On August 7, 2026, a message quickly spread in the US IPO subscription circle: according to sources cited by Bloomberg, London-based AI infrastructure company Nscale is promoting about $51 billion in total contract revenue to potential investors and plans to list on the US stock market as early as this September. This "newcomer in AI computing power rental," established just over two years ago, has once again ignited the AI-themed enthusiasm of the 2026 US IPO after Cerebras, SK Hynix, and Databricks. Today, we will use plain language to explain the ins and outs, profit logic, and potential risks of this company in one go.
From Mining Ruins to "AI Landlord"
Nscale's origin is quite dramatic. In 2024, it was spun off from Australian Bitcoin mining company Arkon Energy, and founder Josh Payne's background spans coal mines, recruitment platforms, and cryptocurrency mining. At that time, AI computing power demand had just exploded, and the accumulated power resources, data center operational experience, and "dare to burn money" gene of mining happened to become the original foundation for transformation.
After the spin-off, Nscale's financing rhythm was so fast that European peers took notice: completed $155 million Series A in December 2024; raised $1.1 billion Series B in September 2025; completed $2 billion Series C in March 2026, with a valuation reaching $14.6 billion; subsequently obtained $1.4 billion in GPU-backed loans, $790 million in debt financing for Norway's Narvik campus, and $900 million in revolving credit. In less than two years, a quasi-public company with a valuation targeting $25 billion took shape, with Silicon Valley political and business celebrities like Sheryl Sandberg and Nick Clegg on the board, making full "face" for the listing.
$51 Billion Contract in Hand: Is it a "Long-term Pie" or "Real Gold"?
The core number that excites Wall Street is about $51 billion in total contract revenue. How exaggerated is this number? For comparison: Nscale's full-year 2025 revenue was only about $33 million, Q1 2026 was about $37 million, and Q2 only surged to over $100 million. If annualized, the current revenue run-rate is about $400-500 million, with a huge "time gap" between this and the $51 billion contract total.
This requires understanding the essence of AI computing power rental business. Nscale's revenue mainly has two categories:
- Long-term reservation contracts: Similar to "chartering a plane seat," customers pay in advance for deterministic computing capacity, covering GPU cluster physical delivery, operations, high-performance network and storage hosting, etc., with Microsoft being the number one customer;
- On-demand cloud services: Public Cloud, Serverless inference, model fine-tuning and other products, "pay-as-you-go" based on actual GPU usage time, inference calls, bandwidth, etc.
Large multi-year contracts count future years' revenue into the total contract amount at the moment of signing, but actual conversion to statement revenue requires GPU installation one by one and tasks running one by one. Therefore, $51 billion is more like "grain in the warehouse," and whether each quarter can "harvest" steadily is the key to testing the company's quality.
The "Luxury Social Circle" Composed of NVIDIA, Microsoft, and OpenAI
Nscale's biggest confidence comes from its super allies behind it. NVIDIA is both a strategic investor and a core chip supplier - most of the company's deployed chips come from NVIDIA's Blackwell series, and it has signed contracts to purchase about 194,000 next-generation Vera Rubin GPUs. As of Q2 2026, Nscale has about 25,000 chips in operation, with a total of about 289,000 chips in operation and under contract, a scale directly comparable to leading "new cloud" vendors.
The gold content of clients is equally high. Microsoft invested $10 billion to build a data center hub in Sines, Portugal, jointly promoted by Nscale and Start Campus, initially deploying about 12,600 GB300 GPUs, with subsequent rumors of expansion to over 66,000 Rubin GPUs; Microsoft is also preparing to lease larger-scale AI computing capacity from Nscale in Norway and the UK. In addition, Nscale has deep cooperation with OpenAI in large-scale AI infrastructure projects in Europe. With both the "chip father" supplying chips and "cloud giants" placing orders, Nscale's positioning as an "AI computing arms dealer" is immediately established.
$25 Billion Valuation, On What Basis?
The valuation target given by investment banks is about $25 billion, a significant jump from the $14.6 billion in Series C in March. The core logic supporting the high valuation is the vertical integration of "power + chips + software":
- Power: The biggest bottleneck for AI data centers is electricity. Nscale currently has 831 megawatts of operational and contracted power and plans to add 10 gigawatts (1 gigawatt can power about 750,000 US households simultaneously), fully expanding in places like Norway's Narvik and West Virginia, USA;
- Chips: The computing power composed of nearly 300,000 operational and contracted GPUs, plus locking of new products like Vera Rubin, ensures future delivery capability;
- Software: Acquired distributed AI software company Anyscale for $1.65 billion in July, supplementing computing power scheduling and cluster management, making customers "get value for money and save usage."
This "from power generation to model running" all-encompassing approach is precisely Nscale's confidence to compare with "NVIDIA's favorite son" CoreWeave (current market value about $46.3 billion), and investors are also willing to bet on it as "the next CoreWeave."
Risks and Questions: Behind the Feast, Cold Thinking is Still Needed
After telling the story, plain language also needs to talk about risks. First, similar AI cloud stocks have recently faced general price pressure, with CoreWeave falling by about 25% in the past month, and sentiment in the high-valuation sector is swinging, making listing pricing face "headwinds"; second, Nscale's customer concentration is high, with large customers like accounting for a large proportion of revenue, and once order delivery rhythms change, earnings elasticity will be amplified; third, whether the profit margin of computing power rental can be sustained, the interest pressure from huge debt financing, and the close competition from rivals like CoreWeave, Lambda, and Crusoe are all unanswerable questions for IPO pricing.
It is reported that Goldman Sachs and JPMorgan are jointly preparing this IPO, but related discussions are still ongoing, and the listing time may also be delayed. For ordinary IPO investors, the most important observation window in the coming period is the pricing range in the prospectus, customer concentration disclosure, and the actual delivery progress of 10 gigawatts of power and 194,000 Vera Rubin GPUs - after all, on the AI computing train, there are both wealth creation myths and valuation de-bubbling lessons from past experience. Whether to "get on board" or "watch the show," it's better to let the bullet fly a little longer.
