Nvidia Prepares to Back a Rival: Report Says a d-Matrix Investment Is in the Works, With the Startup's Next Inference Chip Already Slated for Nvidia's Own Racks

Nvidia has spent the year buying, backing and absorbing the companies building alternatives to its own chips, and The Information reported on Oct 8 that the world's most valuable chipmaker is preparing to do it again: an investment in d-Matrix, a startup whose inference accelerators compete directly with Nvidia's GPUs. The report, citing three people familiar with the matter, did not disclose financial terms, and neither company has announced a deal.

d-Matrix is not a typical Nvidia target. Founded in 2019 by Sid Sheth, a former Marvell executive, the company builds digital in-memory compute silicon designed specifically for inference rather than training — a bet that as models move into production, the bottleneck shifts from raw FLOPS to latency, energy use and cost per token. Its investors include Microsoft's venture arm M12 and Temasek; the company has raised roughly $500 million to date, including a $275 million Series C completed in November 2025 at a valuation of about $2 billion.

What makes the reported investment unusual is how much of the relationship is already public and official. Over the summer, Nvidia announced a collaboration to let its GPUs and d-Matrix's XPUs divide a single inference workload between them, and in September the two companies went further: d-Matrix's next-generation Raptor XPU will natively support Nvidia's NVLink Fusion interconnect and plug into the MGX rack reference architecture, with the first systems expected in the fourth quarter of 2027. Connectivity supplier Astera Labs is part of the same push.

"NVLink Fusion lets partners integrate their custom silicon with Nvidia's NVLink ecosystem, advanced packaging, rack-scale systems and networking," Nvidia CEO Jensen Huang said when the Raptor integration was announced. "As Nvidia AI infrastructure deploys in clouds and local data centers around the world, NVLink Fusion gives partners like d-Matrix a path to integrate seamlessly with our computing platform." Sheth framed it the same way from the other side: customers will be able to run his inference XPUs side by side with the Nvidia AI factory platform. Astera Labs chief executive Jitendra Mohan said the partnership delivers high-throughput connectivity for low-latency inference inside the same ecosystem.

The strategy is easy to read. Rather than treat every specialized inference chip as a threat, Nvidia is wiring them into its own interconnect, rack standard and supply chain — a playbook it has already applied to Groq, which Nvidia absorbed earlier this year in a transaction reported at around $20 billion. An equity stake in d-Matrix would extend that approach from acquisitions into minority investments: the startup's silicon still ships, but inside Nvidia's rails.

For d-Matrix, the tie-up answers the hardest question facing any Nvidia competitor: distribution. Inference is the workload every AI operator is trying to make cheaper, and purpose-built chips are credible challengers on latency and energy — but customers buy systems, not chips, and the MGX rack ecosystem is where those systems get designed. The caveat is just as simple: a reported investment does not mean Nvidia will adopt d-Matrix's technology at scale. It means Nvidia will be inside the room when the next generation of inference silicon gets defined.