d-Matrix Raptor Joins NVIDIA NVLink Fusion

Key takeaways

d-Matrix announced that its next-generation d-Matrix Raptor inference XPU will integrate with NVIDIA NVLink Fusion and the NVIDIA MGX rack ecosystem. The companies described a multi-year roadmap, with initial Raptor-based MGX rack availability expected in the fourth quarter of 2027 rather than an immediate product shipment.
Everyone else is reporting that another accelerator joined NVIDIA’s rack ecosystem; we are explaining the workload split the architecture is trying to enable. d-Matrix specialises in inference, where a deployed model generates answers, while NVIDIA’s GPUs remain central to many compute-heavy phases. A shared rack design can let operators assign different stages to the hardware that fits them.
Availability and controls should be verified against the primary source before deployment.

d-Matrix Raptor: what changed

d-Matrix announced that its next-generation d-Matrix Raptor inference XPU will integrate with NVIDIA NVLink Fusion and the NVIDIA MGX rack ecosystem. The companies described a multi-year roadmap, with initial Raptor-based MGX rack availability expected in the fourth quarter of 2027 rather than an immediate product shipment.

Verified launch facts
Platform	d-Matrix Raptor XPU
Interconnect	NVIDIA NVLink Fusion
Rack ecosystem	NVIDIA MGX
Connectivity partner	Astera Labs
Expected rack availability	Q4 2027

The plan places Raptor alongside NVIDIA Vera CPUs, NVLink switches, BlueField-4 data-processing units, ConnectX-9 network adapters and Spectrum-X Ethernet. Astera Labs is also participating on custom connectivity, according to the d-Matrix announcement.

Everyone else is reporting that another accelerator joined NVIDIA’s rack ecosystem; we are explaining the workload split the architecture is trying to enable. d-Matrix specialises in inference, where a deployed model generates answers, while NVIDIA’s GPUs remain central to many compute-heavy phases. A shared rack design can let operators assign different stages to the hardware that fits them.

d-Matrix says one target is disaggregated AI coding: GPUs handle the compute-intensive prefill stage, while Raptor XPUs handle latency-sensitive decoding. That is a vendor-stated architecture, not an independently benchmarked production result. Buyers should demand end-to-end measurements on their own models before treating the proposed split as an economic advantage.

Raptor uses a memory-centric design that brings DRAM and SRAM compute together in a stacked package. The company says the chip is expected to tape out before the end of 2026 and is being evaluated by hyperscalers and frontier labs. Reuters separately confirmed the NVLink plan, the 2027 rack target and the absence of disclosed financial terms.

NVLink Fusion is the strategic hinge. It gives outside CPU and XPU designers a route into NVIDIA’s scale-up fabric and rack architecture, potentially broadening accelerator choice without forcing operators to invent every networking and mechanical layer. The value depends on software support, qualification and actual availability, not only physical connectivity.

The partnership also illustrates how the inference market is fragmenting by latency, power and workload shape. Interactive assistants, voice agents and coding tools may reward fast token delivery differently from large batch jobs. A mixed system can be attractive only if scheduling, observability and failure recovery do not erase the hardware benefit.

Procurement teams should separate the announced roadmap from products available today. The rack is expected in Q4 2027, while Raptor still faces design completion and qualification. Evaluation should include compiler maturity, model compatibility, memory limits, scaling behaviour and operational support across the combined vendor stack.

For Indian cloud and data-centre operators, heterogeneous racks may eventually offer more ways to tune inference economics. But power density, cooling, import timelines, service coverage and local workload demand will determine whether the design fits a deployment. A future availability date is a planning signal, not capacity that can be booked now.

The d-Matrix Raptor announcement is therefore best read as an ecosystem commitment. It gives the startup a path into a widely used rack framework and gives NVIDIA a custom-inference partner, but customers still need proof that the full system delivers lower latency or cost under production conditions. Published milestones should be checked against tape-out, sampling, qualification and rack availability rather than treated as one launch date.