AI infrastructure company Nscale has filed documents for a New York listing that could value the business at as much as $35 billion. Founded only two years ago, the company remains in a heavy investment phase and reported a net loss of $1.02 billion for the six months ended June 30, 2026. Its management has at one point warned in the filing that there is substantial doubt about the company’s ability to continue as a going concern.
The proposed listing has drawn attention not only because of the scale of Nscale’s losses, but also because of its overlapping commercial and financial ties with Nvidia. The chipmaker is Nscale’s largest supplier as well as an investor, and could become one of its largest shareholders after the IPO. How Nscale explains the relationship between future contracts, capital commitments and realised revenue will be central to investors’ assessment of the AI cloud company.
Nscale bets on turning $103bn of agreements into revenue
Headquartered in London, Nscale is led by founder and chief executive Josh Payne. The company was spun out of Australian cryptocurrency miner Arkon Energy in 2024. It develops or helps finance data centres designed for AI computing, then leases GPUs and related computing capacity to customers.
Its board includes former Meta chief operating officer Sheryl Sandberg, former UK deputy prime minister Nick Clegg and former Yahoo president Susan Decker. In its listing documents, Nscale disclosed approximately $103 billion in contracts and commercial arrangements, underpinning its pursuit of a valuation of up to $35 billion. A significant portion of those agreements has not yet been finalised, however. Their conversion into revenue will depend on customer demand, the pace of data-centre construction and access to financing.
Nscale’s revenue growth has already lagged the expansion of its losses. The latest filing highlights the gap: the company lost $1.02 billion in the six months to June 30. The documents also show that management used the phrase “substantial doubt about going concern”, indicating that Nscale needs continued financing or stronger operating cash flow to maintain its expansion plans.
Nvidia joins a $3.1bn financing round
On September 15, Nscale signed a $3.1 billion financing subscription agreement. Nvidia plans to subscribe for $1 billion of the financing through a combination of convertible notes or non-voting shares. The final investment structure and Nvidia’s ownership percentage will depend on the IPO pricing and issuance structure.
Nscale’s UK accounts also show that Nvidia bought warrants over more than 157,945 Nscale shares for $60 million in October 2025. Nvidia separately agreed to provide up to $860.3 million of guarantees for Nscale’s lease obligations at a facility in Ward County, Texas.
Nvidia chief executive Jensen Huang said publicly in April that companies such as CoreWeave and Nscale would not have reached their current scale without Nvidia’s support. Nvidia’s relationship with Nscale therefore extends beyond chip sales to equity investment, financing arrangements and lease guarantees.
Those arrangements have also raised a broader market question: how much of the AI infrastructure sector’s growth reflects genuine end-customer demand for computing capacity, and how much is supported by financial flows within the supply chain. Nvidia may invest in, lend to or guarantee customers that then use the funds to purchase Nvidia GPUs. In some transactions, it has also invested in AI companies buying capacity from new cloud providers. For investors, separating external demand and firm customer commitments from growth supported by industry financing can be difficult.
CoreWeave offers a public-market comparison
CoreWeave is Nscale’s closest listed comparable. Before its March 2025 IPO, CoreWeave disclosed a net loss of $863 million for 2024. Its revenue was highly concentrated: Microsoft and another customer widely understood to be an Nvidia client together accounted for 77% of revenue.
CoreWeave priced its IPO below the marketed range and ended its first trading session roughly flat, but its shares doubled over the following months. The example shows that losses, customer concentration and close supplier relationships do not automatically prevent a company from attracting a high valuation. They do, however, make the valuation more sensitive to the pace of growth, financing conditions and customer renewals.
The market backdrop for Nscale is different from the one CoreWeave faced when it listed in March 2025. Debate over the pace of AI development and the returns on capital spending has intensified. A semiconductor index fell by nearly 6% in a single recent session, while higher US Treasury yields have made investors more cautious about unprofitable growth companies.
At the same time, data-centre development is facing greater community and political resistance in several regions around the world, with some policy discussions focused on restricting rapid expansion. Nscale’s model depends on continued data-centre construction and large GPU purchases. Permits, energy availability, electricity costs and actual customer utilisation could therefore affect both the fulfilment of its contracts and its cash flow.
Nscale could test public-market appetite for AI infrastructure
Nscale is moving towards an IPO ahead of Anthropic and OpenAI and could become an important reference point for investors valuing AI infrastructure businesses. A stable IPO price and trading performance could influence how the market assesses valuation ranges for later large AI listings. A weaker performance would provide an earlier indication of the constraints facing high-investment, low-profitability models in public markets.
The filing places several central questions on the same balance sheet: how much capital Nscale needs to complete its data-centre expansion, how much of the $103 billion in contracts can become recognised revenue, and how far Nvidia’s financing and guarantees support the company’s reported scale. The final IPO price, post-listing ownership structure and customer revenue concentration will be key measures for the market.
Gemini 4 moves into early post-training
Koray Kavukcuoglu, head of Google DeepMind, said the company’s next flagship model, Gemini 4, has entered an early post-training phase and produced “promising results”. Google plans to release an early version as soon as possible before the end of this year, but has not announced a firm launch date.
The comments marked Kavukcuoglu’s first public appearance since taking over as head of DeepMind. Gemini 4 is being developed as Google, Anthropic and OpenAI compete for the frontier-model market. Model performance, inference costs and the speed of commercialisation will all influence enterprise customers’ choices.
Altman and Amodei call for international AI safety mechanisms
OpenAI chief executive Sam Altman and Anthropic chief executive Dario Amodei addressed the United Nations Security Council separately, calling for stronger international cooperation on AI safety. Altman proposed common standards for assessing AI capabilities and measuring risk, along with secure channels for sharing information about threats so that lessons can be gathered before incidents escalate.
Amodei outlined a phased approach. It would begin with bans on specific uses, including using AI to develop biological weapons, followed by systems for enforcing and verifying commitments. He also called for common testing standards and reporting requirements for serious AI incidents. Both executives focused on model testing, government-to-government information sharing and incident disclosure rather than relying solely on voluntary commitments by individual companies.
Claude agents identify a new phage-enzyme system
Anthropic said about 950 Claude agents spent 21 hours searching large DNA sequence databases and screened more than 200,000 reverse transcriptases down to 20 leading candidates. One agent identified repeated DNA sequences near a gene for an unusual reverse transcriptase.
Anthropic named the system “array-associated reverse transcriptase”, or ART. It contains three components, and its repeated sequence array resembles a CRISPR array. Human scientists still carried out the laboratory validation, and the system’s specific biological function has not been established. Feng Zhang, a CRISPR pioneer at the Massachusetts Institute of Technology and the Broad Institute, said the finding demonstrated one possible route for AI agents to participate in biological research, although its practical value requires further experimental confirmation.
China attracts a larger share of its top AI talent
A tracking study by Carnegie China found that 69% of more than 10,000 leading AI researchers with Chinese undergraduate educational backgrounds now work in China, up from 57% in 2022. Researchers Damien Ma and Binyi Yang said the increase in domestic AI industry jobs and tighter US visa policies may be the main drivers of the shift.
Based on the study’s sample, China now accounts for 41% of the world’s leading AI talent, compared with 27% in 2022. The US share has fallen from 46% to 34%. Researchers with Chinese undergraduate backgrounds account for 57% of the total sample.
The US still retains 89% of researchers with domestic undergraduate backgrounds, well above China’s 69%, and recorded a net increase of 2,145 researchers in 2025. China recorded a net decline of 1,729 over the same period. The gap in talent flows between the two countries is narrowing, however: for every US-background researcher working in China, about 30 China-background researchers work in the US, compared with a ratio of 46 to one in 2022.
For Nscale, talent, computing capacity and capital remain the three conditions underpinning its expansion. Whether the company can turn its not-yet-finalised agreements into revenue, and whether Nvidia-backed financing can cover the cash required for continued expansion, will remain central questions as more details of the proposed listing emerge.