Nvidia teams up with chip rival d-Matrix instead of fighting it

Nvidia is partnering with chip startup d-Matrix to combine its GPUs with d-Matrix's Corsair inference accelerators in a joint system for AI workloads, with cloud firm Parasail as the first customer. The deal reflects Nvidia's strategy of collaborating with rivals to hedge against the rise of specialized inference chips, aiming to keep customers within its ecosystem while reducing costs and energy use for AI inference.

Nvidia has found a new way to handle its chip rivals: work with them. The GPU giant is combining its hardware with inference chips from the startup d-Matrix https://thenextweb.com/news/sambanova-11-billion-valuation-ai-chips , The Information https://www.theinformation.com/articles/nvidias-new-hedge-chip-competitors-partner reported. The two will ship a joint system to run AI models. The AI-cloud firm Parasail will be its first customer, with the system due online later this year. It fits a wider Nvidia habit of partnering with the very companies trying to unseat it. The logic follows the shape of AI work. Answering a prompt splits into two phases. A heavy “prefill” step suits Nvidia’s GPUs. A lighter, repetitive “decode” step runs more cheaply on specialised chips https://thenextweb.com/news/fractile-220m-inference-chip . Pairing the two lets each do what it does best. What d-Matrix brings d-Matrix built its Corsair accelerator around “in-memory computing”, which keeps data next to the logic that crunches it. Founded in 2019, it raised $275m at a $2bn valuation in November, and is raising again. CEO Sid Sheth says its edge is to put compute and memory on one chip. It also skips the pricey, scarce high-bandwidth memory that Nvidia’s chips lean on. The startup claims that pairing Corsair with GPUs runs tokens about 10 times faster https://www.prnewswire.com/news-releases/parasail-to-combine-nvidia-ai-infrastructure-with-d-matrix-accelerators-to-achieve-10x-faster-token-generation-302820178.html , at roughly a third of the cost and up to five times less energy on some jobs. Those are the company’s own figures, not independent results. But they point at the real prize in AI right now. Training built the models. Inference https://thenextweb.com/news/google-inference-chips-nvidia-challenge-supply-chain , running them for millions of users, is where the running costs, and the money, increasingly sit. Nvidia’s frenemy playbook The deal stands out because Nvidia rarely needs help. It dominates AI chips, and startups like d-Matrix exist to chip away at that lead. Rather than fight every one, Nvidia keeps folding rivals into its ecosystem https://thenextweb.com/news/nvidia-marvell-nvlink-fusion-ecosystem-lock-in . It has made the same move with other chip and networking firms. The approach hedges its bets. Say specialised inference chips take off, like the custom silicon https://thenextweb.com/news/amazon-custom-chips-jassy-letter-fifty-billion-trainium now spreading across big tech. Then Nvidia profits from the shift instead of losing to it. If they stall, it has lost little. Either way, customers stay inside Nvidia’s orbit. Why it matters The tie-up signals an AI-chip market maturing past a single winner. Buyers no longer want one giant GPU for everything. They want the cheapest chip for each job, stitched together. Nvidia’s answer is to make sure it sells, or sits beside, whatever wins. Get the TNW newsletter Get the most important tech news in your inbox each week.