AI chip orders rise as financing for compute builds gets harder

On October 28, Google's parent company Alphabet released its Q3 earnings: operating activities generated roughly $42.3 billion in cash, while purchases of equipment and property consumed about $51.4 billion, leaving free cash flow at approximately negative $9.1 billion. Search, advertising and cloud businesses kept growing, but cash from operations no longer covered capital investment in the quarter. For neocloud providers such as CoreWeave and Nebius, which buy GPUs and rent out compute capacity, financing capacity directly determines how fast they can expand.

On October 22, Jensen Huang said demand for compute is extremely strong and these providers' service capabilities are solid, but what they lack is working capital; with AI investment expanding in multiple fields at once, capital needs have outrun the capital available. Three days later, Elon Musk offered a darker view. In recent weeks, he said, debt markets have almost stopped funding companies that lack large-scale revenue beyond AI data centers, and he predicted that more than 200 neoclouds will go through major failures and consolidation, with only a handful surviving. Huang believes the financing problem will be resolved; Musk thinks many of these firms may not last that long.

Over the past few years, CoreWeave, Lambda and other neoclouds have scaled their clusters through GPU-collateralised loans, and funding sources have expanded from private credit firms to large banks. Operators use equipment and customer contracts as the basis for financing, drawing procurement funds upfront and repaying debt with compute rental income. That structure enabled rapid expansion, but it also means the next round of purchases increasingly depends on how lenders value equipment and assess customer credit.

In early October, in the compute financing initiative Nvidia promoted, some banks and asset managers demanded stronger protections than the original plan provided. Lenders care not only whether GPUs can be rented out today, but also how much revenue they can generate over several years, whether customer payments will persist, and how much equipment would sell for after a default. The two sides already disagree on GPU value: Tony Trzcinka, a senior portfolio manager at Impax Asset Management, noted that banks typically value GPUs on 3-to-4-year depreciation schedules, whereas Nvidia argues that high-end GPUs can generate revenue for as long as ten years, and how long equipment stays usable is not the same as whether it keeps producing enough income. So far the dispute remains unresolved, with no firm conclusion on whether financing conditions are tightening.