AI data center racks are about to hold inference accelerators from multiple companies, not only NVIDIA's GPUs.
NVIDIA says d-Matrix will use NVLink Fusion to attach its Raptor accelerators directly into NVIDIA-designed racks. That makes d-Matrix part of a growing partner ecosystem built around the same rack-scale link.
That change matters because it alters how teams deploy and buy inference systems. Instead of bolting alternative chips to separate servers and moving data between boxes, operators can mix different accelerators inside the same rack.
Think of the rack as a shared workshop. NVLink Fusion is a private, high-speed corridor. It lets different processors exchange data instantly, instead of shuttling it between separate servers.
Practical note: d-Matrix plans to finish Raptor's design by late 2026, and systems that pair NVIDIA racks with Raptor accelerators are expected in 2027. This won't be for a laptop. These setups need datacenter racks, software integration and vendor support.
The real test is whether customers will choose mixed racks and create a viable market for third-party inference chips. Or will they stick with all-NVIDIA stacks because of software and ecosystem advantages? The answer will decide whether rack-scale interoperability becomes a real alternative, or just a technical possibility.