Sam Altman, CEO of OpenAI, has expressed skepticism regarding the rapid expansion of specialized GPU cloud providers, known as 'neoclouds.' He noted that many of these companies are building massive compute capacities without clear revenue streams or confirmed buyers, raising concerns about the sustainability of this growth.
Financial data for major neoclouds reveals a complex situation. CoreWeave reported revenue of $985 million for the first half of 2026, yet incurred a quarterly loss of $626 million. These firms often rely on debt financing, using GPU hardware as collateral, with loan-to-value ratios typically ranging between 60% and 70%.
The industry faces a 'refinancing wall' between 2026 and 2028, when significant debt obligations come due. As newer GPU generations like Blackwell emerge, older chips such as the H100 are experiencing declining rental rates, which dropped from over $7 per hour in early 2024 to a median of approximately $3.30 by mid-2026.
To mitigate risks, NVIDIA introduced a 'backstop' mechanism, guaranteeing the rental of idle GPU capacity for certain providers. For instance, a deal with Sharon AI involves 40,000 GPUs over six years. However, NVIDIA recently scaled back a planned guarantee for an OpenAI and SoftBank data center project in Ohio from $250 billion to under $120 billion.
The current model mirrors historical vendor financing schemes, where manufacturers provide credit to customers to purchase their own equipment. Analysts are monitoring whether this structure effectively manages refinancing risks or merely concentrates them on NVIDIA’s balance sheet, especially as the company expands its investments into energy infrastructure.
