d-Matrix Plugs Its AI Chips Into NVIDIA's Factory Platform

Key points

• D-Matrix announced its next-generation Raptor XPUs will connect to NVIDIA's AI factory platform using a technology called NVLink Fusion.
• NVLink Fusion delivers 3 terabytes per second of all-to-all bandwidth per chip and cuts chip-to-chip latency to one-third of standard Ethernet.
• Partners already named for NVLink Fusion include AWS, Arm, Intel, Samsung and Marvell, among others.
• The agreement lets d-Matrix skip building its own rack, cooling and networking infrastructure from scratch.
• D-Matrix also plans to integrate NVIDIA's Vera CPUs, ConnectX-9 network cards and BlueField-4 data-processing units into its systems.

Building a custom AI chip is hard. Getting that chip into a working data centre is often harder and costlier than the chip itself.

That's the gap NVLink Fusion is designed to close. NVLink, in plain terms, is a high-speed connection technology NVIDIA uses to link chips so they can share data very quickly. The Fusion version extends that link to chips made by other companies. We first covered the programme on 26 August when NVIDIA opened NVLink Fusion to outside chipmakers; d-Matrix's adoption is an early test of whether that opening translates into real deployments.

D-Matrix makes what it calls XPUs, pronounced as separate letters: specialised processors built for a specific job. Its Raptor XPU targets AI inference, the step where a trained model actually answers questions or processes requests, rather than the earlier, costlier training phase. Sid Sheth, d-Matrix's cofounder and CEO, said at a press briefing that "demand for inference is soaring, but capital, time and energy remain finite."

What does this mean in practice?

Without NVLink Fusion, d-Matrix would need to design its own rack enclosures, certify its own cooling systems, build high-speed networking and manage its own supply chain before a single data centre could buy its chips. That work takes years.

By adopting NVLink Fusion, d-Matrix's Raptor chips slot into NVIDIA's MGX rack design, a standardised, liquid-cooled enclosure that data centres already buy and operate. The same physical rack can hold NVIDIA GPUs alongside d-Matrix's inference chips. A data centre builds the infrastructure once rather than maintaining separate hardware families. For buyers, that's a meaningful reduction in both upfront cost and ongoing complexity.

The numbers behind the technology are real. Sixth-generation NVLink carries 3 terabytes of data per second between chips, handles ten times more data packets per second than off-the-shelf Ethernet, and cuts chip-to-chip latency to one-third of what standard Ethernet delivers. For inference work, where speed directly affects how quickly a user gets a response, those figures matter.

Who else is involved?

NVIDIA's NVLink Fusion partner list includes d-Matrix, AWS, Arm, Intel, Fujitsu, SiFive, Alchip, Astera Labs, GUC, Marvell, MediaTek, Samsung, Cadence, Synopsys, Ayar Labs and Lightmatter. The platform supports Arm (the architecture inside most smartphones), x86 (which powers most laptops and servers) and RISC-V (an open-source design gaining ground in specialist hardware).

NVIDIA describes the result as a "semi-custom AI factory": a data centre built on NVIDIA's infrastructure but running a mix of NVIDIA and third-party processors matched to each job.

Should you worry about lock-in?

The honest read is that this benefits NVIDIA at least as much as d-Matrix. Every partner chip that relies on NVLink deepens customer dependence on NVIDIA's networking and supply chain. D-Matrix gets a faster route to market; NVIDIA gets a broader platform that's harder to replace. Neither side's hiding that arrangement, which is at least straightforward. What to watch next: whether data centres actually specify Raptor XPUs in new builds, or whether the partnership stays on paper longer than either company would like to admit.