The startup, founded by former Intel and AMD graphics chief Raja Koduri, is shifting from licensing GPU chip designs to building full AI infrastructure, betting that any challenger to Nvidia's grip on AI will live or die on the orchestration layer above the chip, not the silicon…
The most interesting thing about Oxmiq Labs' $35 million Series A is not the size of the round, or even the founder on the cap table. It is what the company is now actually selling. Until recently, Oxmiq was a GPU intellectual property (IP) licensor, the kind of business that sells chip designs to other semiconductor companies. With the new round, the Campbell, California-based startup is repositioning itself as a designer of full AI data-center stacks, with energy generation at one end and a finished "AI factory" reference design at the other. That is a different bet entirely, and one that says a lot about where the bottleneck is moving for anyone trying to build AI infrastructure outside of Nvidia.
Raja Koduri, who most recently served as Intel's senior vice president and general manager of Architecture, Graphics and Software and previously led AMD's Radeon Technologies Group, frames the shift in unusually blunt terms. In discussions of the new round reported by EE Times, he argues that Oxmiq is now designing for the full path "from electron to token," from raw power and silicon through to the AI inference or training output customers actually pay for. The "orchestration gap beyond CUDA" that Koduri invokes, meaning the missing software layer that sits above any non-Nvidia AI chip the way Nvidia's CUDA development platform sits above Nvidia hardware, is, in his telling, the financial feasibility problem for alternative AI factories.
For the past three years, the dominant story in AI infrastructure has been the chip. Training-grade GPUs from Nvidia were the chokepoint, and every alternative, from Tenstorrent and Cerebras to Groq and the custom silicon programs inside the hyperscalers (the very large cloud providers such as Microsoft Azure, Amazon Web Services, and Google Cloud), was framed as a hardware race to close the chip gap. That story is now running out of road, because the buyers of alternative chips face a different problem: they do not run on CUDA, and they do not have a turnkey way to schedule, route, and recover work across 10,000 or more accelerators.
Oxmiq's new product, OxFactory, is positioned as the answer to that gap. It is a data-center-scale orchestration platform that extends the earlier OxCapsule software, originally targeted at edge and consumer hardware at stealth, to coordinate pods of accelerators. Sitting underneath it, OxCore remains the GPU hardware IP core that other companies can license. The thesis is that the three layers, silicon, appliance, and orchestration, belong to one vendor.
The board additions announced with the round reinforce that thesis. Jim Keller, the veteran CPU and GPU architect who now runs Tenstorrent, a direct Nvidia rival building its own AI chips, and Valluri "Bob" Rao, a long-time Intel data-center executive, both joined the board as part of the Series A financing announced by BusinessWire. The Keller seat in particular is a tell: Tenstorrent is simultaneously a potential customer for Oxmiq's orchestration software, a potential licensee of OxCore IP, and a direct competitor building its own silicon. That the CEO of one of the most credible non-Nvidia AI chip companies is willing to take a board seat at an orchestration startup is itself a signal that insiders now expect value to accrue in the software layer above the chip.
The competitive context is crowded. The "AI factory" framing has been adopted across the industry, from Nvidia's own marketing to CoreWeave's positioning as a GPU cloud provider to the various Stargate-scale projects (multi-billion-dollar AI infrastructure initiatives announced in the US and elsewhere) that treat clusters of accelerators as production lines rather than servers. Oxmiq's differentiation is the "non-Nvidia" part of that market, and the claim that its orchestration software is accelerator-agnostic, working across OxCore IP and third-party silicon. The risk is execution. There are no public customers, no disclosed design wins, and no revenue yet. The local coverage of Oxmiq's stealth exit confirms the company has been operating quietly for roughly two years, but does not resolve whether the orchestration thesis will translate into a paying customer base.
What to watch is not inside Oxmiq. It is whether any of the announced multi-billion-dollar AI factory projects outside Nvidia, whether sovereign (state-funded national AI infrastructure), hyperscaler, or independent cloud, end up selecting orchestration software that is not anchored to CUDA. If they do, the post-Nvidia orchestration thesis becomes more than a Series A story, and the company that owns the orchestration layer for alternative silicon becomes, in effect, the platform vendor for the second-tier AI factory market. For any institution evaluating vendor risk on AI infrastructure exposure, the question is no longer which chip but which software layer sits above whichever chip we pick, and that layer is, for the first time, the place where non-Nvidia builders are now competing.