D-Matrix Corsair Enters Production, Hits 10x Inference Speed
Key insights
- Gimlet Labs independent testing showed Corsair paired with GPUs cut inference response time from 24 seconds to under two seconds.
- SquadRack combines Corsair accelerators, JetStream networking, and Aviator software in a rack-scale unit that requires no liquid cooling.
- Corsair uses TSMC N6 SRAM-based chiplets with LP-DDR5 memory, bypassing HBM supply chain risks that affect competing inference hardware.
Why this matters
Heterogeneous disaggregated inference, splitting prefill onto GPUs and decode onto dedicated accelerators, now has a production-validated architecture in SquadRack, giving hyperscalers a concrete alternative to GPU-only clusters at a moment when agentic AI workloads are making decode latency a first-order concern. The Gimlet Labs result of reducing response time from 24 seconds to under two seconds sets a public benchmark that GPU-only inference vendors will have to respond to on latency grounds. For infrastructure teams building real-time AI products, Corsair's full production status means disaggregated inference is a procurement option today, not a futures bet.
Summary
D-Matrix's Corsair inference accelerator platform entered full production on June 9, 2026, with volume shipments going to priority hyperscalers, neoclouds, and frontier AI labs.
Corsair handles the decode phase of inference while GPUs take prefill tasks. Independent testing by Gimlet Labs showed pairing the two cut response time from 24 seconds to under two seconds, roughly a 10x speedup over GPU-only approaches. Demand is being driven by agentic workloads including Claude Code and OpenClaw, both launched in late 2025.
Essentially: (d-Matrix, Gimlet Labs) put a production-validated heterogeneous inference stack on the market, not a prototype.
- SquadRack packages Corsair with JetStream networking and Aviator software in a rack-scale unit requiring no liquid cooling and deploying within days.
- Corsair is manufactured at TSMC N6 via Alchip Technologies, using SRAM-based chiplets with LP-DDR5 memory, avoiding HBM supply dependencies.
- d-Matrix acquired GigaIO's data center business in April to deepen rack-scale integration expertise.
With SquadRack now shipping to production customers, the disaggregated inference architecture moves from benchmark claim to deployable product.
Potential risks and opportunities
Risks
- GPU vendors could accelerate decode-optimized updates to next-generation products, narrowing Corsair's performance differentiation before d-Matrix achieves volume scale across hyperscaler customers.
- If priority hyperscaler customers redirect spending toward in-house silicon programs, d-Matrix's summer 2026 production ramp could outpace near-term demand and create inventory risk.
- Integration of GigaIO's data center business acquired in April adds software complexity to SquadRack; delays unifying that stack with Corsair could undermine the 'deploys within days' claim central to the platform's pitch.
Opportunities
- Alchip Technologies, co-manufacturer of Corsair at TSMC N6, is positioned for volume contract expansion as d-Matrix scales from priority customers to broader hyperscaler deployments.
- Neoclouds that adopt SquadRack early gain a differentiated low-latency inference offering for agentic workloads, ahead of competitors still relying on GPU-only infrastructure.
- Enterprise AI teams building real-time voice and agentic applications now have a production-validated alternative to GPU-only decode infrastructure through SquadRack as a concrete reference architecture.
What we don't know yet
- Which specific hyperscalers or neoclouds are among the 'priority customers' receiving initial volume shipments is not disclosed in the announcement.
- Pricing and total cost of ownership for SquadRack versus GPU-only racks is absent from the announcement.
- How Corsair's SRAM-based architecture performs on workloads beyond the agentic and real-time voice use cases highlighted is not addressed.
Originally reported by d-matrix.ai
Read the original article →Original headline: D-Matrix Corsair AI Inference Accelerator Enters Full Production — Claims 10× Speed and 5× Energy Efficiency Over GPU-Only Inference, Targets Hyperscalers and Frontier AI Labs