Hardware Vendors Capturing Token Revenue

This shift changes the business from selling a box once to metering every prompt that runs on it. A chip vendor that also runs the serving layer owns the API, scheduler, and billing path, so it can keep earning as workloads grow after the hardware lands. Cerebras, Groq, and SambaNova already show the pattern, with token priced cloud or managed inference offers sitting directly on top of custom silicon.

• Cerebras has built the clearest bridge from hardware into service revenue. It offers self serve inference with pay per token pricing and describes its inference cloud as OpenAI API compatible, which lowers the switching cost from buying alternative hardware to simply routing live traffic onto Cerebras capacity.
  2 sacra 3 cerebras 7 cerebras
• Groq and SambaNova are doing the same move in different forms. GroqCloud sells tokens as a service, while SambaNova sells pay as you go cloud access and also a managed product that lets telecoms, data centers, and enterprises stand up their own branded inference cloud on SambaNova systems.
  4 sacra 5 groq 6 sambanova 8 sambanova
• For Fractile, software is what makes this expansion possible. Fleet management, runtime optimization, and developer tooling are the pieces that decide model placement, throughput, latency, and uptime in production. Those are the controls that turn a fast chip into a service customers keep paying for every month.
  1 sacra

The next step in AI chips is not just winning a benchmark, it is becoming the place where inference traffic actually runs. If Fractile gets early systems into frontier labs or neoclouds, the natural path is to layer on hosted capacity, usage based pricing, and developer APIs, and move from hardware ASPs toward a growing share of token revenue.