d-Matrix Plugs Raptor XPUs Into NVIDIA NVLink Fusion
The inference chipmaker wants Raptor accelerators inside NVIDIA's MGX racks. The announcement names no launch date, no specs and no customers.
In short
d-Matrix will connect its next-generation Raptor XPUs to NVIDIA's rack platform over NVLink Fusion, letting them run in standard MGX racks alongside NVIDIA GPUs and CPUs.
At a glance
- Announced September 10, 2026 on NVIDIA's corporate blog, written by Jesse Clayton.
- Interconnect: NVLink Fusion for scale-up, Spectrum-X for scale-out, MGX as the rack architecture.
- The move covers the next-generation Raptor XPU line, not the d-Matrix part shipping today.
- NVIDIA places Vera Rubin NVL72, Vera CPUs, ConnectX-9 SuperNICs and BlueField-4 DPUs in the same platform context.
- Not disclosed: ship date, clock speeds, memory capacity, bandwidth, pricing, customers.
d-Matrix has decided its next chip generation will live inside someone else's rack. The company says its Raptor XPUs will attach to NVIDIA's AI platform through NVLink Fusion and deploy in standardized MGX racks, a commitment NVIDIA published on its corporate blog on September 10, 2026.
What NVLink Fusion buys a non-NVIDIA chip
NVLink Fusion exposes NVIDIA's scale-up interconnect to accelerators NVIDIA did not design. For d-Matrix that means Raptor lands in validated MGX racks, riding on the liquid cooling, power delivery and supply chain already qualified there, with Spectrum-X carrying scale-out traffic. The chip no longer has to arrive with a rack ecosystem of its own; it only has to fit one operators already install.
The case is economic rather than architectural. Cofounder and CEO Sid Sheth argues that inference demand keeps climbing while capital, time and energy stay finite, so borrowing a deployed, liquid-cooled design is the faster and lower-risk route to scale.
Why NVIDIA courts the company
Each third-party accelerator on NVLink Fusion strengthens NVIDIA's claim that the rack, not the individual chip, is the thing it sells. The post slots d-Matrix into the same lineup as Vera Rubin NVL72, Vera CPUs, ConnectX-9 SuperNICs and BlueField-4 DPUs. That shifts the contest away from whose accelerator posts the better benchmark and toward whose rack a data center can realistically power, cool and procure.
What the announcement leaves open
There is no ship date, no clock speed, no memory capacity, no bandwidth figure, no price and no named customer. It also does not say whether Raptor is meant for mixed racks beside NVIDIA GPUs or for XPU-only configurations. Anyone trying to size the engineering payoff is working without a single measurable number, only a choice of architecture.
A note on sourcing: this report rests on one source, NVIDIA's own blog, which is a party to the deal. We found no independent confirmation from d-Matrix or from a third party at the time of publication.
FAQ
What is NVIDIA NVLink Fusion?
It opens NVIDIA's scale-up interconnect to accelerators NVIDIA did not build, so third-party silicon such as the d-Matrix Raptor can operate inside NVIDIA racks.
When does the d-Matrix Raptor XPU ship?
The announcement gives no date. Raptor is described only as the next-generation d-Matrix XPU, with no availability or volume timeline attached.
Is d-Matrix still an NVIDIA competitor?
In inference accelerators, yes. In rack infrastructure it becomes a customer, since Raptor is meant to run inside NVIDIA's MGX architecture, power and cooling.
