AI inference startups that rent cloud computing capacity and resell it to customers are beginning to buy chips themselves as compute becomes harder to secure. According to Sina Finance, citing The Information, the shift could change these companies’ cost structures, while reliance on smaller cloud providers has also exposed them to security risks. A cyberattack on Fireworks was reportedly traced to weak security at a small cloud provider, prompting some customers to consider switching providers.

Fireworks said its annualized revenue exceeded $1 billion in July, five times the year-earlier level, and its latest valuation is $17.5 billion. CoreWeave reportedly decided earlier this year to stop selling computing capacity to Baseten after launching its own inference service in March, although a person close to Baseten disputed that CoreWeave had reduced its supply. DigitalOcean CEO Paddy Srinivasan said the company prioritizes customers with end-user demand over firms that resell computing capacity.

Modal added 30 neocloud partners over the past year after struggling to obtain the desired quantity and pricing of Nvidia Blackwell chips from larger cloud providers. Fal has begun buying Nvidia chips for deployment in leased data centers and raised $1.65 billion earlier this month to expand computing capacity, including GPU purchases and AWS data center rentals. Together plans to have 250 megawatts of capacity in its own data centers by the end of 2026, while Baseten also plans to buy Nvidia GPUs for facilities it operates. Fireworks, by contrast, has told investors it does not currently plan to buy hardware, arguing that its software is its differentiator. Modal has discussed buying more hardware but CEO Erik Bernhardsson said the company sees software, not hardware ownership, as its main source of value.