AI Chip Startups Surge: 5-10 New Firms Yearly

Albert Reuther has been tracking AI chips since 2018, and the count keeps climbing. Reuther, a staff member at MIT Lincoln Laboratory Supercomputing Center, leads the Lincoln AI Computing Survey, known as LAICS, now six papers deep into cataloging commercial AI accelerators and measuring their peak performance against peak power draw. The project started because government sponsors of the lab's work kept asking about a sudden surge of new chips designed to speed up neural networks and machine learning tasks.

These accelerators don't just run AI models. The same hardware can model molecular behavior or speed up fluid dynamics simulations, work that eats enormous computing resources regardless of whether it touches machine learning at all. The hardware itself splits into several categories: CPUs for general-purpose jobs, GPUs, application-specific integrated circuits built for narrow tasks, field-programmable gate arrays that can be reconfigured after manufacturing, and dataflow accelerators built around moving data efficiently rather than executing fixed instructions.

Each architecture trades off flexibility against speed differently, and that tradeoff shapes which chips end up in data centers, defense systems, or research labs. Reuther's survey exists to keep track of that shifting landscape, one that the Lincoln Laboratory team started documenting before most people outside the chip industry noticed it happening.