D-Matrix, an AI inference hardware startup, has entered full production with its Corsair platform, manufactured through TSMC’s N6 fabrication process. The chip claims performance figures that should make Nvidia investors pause: up to 10x faster AI inference and 5x better energy efficiency compared to Nvidia’s standalone GPUs for generative AI workloads.
What D-Matrix is actually building
The Corsair platform uses an architecture called 3DIMC, which stands for digital in-memory compute. Traditional GPUs have to constantly shuttle data between memory and processing units, creating a bottleneck. D-Matrix’s approach essentially puts the pantry inside the kitchen.
The result, according to the company’s benchmarks, is the ability to process 30,000 tokens per second at 2 milliseconds per token for Meta’s Llama 70B model.
D-Matrix raised $275 million in a Series C funding round in November 2025, pushing the company’s valuation to $2 billion. Microsoft’s venture arm M12 participated in the round.
