D-Matrix, a startup located three miles from Nvidia's Silicon Valley headquarters, has begun shipping its Corsair inference chip this month, claiming it can run AI inference workloads 10 times faster and using five times less energy than Nvidia's standalone graphics processing units, at least for smaller tasks. The key difference: instead of relying on scarce DRAM memory like most AI chips, Corsair builds memory directly onto the chip itself, sidestepping a critical bottleneck that's slowing the entire industry.

The AI chip market is increasingly crowded, with startups racing to carve out niches in a field dominated by Nvidia. D-Matrix co-founder and CEO Sid Sheth told CNBC that "this is a $1 trillion market in the making," and he has no intention of selling the company despite acquisition interest from larger players. The startup has raised around $500 million and is valued at approximately $2 billion, with Microsoft backing it through its M12 venture arm.

Sid Sheth

Why Is Memory Such a Bottleneck for AI Chips?

Most AI chips rely on DRAM, a type of memory that's in critically short supply. Only three companies manufacture DRAM at scale: Micron, Samsung, and SK Hynix. Demand for AI computing has far outpaced their production capacity, creating a traffic jam that's slowing innovation across the entire field. This shortage has become one of the most pressing constraints for data centers trying to scale their AI infrastructure.

D-Matrix took a fundamentally different approach. Instead of relying on external DRAM, Corsair uses SRAM, a faster type of memory that can be manufactured and integrated directly onto the chip itself at logic fabs like Taiwan Semiconductor Manufacturing Company (TSMC). "We're not running into a chokepoint around DRAM with our product because our product doesn't really rely on DRAM to be successful," Sheth explained.

This design choice mirrors the approach taken by other inference chip startups like Cerebras and Groq, which also integrate memory directly onto their chips. Cerebras held a blockbuster initial public offering last month, raising over $5.5 billion and reaching a valuation exceeding $50 billion. Groq's assets were acquired by Nvidia for $20 billion in December, making it the AI giant's largest purchase to date.

What Are the Trade-offs of D-Matrix's Design?

While Corsair's on-chip SRAM approach delivers impressive speed and efficiency gains, it comes with a significant limitation. SRAM cannot handle the massive reasoning models that now define cutting-edge AI. Rick Bahr, an adjunct professor of electrical engineering at Stanford University, explained the constraint: "That number of parameters just simply can't be put onto an SRAM-based design. That's the big challenge".

The largest language models from companies like OpenAI and Anthropic contain trillions of parameters, far too much data to fit on an SRAM-based chip. This means Corsair is optimized for a specific slice of the AI inference market: smaller, faster tasks where speed and energy efficiency matter more than raw model size. Think chatbots, voice agents, and agentic tools like Claude Code and OpenClaw, rather than the heaviest computational workloads.

When paired with an Nvidia Blackwell GPU, D-Matrix claims that Corsair can run inference 10 times faster, three times cheaper, and up to five times more energy efficiently than a standalone GPU, according to research from Gimlet Labs. This complementary approach suggests that D-Matrix's chips are designed to work alongside Nvidia's hardware, not replace it entirely.

How Is D-Matrix Bringing Corsair to Market?

• Hardware Configuration: D-Matrix sells four Corsair chips packaged together inside a card that slides into standard data center server racks, with each card costing tens of thousands of dollars.
• Memory Density: Sheth calls Corsair "the densest SRAM solution in the market today," with up to 128 gigabytes of SRAM memory in a single server, enabling significant compute density without relying on external memory stacks.
• Ecosystem Partnerships: D-Matrix has teamed up with Arista, Broadcom, and Super Micro to build a full rack-scale system called SquadRack for deploying its chips in AI data centers, offering a plug-and-play solution that differentiates it from competitors like Cerebras and Groq.
• Manufacturing and Roadmap: The Corsair chip is manufactured in Taiwan on TSMC's 6-nanometer process node, with the next-generation Raptor chip scheduled to launch next year on TSMC's more advanced 4-nanometer process, potentially produced at TSMC's Arizona facility.

D-Matrix began shipping Corsair to customers this month, with commitments from high-profile hyperscalers, neoclouds, and frontier AI labs eager to secure additional computing resources. About 90 percent of early customers are in the United States, while overseas customers are located in the Middle East and Southeast Asia.

"Quite often they sell to customers to use this stuff in conjunction with Nvidia," said Stacy Rasgon, a semiconductor analyst at Bernstein Research, adding that different chips excel at different tasks. "Sounds like he's got a fair number of actual, real customer engagements".

Stacy Rasgon, Semiconductor Analyst at Bernstein Research

Where Does D-Matrix Fit in the Broader AI Chip Landscape?

The inference chip market is becoming increasingly competitive, with multiple startups challenging Nvidia's dominance. Nvidia CEO Jensen Huang has emphasized that his company remains the leader in low-cost inference with its Vera Rubin system, arguing that speed alone doesn't determine superiority. "The reason for that is we integrate everything, we design everything from the ground up, we simulate the entire system and we use extreme co-design," Huang said at Computex in Taiwan.

D-Matrix's strategy differs from Nvidia's vertically integrated approach. Rather than attempting to replace Nvidia entirely, the startup is positioning Corsair as a complementary solution for specific inference workloads where latency and energy efficiency are paramount. This niche-focused strategy mirrors the approach of other successful AI chip startups that have found success by specializing rather than competing head-to-head across all use cases.

Microsoft's backing through its M12 venture arm is particularly significant, given the company's own ambitious chip development efforts, including its Maia 200 chip for AI inference, new PC processors built with Nvidia, and an in-house quantum computing chip announced recently. The investment signals confidence in D-Matrix's technology while also reflecting Microsoft's broader strategy of diversifying its AI infrastructure across multiple chip architectures and suppliers.

As the AI industry matures, the race is no longer solely about raw speed or raw power. Instead, it's increasingly about solving real infrastructure constraints like DRAM shortages, managing energy consumption in data centers, and delivering specialized solutions for specific workloads. D-Matrix's Corsair represents a pragmatic response to these challenges, offering a path forward for companies that need faster, more efficient inference without waiting for the global DRAM supply chain to catch up.