Overview
- D-Matrix began volume production of its Corsair inference cards in June 2026 and has started shipping hardware to hyperscalers, neoclouds and AI labs.
- Corsair tightly integrates compute and on-chip SRAM to move data shorter distances, a design meant to lower inference latency and energy use compared with GPUs.
- The company and cited Gimlet Labs research say Corsair can run small inference workloads up to 10 times faster and use up to five times less energy than a standalone Nvidia GPU.
- Experts caution SRAM-based chips cannot hold the trillions of parameters used by the largest reasoning models, so Corsair targets interactive workloads rather than the biggest training or model-hosting tasks.
- D-Matrix sells four-chip cards and rack systems with partners such as Arista, Broadcom and Super Micro, has Microsoft M12 among its investors, and plans a 4 nm successor called Raptor next year.