This Microsoft-backed startup wants to rival Nvidia in AI chips: Meet its Indian-American founders

D-Matrix, a California-based startup co-founded in 2019 by Indian-American entrepreneurs Sid Sheth and Sudeep Bhoja, is positioning itself as an alternative to dominant GPU vendors in AI inference. Backed by Microsoft’s venture arm M12, the company has raised roughly $500 million and is currently valued at about $2 billion. D-Matrix has begun shipping its flagship SRAM-based inference chip, Corsair, to selected customers — including hyperscalers, neoclouds and frontier AI labs — with most early adopters in the United States and others in the Middle East and Southeast Asia.

“We’re not running into a chokepoint around DRAM with our product because our product doesn’t really rely on DRAM to be successful,” Sid Sheth said, describing the architectural difference D-Matrix believes gives Corsair an edge on latency and energy efficiency.

Context and technical approach

D-Matrix’s strategy centers on inference workloads that demand low latency and smooth user interactions — chatbots, voice agents and AI coding assistants — rather than the large-scale training workloads that have driven GPU adoption. Corsair’s architecture uses on-chip SRAM to keep memory and compute closer together, a design the company says reduces data movement and power draw compared with conventional GPU-plus-DRAM approaches.

The company cites independent research by Gimlet Labs showing that Corsair, when deployed alongside Nvidia’s Blackwell GPUs, can run AI inference workloads up to 10 times faster, at one-third the cost and with as much as five times greater energy efficiency than a standalone GPU. Corsair is manufactured by Taiwan Semiconductor Manufacturing Company (TSMC) on a 6-nanometre node. D-Matrix plans a next-generation product, Raptor, on TSMC’s 4-nanometre process expected next year.

• Founders: Sid Sheth (CEO) and Sudeep Bhoja (CTO)
• Founded: 2019
• Capital raised: roughly $500 million
• Valuation: around $2 billion
• Manufacturing: TSMC 6nm (Corsair); Raptor planned on 4nm
• Partners: Broadcom, Arista Networks, Super Micro Computer (SquadRack rack-scale systems)

Sheth, a veteran semiconductor executive who previously led the Networking Business Unit at Inphi Corporation — helping grow it to more than $1 billion in revenue — and CTO Sudeep Bhoja, former CTO of Inphi’s Datacenter Business Unit with prior roles at Broadcom, are also building an ecosystem beyond individual chips. D-Matrix is working with Broadcom, Arista Networks and Super Micro Computer on larger rack-scale systems it calls SquadRack for AI data center deployment.

Outlook and challenges

Industry observers note the trade-offs of SRAM-centric designs. Rick Bahr, adjunct professor of electrical engineering at Stanford University, warned that SRAM-based chips may struggle to support the largest reasoning models: “That number of parameters just simply can’t be put onto an SRAM-based design. That’s the big challenge,” he said. The comment highlights a fundamental tension: SRAM architectures can dramatically improve latency and efficiency for many inference tasks, but they face limitations when confronting trillion-parameter models that dominate frontier generative AI.

D-Matrix has already started shipping Corsair and is targeting a mix of hyperscalers, neoclouds and AI labs while preparing Raptor and broader systems integration. “Building a computing solution for AI inference is going to be the grand prize,” Sheth added, framing the company’s ambition to claim a distinct role alongside incumbents rather than displace them outright.