Here's a contrarian bet that just paid off: While everyone else was obsessing over training AI models, a couple of Silicon Valley veterans decided the real money would be in running those models efficiently. On November 12, their company d-Matrix announced a $275 million Series C funding round at a $2 billion valuation, with some eye-popping claims about outperforming Nvidia (NVDA) GPU systems.

The funding round was co-led by BullhoundCapital, Triatomic Capital, and Singapore's Temasek sovereign wealth fund. Microsoft (MSFT) participates through its M12 venture arm, and the round attracted investors from four continents including Qatar Investment Authority and Singapore's Economic Development Board Investments. Existing backers Nautilus Venture Partners, Industry Ventures, and Mirae Asset also joined. That brings d-Matrix's total capital raised to $450 million since the company's 2019 founding.

The Memory-First Architecture Play

d-Matrix founder and CEO Sid Sheth and Chief Technology Officer Sudeep Bhoja started the Santa Clara, California-based company six years ago with what seemed like a contrarian thesis at the time. Everyone was focused on training AI models, but they built technology specifically for inference—actually running those models at scale in production environments.

"When we started d-Matrix six years ago, training was seen as AI's biggest challenge, but we knew that a new set of challenges would be coming soon," Sheth said. "We predicted that when trained models needed to run continuously at scale, the infrastructure wouldn't be ready."

Turns out they were onto something. The company now employs 250 people across offices in Santa Clara, Canada, Australia, India and Serbia. Both founders previously shipped over 100 million chips, so they're not exactly newcomers to the semiconductor game.

The Performance Claims

Now for the bold part. d-Matrix claims its inference platform delivers ten times better performance, triple the cost efficiency, and up to five times superior energy efficiency compared to GPU-based systems—which typically means systems built around Nvidia hardware.

The company's Corsair accelerators can apparently generate 30,000 tokens per second with 2-millisecond latency on Llama 70B models. That speed enables running 100-billion-parameter models within a single server rack, which is notable because it means you're not spreading compute across multiple racks and dealing with all the networking headaches that creates.

"The explosion in AI inference demand shows us that efficiency and scalability can be key contributors to revenue capture and profitability for hyperscalers and AI factories," said M12 Managing Partner Michael Stewart. "d-Matrix is the first AI chip startup to address contemporary unit economics in LLM inference for models of a range of sizes that are growing the fastest."

The Sustainability Angle

d-Matrix is framing its technology as addressing AI's sustainability challenges, claiming that its platform allows one data center to accomplish what traditionally requires 10 facilities. That's a compelling pitch in an era where data center power consumption is becoming a real political and economic issue.

"AI inference is becoming the dominant cost in production AI systems, and d-Matrix has cracked the code on delivering both performance and sustainable economics at scale," said Triatomic Capital General Partner Jeff Huber. "Their digital in-memory compute architecture is purpose-built for low-latency, high-throughput inference workloads that matter most."

Building the Ecosystem

d-Matrix recently unveiled its SquadRack reference architecture through partnerships with Arista (ANET), Broadcom (AVGO), and Super Micro Computer (SMCI). This ecosystem approach is designed to accelerate adoption among hyperscale, enterprise, and sovereign customers—basically showing potential buyers that d-Matrix chips can slot into existing infrastructure without requiring them to rebuild everything from scratch.

The real test, of course, will be whether these performance claims hold up in production environments and whether customers actually adopt the technology at scale. But with $450 million in funding and backing from Microsoft and major sovereign wealth funds, d-Matrix has the runway to find out.