D-Matrix is betting that running AI models (inference) has different hardware requirements than training them, and that NVIDIA's GPU dominance in training does not automatically translate to inference. Microsoft invested. The chip ships to customers. Whether a compute-in-memory architecture can actually take market share from CUDA's gravitational pull is the $242M question.
Founded
2019
HQ
Santa Clara, CA
Total Raised
$242M raised
Founder
Sid Sheth, Sunil Gowda
Status
Active — shipping Corsair chips to early customers
Website
dmatrix.aiTHE ORIGIN STORY
Sid Sheth and Sunil Gowda were both veterans of Intel and AMD. They saw the same problem from the inside: the transformer architecture used in LLMs has a memory bandwidth problem.
When you run inference on a GPT-style model, you are constantly reading enormous weight matrices from memory, and the memory bandwidth bottleneck limits speed and drives up energy costs. NVIDIA's GPU architecture is great for training (parallel math) but less optimal for inference (random memory access patterns).
D-Matrix's Corsair chip uses a fundamentally different architecture — compute-in-memory — that embeds compute units directly in the memory arrays, eliminating the bottleneck.
WHAT THEY ACTUALLY DO
D-Matrix sells AI inference accelerator chips and systems designed specifically for large language model inference — running AI models at deployment, not training. Its chips use a compute-in-memory (CIM) architecture, which reduces the energy cost of moving data between memory and processing units, one of the fundamental bottlenecks in transformer model inference.
Revenue model is chip and system sales to data centers and cloud providers.
THE PRODUCTS
Corsair compute-in-memory AI inference accelerator, d-Matrix software stack (model compatibility layer), reference server designs
HOW THEY GREW
D-Matrix targeted hyperscalers and large enterprises deploying LLMs at scale, where inference cost is a real operating line item. Microsoft's strategic investment signals potential cloud deployment.
It positioned Corsair as the TCO (total cost of ownership) winner for inference workloads — not claiming to beat NVIDIA on raw performance, but on cost per inference query at scale.
THE HARD PART
NVIDIA has 90%+ market share in AI accelerators and enormous software ecosystem advantages (CUDA). Any alternative chip architecture requires software compatibility work to integrate with existing AI frameworks (PyTorch, TensorFlow).
Early AI chip companies (Graphcore, Cerebras) found this integration cost higher than expected. D-Matrix's inference-specific positioning is a deliberate attempt to find a wedge market where NVIDIA is not the obvious default.
MONEY TRAIL
Series A
2020 · Led by Playground Global, Nautilus Venture Partners
$44M raised
Series B
2022 · Led by Microsoft, Standard Investments
$110M raised
Series C
2024 · Led by SK Telecom, Microsoft
$110M raised
WHO BACKED THEM
Microsoft (strategic), Playground Global, Standard Investments, SK Telecom, Nautilus Venture Partners
Related Profiles
Head-to-Head
Compare d-Matrix vs another company.