d-Matrix Moves Its Corsair Inference Accelerator Into Full Production, Betting On-Chip SRAM Beats the GPU at Decode
d-Matrix says its Corsair inference accelerator is in full production, pairing on-chip SRAM with GPUs to cut AI response times by more than 10x.
Overview
The AI-chip startup d-Matrix announced on June 9, 2026 that its Corsair inference accelerator platform is in full production, with hardware set to ship in volume to priority hyperscalers, neoclouds, and frontier labs, according to the company’s release distributed through PR Newswire. Rather than chase Nvidia in raw training throughput, d-Matrix is targeting the narrower problem of inference with an architecture that embeds processing inside the memory — a design it pitches as a faster, more efficient complement to the GPU.
What We Know
d-Matrix says Corsair is built to accelerate the decode phase of large language model inference, leaving the compute-heavy prefill stage to GPUs. “GPUs dominate the compute-intensive prefill portion of the workload, while Corsair excels at the decode phase,” the company stated in the release. Pairing the two in a heterogeneous cluster can “speed up AI model response time by more than 10x,” according to the same statement.
That figure traces to a third-party benchmark. “Independent testing by Gimlet Labs demonstrated that a baseline 24-second response time was reduced to less than two seconds when pairing Corsair accelerators with GPUs,” the release said. The same announcement, carried by Yahoo Finance, named TSMC and Alchip Technologies as manufacturing partners and Arista, Broadcom, and Supermicro as ecosystem partners. Founder and CEO Sid Sheth framed the launch around the shift in AI workloads, saying d-Matrix “built Corsair specifically for this moment, the Age of AI Inference” and that “Corsair takes off from where the GPU leaves off,” per the release.
The distinguishing technical bet is where the memory sits. Instead of pairing a processor with separate high-bandwidth memory the way a graphics card does, Corsair “embeds processing components into the memory,” and d-Matrix’s engineers “have repurposed some of Corsair’s SRAM circuits to carry out vector-matrix multiplications,” SiliconANGLE reported. According to The Register, each Corsair card contains 2GB of SRAM good for 150 TB/s alongside 256GB of LPDDR5 capable of 400 GB/s — a hierarchy that “pairs a rather large amount of blisteringly fast SRAM with much, much slower, but higher capacity LPDDR5 memory” rather than relying on HBM. The same outlet reported the card delivers 2.4 petaFLOPS on the MXINT8 data type and 9.6 petaFLOPS on the lower-precision MXINT4 format. SiliconANGLE put the MXINT4 figure at “9,600 trillion calculations per second” and reported that a single SquadRack enclosure can run AI models with up to 100 billion parameters entirely in SRAM.
The company is well capitalized. d-Matrix raised $275 million in a Series C round that valued it at $2 billion, led by Bullhound Capital, Triatomic Capital, and Singapore’s Temasek sovereign wealth fund, SiliconANGLE reported. Microsoft participates in the company’s investor base through its M12 venture arm, with the Qatar Investment Authority and Singapore’s Economic Development Board Investments joining as new participants, according to Yahoo Finance. The company says the platform delivers tenfold better performance, triple the cost efficiency, and up to five times superior energy efficiency compared to GPU-based systems, the same report noted.
What We Don’t Know
d-Matrix has not published a named list of Corsair customers beyond the hyperscaler, neocloud, and frontier-lab categories cited in its announcement. The performance comparisons against Nvidia GPUs are the company’s own claims or come from a single third-party test by Gimlet Labs; broad, standardized benchmarks across a range of models had not been published at the time of the announcement. It also remains to be seen how the decode-only positioning holds up as GPU vendors and other inference specialists ship competing parts.
Analysis
d-Matrix’s pitch is less about beating the GPU outright than about dividing the inference workload and claiming the decode half that general-purpose accelerators handle least efficiently. By keeping the hot working set in on-die SRAM and falling back to LPDDR5 for capacity, the company sidesteps the high-bandwidth memory that dominates competing AI silicon — a notable design choice in a year when memory has become one of the industry’s tightest supply constraints. Whether the bet pays off will depend on the volume shipments d-Matrix has promised landing on schedule, and on how decode acceleration performs against a broadening field of inference-focused chips.