Nvidia doesn't have to win the inference chip race to win the inference chip race

AI inference, the running of trained models rather than the training of them, is becoming a toll road, and the toll is collected at the rack.

d-Matrix, a chip startup that has been publicly building silicon to compete with Nvidia's GPUs on inference, announced on 10 September 2026 that it will wire its next-generation Raptor XPUs into Nvidia's NVLink Fusion fabric. Systems are targeted for the fourth quarter of 2027. The rival is now a tenant of the platform.

Nvidia's play is structural. By opening NVLink Fusion to third-party accelerators, Nvidia turns its interconnect into the connective tissue every serious inference chip has to plug into. The NVLink Fusion ecosystem already counts AWS, Arm, Intel, Fujitsu, and MediaTek as partners. The economics follow the fabric, not the silicon. A customer buying an inference server is still buying Nvidia's rack, networking, and software plumbing even when the math runs on a competitor's chip.

The $110 million d-Matrix raised from Microsoft, reported by Reuters via Investing.com, looked like validation of an alternative to Nvidia. Read against the same week's announcement, it is better understood as a subscription to Nvidia's rack.