D-Matrix Begins Shipping SRAM-Based Corsair Inference Chip for Small Models, Claims Speed and Efficiency Gains vs Standalone Nvidia GPUs

D-Matrix begins shipping its Corsair inference chip to customers this month. The startup, located three miles from Nvidia's Silicon Valley headquarters, says the chip runs small inference workloads 10 times faster and uses five times less energy than a standalone Nvidia graphics processing unit. Corsair integrates memory and compute on a single chip using SRAM rather than DRAM.

Sid Sheth, D-Matrix co-founder and CEO, said the design avoids supply constraints that affect Micron, Samsung and SK Hynix DRAM production. The chip is manufactured at Taiwan Semiconductor Manufacturing Company on a 6-nanometer process. Four Corsair chips are packaged together on a card that fits standard data-center server racks and costs tens of thousands of dollars.

The card provides up to 128 gigabytes of SRAM memory per server. D-Matrix also partnered with Arista, Broadcom and Super Micro to offer a complete rack-scale system called SquadRack. Sheth said Corsair targets interactive AI applications such as chatbots, voice agents and tools like Claude Code and OpenClaw.

When paired with an Nvidia Blackwell GPU, the company cited Gimlet Labs research showing inference runs 10 times faster, three times cheaper and up to five times more energy efficient than a standalone GPU. D-Matrix has raised about $500 million and is valued at roughly $2 billion. Microsoft invested through its M12 venture arm.

Sheth said the company holds commitments from hyperscalers, neoclouds and frontier AI labs, including overseas customers in the Middle East and Southeast Asia. 5 billion in an IPO last month and is now valued above $50 billion. Nvidia acquired Groq's assets for $20 billion in December and released a new language processing unit based on that technology at its March GTC conference.

Rick Bahr, adjunct professor of electrical engineering at Stanford University, said SRAM cannot accommodate the trillions of parameters in the largest reasoning models. Sheth acknowledged the limitation and positioned Corsair for speed-focused inference rather than massive model size.

Nvidia CEO Jensen Huang said last week that the company remains the leader in low-cost inference with its Vera Rubin system.

D-Matrix's next chip, Raptor, is scheduled for launch next year on TSMC's 4-nanometer process and could be produced at the foundry's Arizona factory.