Seven years in, d-Matrix has a rack to call home. The Santa Clara startup announced Thursday that its forthcoming Raptor memory-centric XPU will plug into Nvidia's NVL144 MGX liquid-cooled rack — the same NVLink switch trays, ConnectX-9 SuperNICs, and Spectrum-X Ethernet that carry Vera-Rubin silicon. A single rack will hold 144 Raptor cards, each packing 2.3 TB of 3D-stacked DRAM at 100 TB/sec internal bandwidth, for an aggregate 7.2 petabytes per second per rack.

For d-Matrix CEO Sid Seth, who co-founded the company in 2019 "when not many people even knew what inference was," the deal sidesteps the cost of building his own liquid-cooled chassis. "This is truly the fastest way we feel of getting this amazing path-breaking technology that we are building at d-Matrix into the market with quick scale, with a robust supply chain to back it up and with a partner who is deployed in every datacenter across the world," Seth told reporters this week.

The company has raised more than $500 million in funding — Microsoft M12 among its backers — and shipped its first inference accelerator, Corsair, in June. Corsair claims tenfold speed-ups over Nvidia GPUs alone on token-decode workloads, the phase where the model generates responses rather than ingesting context. Seth has positioned the design against the explosion in low-latency inference driven by agentic tools like Anthropic's Claude Code. "Over the last twelve months, there has been an explosion of low-latency inference across many different applications," he said. "Now, with the arrival of cybersecurity applications, low latency tokens are in huge demand."

What's actually inside Raptor?

Raptor, detailed at Hot Chips 2026, fuses a 4-nanometer TSMC compute die with a custom DRAM die at a 36-micron pitch — a chip-on-wafer variant borrowed from HBM packaging. Seth said Raptor will tape out before year-end and ship in the fourth quarter of 2027. Throughput claims from the Hot Chips demo: 3,000 tokens per second per user on Z.ai's GLM 5.2, and 1,000 tokens per second per user on Moonshot AI's Kimi K3, with the vendor scaling the workload across eight racks.

The Corsair reference design — SquadRack — already shipped with Arista Networks, Broadcom, and Supermicro after d-Matrix's April acquisition of GigaIO's datacenter business.

Why Nvidia, why now?

Nvidia's AI factory business is the prize. CFO Colette Kress disclosed during the chipmaker's Q2 2027 earnings that the per-gigawatt revenue opportunity has climbed from roughly $18 billion under 2022's Grace-Hopper rackscale systems to about $40 billion under the upcoming Vera-Rubin generation. d-Matrix's MGX plug-in inherits that trajectory. "The beauty of this solution is we take the Raptor trays, plug them into the same NVL144 MGX rack architecture, which is widely deployed across many datacenters, and we get instant access to those datacenters," Seth said.

NVLink Fusion ports on the MGX platform also accommodate Groq 3 LPX accelerator racks and third-party silicon. Jesse Clayton, principal product manager for Nvidia's Data Center GPU business, framed the arrangement as a "completely fungible" platform — "vertically integrated, but horizontally open." Vendors choose how much of Nvidia's stack to consume, then drop in their own compute.

What comes after Raptor?

Lightning, the next entry on d-Matrix's roadmap, layers a multi-high DRAM stack on top of the 3D-RAM foundation. It, too, will reach market through Nvidia's MGX channel, Seth confirmed.

He framed the bet in capital terms. "Our entire approach is predicated on doing more with the capital people deploy in our compute," Seth said. "We are able to run really, really fast compute. We do more inference with a little amount of time, and we have made a very energy-efficient solution with the memory-centric computing. That allows us to do more with less of these resources — money, time, and energy. We can hopefully, over time, alleviate the need to build out more datacenters."

Whether the 7.2 PB/sec story holds at hyperscaler scale — and whether d-Matrix hits its year-end tape-out on the way to a Q4 2027 Raptor launch — will set the cadence for Lightning's arrival.