Raja Koduri's Oxmiq raised $35 million to license a GPU core that runs on an FPGA today

Raja Koduri, the GPU architect who ran graphics at AMD and then at Intel, closed a $35 million Series A for his startup Oxmiq Labs on July 1. That takes total funding to $60 million. The money goes toward OxCore, a licensable GPU core Oxmiq wants chipmakers and cloud builders to license so they can design their own AI silicon without paying for a full chip program.

Start with the part the press release puts near the bottom: OxCore runs on an FPGA today. An FPGA prototype proves the logic works. It is not a taped-out chip, and it tells you nothing yet about clock speed, power draw, or yield in real silicon. Oxmiq is selling architecture, not parts. There is no product to benchmark, and the company did not publish one.

What OxCore actually is

OxCore folds three things that normally sit on separate chips into one licensable core: a GPU engine Oxmiq calls CUDA-compatible, a tensor engine for the matrix math that dominates AI workloads, and a CPU that Oxmiq calls an orchestration engine for scheduling work across the system. The design point is near-memory compute, keeping the math close to the memory so the system moves less data. Data movement is where a large share of the energy in AI inference goes, so cutting it is a real target, not marketing. Whether OxCore's version of it holds up is a silicon question no one can answer from a press release.

The claim to interrogate is "CUDA-compatible." Oxmiq says a layer called OxPython runs existing CUDA and PyTorch code on OxCore without code changes, and that it has shown this on third-party hardware in live demos. Nvidia's real moat is CUDA, and a long line of companies has promised to run CUDA code somewhere else. Compatibility is easy to demo and hard to ship at full speed. Until there is an independent benchmark that says what the translation costs in performance, treat it as a demo.

A second piece, OxQuilt, is Oxmiq's chiplet packaging layer. It is meant to let a customer mix compute chiplets and memory types across foundries instead of locking a design to one fab and one memory standard. That flexibility is the selling point for a licensing business, because it widens the set of customers who can use the IP.

The business model is the actual pitch

Oxmiq is not trying to build and sell a GPU. It licenses the design and lets the customer build the chip, the same shape of business as Arm or Imagination. That keeps Oxmiq's costs down: it books revenue from customer engagements while it funds the rest of the stack, instead of spending nine figures to tape out an SoC it has to sell.

The catch is that a licensing company is worth what its license book is worth, and Oxmiq has not named a customer. It says it is working with semiconductor companies, neoclouds, and robotics firms "ready to own their compute roadmap." That is a pipeline, not a signed deal. The number that will matter in a year is the first named licensee, not the $35 million.

The people and the money

Jim Keller joins the board. Keller's résumé is the reason that line is in the release: he worked on Apple's A4 and A5, led the AMD Zen architecture, ran Tesla's Autopilot silicon, and now runs Tenstorrent. His presence does not prove the product works, but it is a real signal about the team.

The round was co-led by Fundomo and the Samsung Catalyst Fund, with MediaTek reinvesting from the seed, plus Pegatron Venture Capital, CDIB-TEN, Darwin Ventures, Morgan Creek Digital, and AM Intelligence Labs. Intel Capital is in as a strategic IP partner, which is its own footnote: Koduri left Intel in 2023, and Intel's investment arm is now backing his company. The strategic names carry more weight than the dollar figure here. Samsung and MediaTek can fabricate and ship, and for an IP business, money from investors who own the supply chain is worth more than money that does not.

What to watch

Three things tell you whether this is more than a well-funded architecture deck: a first named licensee, a benchmark with a stated methodology, and OxCore in real silicon rather than on an FPGA. Koduri has built credible GPUs before, the team is real, and $60 million buys time to prove it. None of that is the same as a working chip in a customer's hands, which is the only test that counts.