Inference startups are becoming the capital-intensive businesses they were meant to abstract away

AI inference providers built an attractive story around software: rent chips from clouds, schedule models efficiently and resell the resulting capacity without owning the underlying hardware. The Information reports that supply constraints are now pushing several of them into buying Nvidia GPUs and managing data-centre capacity themselves.

Fal has begun buying chips for leased facilities and is hiring a data-centre team; it may use debt as well as part of a recent $1.65bn equity raise. Together planned to reach 250MW of owned capacity by the end of 2026. Baseten is considering Nvidia purchases for sites it would manage. Modal, meanwhile, expanded from a handful of large cloud partners to roughly 30 additional neoclouds because it could not obtain enough Blackwell capacity at the price and volume it wanted.

The pressure is structural. Clouds prefer direct software customers to resellers, and some now offer competing inference services. Scarcity therefore exposes the weakness in an asset-light intermediary: its promised availability depends on suppliers that can ration it or move up the stack.

Owning chips solves that dependency by creating another. These companies must finance equipment before demand arrives, keep depreciating GPUs occupied and operate securely across physical facilities. The distinction between inference software and a neocloud begins to collapse. The durable advantage will have to appear in utilisation and model performance—not merely in access to capital that lets a startup acquire the same hardware as everyone else.