d-Matrix pairs Corsair chip with Nvidia Blackwell in hybrid rack

The chip story worth reading twice this week is not another Nvidia flagship, it is that a Microsoft-backed startup slid its inference accelerator into a rack next to Nvidia's Blackwell rather than trying to replace it. d-Matrix said on June 9 that its Corsair inference accelerator entered full production and began shipping in volume to hyperscalers, neoclouds, and frontier AI labs, per reporting from The Information and CNBC.

The specific claim d-Matrix is making is that when Corsair is paired with an Nvidia Blackwell GPU, the combined rack runs inference roughly 10 times faster, three times cheaper, and up to five times more energy efficiently than a standalone GPU. Take those multiples as reported, not settled: d-Matrix is citing research from Gimlet Labs rather than independent hyperscaler telemetry, and it is the company's own framing. The architectural bet underneath is more interesting than the marketing numbers. Instead of shuttling activations in and out of HBM, Corsair keeps compute on-die in SRAM, and in disaggregated serving the GPU handles the compute-heavy prefill while Corsair takes the decode phase.

Why this matters if you are not buying chips yourself: inference is on track to be two-thirds of all accelerator spending, and Nvidia still holds roughly 80% of the AI accelerator market. A cooperative pairing, rather than a challenger substitution, gives hyperscalers a credible second stack that runs alongside Blackwell rather than displacing it, which is a very different negotiating posture than the Nvidia-or-nothing one they have been in. d-Matrix has raised about $500 million at roughly a $2 billion valuation, with Microsoft in via its M12 venture arm, so the balance-sheet backing to actually ship at scale is there.

The honest caveat is that the eye-catching multiples came from a single cited study rather than customer benchmarks, and the SRAM-heavy design concentrates supply risk on TSMC's N6 process and Alchip on manufacturing. What the reporting does not give you is what Nvidia formally commits to on its side, whether there is joint engineering or just interoperability, and how Corsair holds up on frontier models whose state does not fit neatly on the card.

The forward-looking read is that the integrators quietly win here. Supermicro, Arista, and Broadcom are packaged into d-Matrix's SquadRack rack-scale reference design, with configurations available through Supermicro from Q1 2026, and any AI infra buyer who was hoping for a second credible inference vendor to walk into procurement with just got one.