Model-Specific Inference Chips

OpenAI has moved custom inference silicon from an internal project to a benchmarked production architecture. Jalapeno was co-designed with Broadcom around model-serving economics and reports higher throughput per watt and lower latency than general-purpose alternatives. Added to earlier transformer-specific chips and Waymo's workload-specific design, the change is structural: leading AI operators are beginning to shape silicon around their own models instead of adapting every workload to a general GPU.

Specialized AI silicon is appearing beyond one acquisition. A current comparison positions Etched's Sohu as a transformer-specific alternative to general GPUs, while Waymo is developing a custom chip around the inference needs of robotaxis and a research practitioner describes work on automated model-and-chip co-design. These are distinct approaches rather than one settled category, so the line remains on watch until independent benchmarks, production volumes or repeatable model-to-silicon workflows appear.

AMD acquired Taalas, whose model-specific integrated circuits encode model weights in silicon and claim inference rates up to 17,000 tokens per second. The mechanism could create a specialized hardware tier between general accelerators and fixed application ASICs, especially for high-volume agent workloads. It remains a weak signal because the evidence is one acquisition and early demonstrations; independent benchmarks, customer deployments or additional model-specific chips would confirm the category.