contract, and hopes deliveries land on schedule. Nscale's S-1 lists a stack of facilities: about $1.4 billion of GPU financing, a $900 million revolving credit facility, up to $1.85 billion against GPUs in Ward County, Texas, up to $1.2 billion in North Carolina, and up to $790 million for its Kvandal South data center in Norway. Total liabilities stood at $12.4 billion on June 30.
Public equity is the cheapest money left. It funds the next campus without another tranche of secured debt, and it gives lenders a market-priced equity cushion beneath their loans. That is why the pipeline behind Nscale is so crowded.
There is a timing argument too. Our view at DataStorage.com is that the GPU supply balance is shifting: Meta Compute selling surplus capacity is one signal, and every gigawatt that comes online gives buyers more leverage. Neoclouds that list while scarcity still sets prices lock in valuations built on today's tight market. Those that wait risk listing into a buyer's market.
NVIDIA sits in the middle of all of it. It is already an Nscale shareholder through non-voting shares, is taking another $1.0 billion in this round, has backed Lambda, and supplies the hardware every one of these companies buys. When the chip supplier is also the investor, read backlog numbers with extra care, because demand and financing are partly coming from the same place.
CoreWeave set the template when it began trading on Nasdaq on March 28, 2025, selling 36.6 million shares at $40 and raising about $1.5 billion after cutting the size of the deal. The Information reports CoreWeave passed $5 billion of revenue in 2025 and has guided to more than $12 billion for 2026. Nebius also trades on Nasdaq. Behind them, three groups are lining up.
Bloomberg reported on August 24 that Lambda is in talks to raise up to $3 billion at a valuation of $12 billion or more, with an IPO as soon as 2027. The company expects more than $1.5 billion of revenue in 2026, appointed Michel Combes as CEO in May, and added a $1 billion credit facility and a $926 million term loan this year. Its timeline has already slipped once: Lambda hired Morgan Stanley, J.P. Morgan and Citi in 2025 for a listing as early as the first half of 2026.
Bloomberg reported in July that Crusoe was in talks to raise about $3 billion at a valuation around $30 billion, close to triple the more than $10 billion it was valued at in October 2025. Axios reported in August that Crusoe had held preliminary IPO conversations with banks. It has not filed, and it is still weighing private capital as an alternative.
Beneath the GPU clouds sits the real estate. Data Center Dynamics reports that operators Vantage, CyrusOne and DayOne are also preparing public offerings. They matter less directly to GPU buyers, but they compete for the same capital and the same power, and their pricing sets the floor for what neoclouds pay to host your clusters.
| Company | Status | Valuation marker | Scale signal |
|---|---|---|---|
| CoreWeave | Public, Nasdaq (Mar 2025) | IPO at $40/share, about $1.5B raised | Revenue above $5B in 2025; above $12B guided for 2026 |
| Nscale | S-1 filed Sep 18 2026, NYSE: NSCL | $14.6B Series C; notes cap at $30B | $103.4B active and contracted TCV |
| Lambda | Private, pre-IPO talks | Up to $3B raise at $12B or more | Above $1.5B revenue expected in 2026 |
| Crusoe | Private, early bank talks | About $3B raise near $30B | Over $10B valuation in Oct 2025 |
Most coverage of the Nscale filing asks whether the stock is worth owning. The more useful question for infrastructure teams is what these filings change about how you buy compute.
The first question to ask any provider, as Massed Compute founder and CTO Sunny Smith put it on Episode 7 of the DataStorage.com Podcast, is whether it owns the GPU. Many neocloud brands are brokers reselling someone else's capacity. A public filing is the first time you can verify ownership, delivery and financing claims instead of taking a sales deck on trust. Before you sign, pull these five things from the prospectus or ask the provider for them directly:
Nscale says many of its contracts give customers a right of first refusal on additional compute. In practice, when a new hall comes online, Microsoft and Anthropic see it before you do. Mid-size buyers are the flex layer. Plan on the assumption that on-demand Blackwell and Vera Rubin capacity stays tight at providers whose backlog is dominated by frontier labs, track live rental prices across providers, and keep at least two suppliers qualified.
Nscale's contract terms start only when GPU clusters are delivered, which protects the provider as much as the customer. If your roadmap depends on a go-live date, negotiate delivery-date credits, the right to reduce or exit a commitment if capacity slips past an agreed date, assignment and change-of-control terms (listings are often followed by acquisitions), and clear SLA credits. As supply loosens, shorter terms and multi-provider strategies gain value, so price the discount on a three-year commitment against the flexibility you give up.
The most expensive part of switching neoclouds is rarely the GPU contract. It is moving the data. Suppose your training corpus lives in Amazon S3 and you need to push 500 TB to a new GPU provider in a single month. On AWS's standard internet egress tiers ($0.09 per GB for the first 10 TB, $0.085 for the next 40 TB, $0.07 for the next 100 TB and $0.05 above 150 TB), that one copy costs about $28,800. Do it twice a year as capacity moves between providers and you have spent $57,600 on a toll booth, not on compute.
| Tier | Volume | Rate per GB | Cost |
|---|---|---|---|
| First 10 TB | 10 TB | $0.09 | $900 |
| Next 40 TB | 40 TB | $0.085 | $3,400 |
| Next 100 TB | 100 TB | $0.07 | $7,000 |
| Remaining 350 TB | 350 TB | $0.05 | $17,500 |
| Total, one 500 TB move | 500 TB | blended | $28,800 |
Keeping the canonical dataset on S3-compatible storage with no or minimal egress fees, such as Backblaze B2, Wasabi or Cloudflare R2 (each subject to its own fair-use terms), turns a provider switch into a network question instead of a budget question. It is the same logic behind our view that storage, not compute, is the anchor of the AI stack: GPUs are the part you can move. Your data is the part that decides whether you can afford to.
A neocloud IPO is a capital event for the provider. For the buyer, it is the first chance to read the fine print.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
