d-Matrix Brings Raptor Inference XPUs to NVIDIA MGX Racks With NVLink Fusion

d-Matrix has announced a multiyear product-roadmap collaboration with NVIDIA to integrate its upcoming Raptor inference XPUs with NVIDIA NVLink Fusion. The first planned implementation is a Raptor-based system built on NVIDIA’s MGX rack reference architecture, aimed at large-scale, ultra-low-latency AI inference. Initial availability of Raptor in an MGX rack is targeted for the fourth quarter of 2027. 1211

What d-Matrix announced

NVLink Fusion is NVIDIA’s platform for linking third-party CPUs and accelerators into its NVLink scale-up networking and broader AI infrastructure. Under the partnership, d-Matrix plans to connect Raptor XPUs through NVLink while using Spectrum-X Ethernet for scale-out networking. 1

In practical terms, this gives d-Matrix a path to build its custom inference silicon into an established NVIDIA rack design rather than create and qualify an entirely separate rack, networking and cooling platform. NVIDIA and d-Matrix describe that as a faster, lower-risk route from custom silicon to large-scale deployment. 12

The planned MGX hardware stack

The planned Raptor system is intended to use NVIDIA’s latest MGX reference architecture. d-Matrix says that design will include:

• NVIDIA Vera CPUs
• NVIDIA NVLink switches
• NVIDIA BlueField-4 DPUs
• NVIDIA ConnectX-9 SuperNICs
• NVIDIA Spectrum-X Ethernet networking 2

MGX is designed as a modular rack architecture, and NVIDIA says the integration would place Raptor in a broadly deployed liquid-cooled AI infrastructure. 1

The announcement supports interoperability with NVIDIA’s rack-scale platform, but the cited company announcements do not establish a finalized, shipping Raptor configuration specifically branded as an “NVL144” system. Any 144-accelerator or memory-bandwidth figures reported ahead of launch should therefore be understood as roadmap projections, not deployed or independently benchmarked specifications.

Why a specialized XPU can complement GPUs

The partnership centers on heterogeneous inference: using different processors for different parts of serving a generative AI model.

A common split is to use GPUs for prefill, the compute-intensive stage in which a model processes an input prompt, and a specialized accelerator for decode, the latency-sensitive stage that generates tokens one at a time. d-Matrix positions Raptor for that decode-oriented role. 1819

Raptor follows the company’s production Corsair platform and extends d-Matrix’s memory-centric design. d-Matrix says Raptor uses a 3D DRAM-stacking approach that brings a DRAM memory die and an SRAM compute die together in one package. The company’s premise is that tighter coupling of memory and compute can better serve low-latency inference workloads. 2

That positioning matters: Raptor is not presented as a wholesale GPU replacement. It is intended to become a specialized component in an NVIDIA-based inference cluster, with each processor assigned work suited to its architecture.

Why NVIDIA is opening the rack to third-party silicon

For NVIDIA, supporting a third-party XPU through NVLink Fusion expands the range of hardware that can operate within its networking, rack and infrastructure ecosystem. For customers, the immediate appeal is integration: a specialized inference accelerator can be deployed with NVIDIA’s validated MGX design, NVLink connectivity and Spectrum-X networking rather than assembled as a standalone environment. 1

That makes the arrangement a platform strategy as much as a chip partnership. NVIDIA can keep infrastructure buyers on its rack-scale foundation while allowing specialized silicon to address a distinct inference bottleneck. The commercial outcome will depend on real-world availability, software support, cost and workload performance once Raptor ships.

Timeline and commercial context

d-Matrix expects Raptor to tape out by the end of 2026 and targets initial MGX rack availability in Q4 2027. These are forward-looking roadmap dates and may change. 911

The company’s current Corsair XPU platform is already in production. In a separate deployment announcement, Parasail said it would pair Corsair accelerators with NVIDIA Hopper and Blackwell GPUs, using GPUs for prefill and Corsair for decode. Parasail and d-Matrix claimed up to 10x faster token generation; that is a company-reported performance claim, not an independent benchmark for Raptor. 19

D-Matrix also reported a $275 million Series C financing in November 2025, valuing the company at $2 billion and bringing its total funding to $450 million. 16

The takeaway

The important development is not simply that d-Matrix has adopted an NVIDIA interconnect. It is that Raptor is being designed to enter NVIDIA’s MGX rack ecosystem as a purpose-built inference component. If the roadmap is delivered, operators could combine NVIDIA CPUs, networking and GPUs with a memory-centric XPU intended for the decode phase of AI serving—without building a separate rack-scale platform from scratch. 12

For now, the partnership is a roadmap with a 2027 availability target. The eventual value proposition will need to be tested against shipping hardware, real inference workloads and independently measured latency, throughput and total cost of ownership.