d-Matrix Says Corsair AI Inference Platform Has Entered Full Production

d-Matrix said its Corsair AI inference accelerator platform has entered full production, marking a manufacturing milestone for the company’s push into low-latency AI serving infrastructure.

Corsair moves into production

In an announcement published by d-Matrix and distributed through PR Newswire, the company said Corsair is now in full production and is aimed at rack-scale inference deployments for hyperscalers, neocloud providers, and frontier AI labs. The company described Corsair as an inference platform designed for high-throughput, low-latency generative AI workloads.

The PR Newswire release said d-Matrix plans volume shipments for priority customers, and included partner comments from TSMC and Alchip about the production ramp. Those comments indicate that d-Matrix is presenting the milestone not only as a product update, but as evidence that its manufacturing and supply-chain partners are supporting broader availability.

The company’s positioning reflects a broader industry focus on inference, the stage where trained AI models generate outputs for users. As AI services move from development into large-scale deployment, chipmakers and infrastructure companies are competing to reduce latency, power use, and cost per request. d-Matrix is one of several firms targeting this part of the market with specialized hardware rather than general-purpose graphics processors alone.

A focus on low-latency AI serving

According to d-Matrix, Corsair is built for low-latency inference at rack scale. The company has emphasized use cases where responsiveness matters, including interactive generative AI applications. The announcement did not provide independently audited performance figures, so claims about production status and target markets should be read as company statements.

A separate technical post from Gimlet Labs described testing that combined d-Matrix Corsair with GPU-based speculative decoding. Gimlet Labs reported 2–5x end-to-end request speedups for configurations optimized around interactivity, and up to 10x speedups for energy-optimized configurations. Those results come from Gimlet Labs’ own testing and should not be generalized to all models, workloads, or deployments without additional benchmarking.

Speculative decoding is a technique used to accelerate model output by pairing a faster draft process with verification by a larger model. In practical deployments, the benefit depends on factors such as model architecture, batch size, interconnect, memory behavior, and service-level latency targets. Gimlet Labs’ post is notable because it places Corsair in a real inference-serving workflow rather than only describing the chip in isolation.

Why the milestone matters

The production announcement matters because inference hardware must prove not only performance, but availability, integration, and reliability. Customers deploying AI services at scale generally need predictable supply, software support, and the ability to fit new accelerators into existing data center operations.

The d-Matrix and PR Newswire announcements frame Corsair as ready for wider customer deployment, while the Gimlet Labs post provides an early example of how the platform may be used in latency-sensitive serving setups. However, the available sources do not disclose customer names, shipment volumes, pricing, or broad third-party benchmark comparisons against competing accelerators.

For now, the most supportable conclusion is that d-Matrix has moved Corsair from development into full production and is preparing volume shipments for selected customers. The next test will be whether customers and independent evaluators report consistent gains in production AI services across a range of models and workloads.