Nscale Goes Public: $140M in Revenue, $103 Billion in Promises
The S-1 of a two-year-old British AI hyperscaler shows a backlog 700 times larger than its actual revenue.
In brief
Nscale, a London-domiciled AI compute company, has filed its S-1 with the SEC for a NYSE listing under the ticker "NSCL." The filing reveals $140.6M in revenue for the first half of 2026 against $1.02B in net losses, versus $103.4B in signed take-or-pay contracts, dominated by Microsoft and Anthropic. It's the most detailed document published to date on the real economics of an AI "neocloud."
🍺 Bar-stool version
Nscale has existed since May 2024, makes $140 million in revenue, and loses a billion dollars in six months. So far, nothing out of the ordinary for the sector. What is out of the ordinary is that it has signed $103 billion in future contracts, roughly Croatia's GDP, promising to deliver data centers it hasn't built yet with financing it admits, in black and white, isn't locked in. The prospectus is worth reading because it puts in writing, under legal liability, what everyone has been saying out loud for two years about AI infrastructure.
Key takeaways
- 1
Revenue of $140.6M in H1 2026 versus $10.4M a year earlier (+1,252%), with a net loss of $1,020.1M and an accumulated deficit of $1,860.9M.
- 2
$103.4B in active and contracted take-or-pay contracts as of August 31, 2026, of which only $2.6B is already active, with a weighted average duration of 5.7 years.
- 3
Two clients dominate: Microsoft (up to $43.8B through 2033) and Anthropic (up to $44.6B); the top client accounted for 52% of revenue in H1 2026 and 73% in 2025.
- 4
Nscale acknowledges it has not secured any binding financing commitments for the deployments owed to Anthropic at the Monarch Compute Campus in West Virginia (2,250 acres, up to 8 GW gross).
- 5
Only 25,000 active GPUs, versus 461,000 active and contracted; 1.37 GW of active and contracted capacity, with a line of sight to 10 GW.
- 6
The acquisition of Anyscale, the commercial company behind Ray (740 million cumulative downloads), is set to close at IPO time and add ~200 employees.
- 7
Two material weaknesses in internal controls are disclosed, unresolved, with remediation spread across 2026-2027.
What the filing says
Nscale Limited, an English and Welsh company based in Mayfair, filed its S-1 on September 18, 2026. It will rebrand as Nscale plc ahead of the offering and will seek to list on the NYSE under the ticker "NSCL." Goldman Sachs, J.P. Morgan, and Morgan Stanley lead a syndicate of more than twenty banks.
The company describes itself as a "full-stack AI hyperscaler": it buys electrified land, generates its own power behind-the-meter, builds water-cooled modular data centers, deploys NVIDIA GPUs, and operates the software layer on top. Founder Josh Payne's central argument is that every layer is a bottleneck and that only vertical integration can bring down the cost per token.
The company was spun off from Arkon Energy in May 2024. In two and a half years, it claims to have grown from $100M to over $103B in total contract value, from 750 MW to over 10 GW of power pipeline, and from 40 to over 1,000 employees.
The portfolio is concentrated in regions with cheap, renewable electricity: Norway, Iceland, Portugal, plus US and APAC sites. Nscale claims it pays roughly 70% less for electricity there than in major US markets.
The gap between the backlog and the income statement
The GAAP figures are brutal. $140.6M in revenue for the first half of 2026, $189.6M in cost of revenue excluding depreciation, $217.7M in general and administrative expenses, $174M in depreciation. Operating loss: $492M. Net loss: $1,020.1M, including $457.1M in fair-value adjustments. Adjusted EBITDA remains negative at -$199.2M.
On the other side, $103.4B in signed TCV. But the prospectus is explicit about how to read it: only $2.6B is "active," meaning backed by capacity already in service and generating revenue. The rest depends on data centers yet to be built, GPUs yet to be delivered, and performance milestones yet to be met.
Concentration is extreme. A single client accounted for 52% of revenue in H1 2026, 73% in 2025, and nearly all of it in 2024. The Microsoft ($43.8B through December 2033) and Anthropic ($44.6B) agreements together make up the bulk of the backlog.
One detail not to miss: the Anthropic agreements require Nscale to use "best efforts" to obtain financing for the GPUs and infrastructure of the Monarch Compute Campus. As of the prospectus date, no binding commitment has been secured. If a tranche is delayed, Anthropic can terminate it without penalty.
Power-first, Ray, and the sovereignty card
Nscale's industrial thesis is that the real rationing isn't about chips but about power. Hence the March 2026 acquisition of AIPCorp and the Monarch Compute Campus in Mason County, West Virginia: 2,250 acres, a gas-fired micro-grid built with Caterpillar, a target of 2 GW gross for the first half of 2028, and an announced trajectory toward 8 GW gross by 2031.
On the software side, Nscale agreed in July 2026 to acquire Anyscale, the commercial company behind Ray, the open source distributed computing framework (740 million cumulative downloads, including 174 million in Q2 2026 alone). The idea: own the layer that decides how resources are consumed, in addition to the resources themselves.
The "open-by-design" card is embraced as a low-cost acquisition channel: developers adopt Ray for free, then migrate to the paid control layer and Nscale infrastructure. The prospectus does warn, however, that Ray is governed by the PyTorch Foundation and that Nscale does not control it.
The UK domicile is presented as a commercial asset for the European sovereign market. Nscale sees it as a differentiator against American hyperscalers, while acknowledging that this segment remains embryonic, with long sales cycles and political volatility.
The red flags in the prospectus
Nscale discloses two material weaknesses in its internal controls: lack of formal design and documentation of controls (including ITGC) and insufficient accounting staff. Remediation is spread across 2026 and 2027, with no guarantee of success.
The model is extremely capital-intensive and financed through stacked debt: a GPU Financing Facility of ~$1.4B, a Revolving Credit Facility of $900M, Kvandal South, Macquarie Iceland, Ward County ($1.85B), North Carolina ($1.2B), plus ~$2.54B in upfront lease payments via Dell Financial Services master agreements. In September 2026, a Subscription Agreement of at least $3.1B includes $1B earmarked for NVIDIA in non-voting shares.
One episode worth noting: in April 2026, OpenAI withdrew from the Stargate Norway and Stargate UK partnerships, with Microsoft subsequently taking over the extra capacity in Norway. The prospectus mentions this under the risks tied to strategic alliances.
Add to that the sector's usual but concrete risks: dependence on NVIDIA (three suppliers accounted for 53%, 37%, and 10% of purchases in H1 2026), data center moratoriums in New York State, audit directives in Texas, local opposition, unstable tariffs, and 18-to-24-month lead times on electrical equipment.
“Infrastructure is the defining constraint of the age, and it will decide which nations lead, which companies win, and which products become ubiquitous.”
“In only two and a half years, we have scaled from $100 million in total contracted value to over $103 billion.”
“As of the date of this prospectus, we have not obtained binding commitments for any of the financings required to fund performance under the Anthropic Services Agreements.”
Why it matters
An S-1 is the one moment when a company in this sector must put everything in writing under legal liability. Nscale's offers a rare X-ray of AI compute economics: a 700-to-1 ratio between signed contracts and realized revenue, a backlog resting on two clients, and future deliveries whose financing remains, by the issuer's own admission, unsecured. It's the mechanics of the entire neocloud wave, laid bare. The critical angle is obvious: TCV is not revenue, it's an option on execution. It assumes permits get approved, turbines arrive, GPUs get delivered, banks follow through, and Microsoft and Anthropic don't change their minds, even though their contracts specifically include termination rights in case of delay. OpenAI's withdrawal from Stargate Norway and Stargate UK is a reminder that these commitments are reversible. If the IPO goes through, public markets will have to weigh the power-first thesis, appealing and well-documented as it is, against the reality of a balance sheet burning a billion dollars every six months.
Read next
McKinsey: 13 Frontier Technologies, and the AI That Absorbs Everything
The fifth edition of the Technology Trends Outlook puts $223 billion on energy, $124 billion on AI — and reveals a far less discussed signal on the jobs front.
DHH: "I Haven't Written a Single Line of Omarchy Quattro's Code"
The creator of Ruby on Rails, long a skeptic, tells Lex Fridman how in nine months he went from hand-chiseling Ruby to running 16 agents in parallel — and why he thinks Linux is going to win the desktop.