The AI chip market is splitting. General-purpose GPUs still win the flexible work, but the predictable, high-volume inference jobs are moving to silicon built for one customer and one workload. OpenAI’s Jalapeno is the clearest sign of that split.
OpenAI and Broadcom‘s Jalapeno chip posts up to 1.9x Blackwell inference throughput. But it does inference only, and it does not train models.
- 1.5x-1.9x more throughput on inference versus Nvidia‘s GB200 and GB300 racks.
- Inference only. Nvidia still owns training, where the big spend is.
- Broadcom is the real shift. AI chip revenue up 221% to $16.7B, custom silicon driving 73%.
Get the next one before it is old news
Independent analysis of cloud-native infrastructure, Kubernetes and data centre economics. No vendor spin.