The technical bet here is squarely on software compatibility, not just raw silicon performance. Reuters quotes Koduri wanting OXMIQ to "be the Arm of this next era", licensing GPU IP the way Arm licenses CPU designs rather than selling chips directly. For ML infrastructure engineers evaluating non-Nvidia hardware, the practical blocker has rarely been raw compute; it has been that most of the PyTorch/HuggingFace ecosystem assumes torch.cuda and Nvidia-specific tooling. OXMIQ's OxPython layer is explicitly built to let existing CUDA and PyTorch code run unmodified on other hardware (Tenstorrent first, per earlier EE Times reporting), which is the harder and more consequential problem to solve than the hardware IP itself.
What happened
Reuters reported that OXMIQ Labs, a Campbell, California startup founded by Raja Koduri (former Intel chief GPU architect and ex-AMD graphics executive), closed a $35 million Series A on July 1, 2026, bringing total funding to $60 million. The round was co-led by Samsung Catalyst Fund and Fundomo, with MediaTek and Pegatron Venture Capital among the participants, per Reuters and the company's own announcement. Koduri told Reuters the funding will complete the first batch of OXMIQ's licensable IP and expand engineering headcount; the company also said it plans to enter the custom chip market where Broadcom, Marvell, and MediaTek already compete.
Technical context
According to earlier EE Times reporting on OXMIQ's 2025 stealth launch (when it raised a $20 million seed round), OxCore, the company's licensable GPU hardware IP, combines a CUDA-compatible GPU engine, a tensor processing engine, and an orchestration engine in one design, using RISC-V cores and near-memory compute concepts while still conforming to standard SIMT GPU programming models. OxQuilt is OXMIQ's chiplet-configuration toolset, letting customers assemble compute, memory, and interconnect chiplets already in OXMIQ's library, which Koduri said is "20 to 100 times less expensive" than designing and taping out custom silicon. The company says OxCore is running on FPGA today, ahead of silicon availability.
Market context
OXMIQ's premise, that IP-first, chiplet-friendly licensing can shorten the multi-year, multi-hundred-million-dollar cycle of building a custom AI chip, mirrors a broader industry pattern of foundry-adjacent and strategic corporate investors backing IP alternatives to full custom silicon programs. The investor mix here, Samsung's chip-investment arm alongside MediaTek, a fabless SoC vendor, fits that pattern.
What to watch
Whether OxCore ships silicon-class (not just FPGA) performance benchmarks, and whether any named design wins emerge from MediaTek or Pegatron; adoption of OxPython by developers running CUDA/PyTorch workloads on non-Nvidia hardware, the key test of whether the software-compatibility bet works; and whether OXMIQ's planned move into the custom chip market puts it in direct competition with Broadcom, Marvell, and MediaTek rather than purely licensing IP to them.
Key Points
- 1OXMIQ's core bet is software compatibility (CUDA/PyTorch code running unmodified via OxPython), the harder problem that has blocked prior Nvidia GPU challengers.
- 2Founder Raja Koduri (ex-Intel chief GPU architect) wants OXMIQ to license GPU IP the way Arm licenses CPU designs, rather than sell chips directly.
- 3OxCore uses RISC-V cores and near-memory compute and is running on FPGA today; OxQuilt's chiplet approach claims 20-100x lower R&D cost than custom silicon tapeout.
Scoring Rationale
A credible, well-attributed funding story ($35M Series A, $60M total) from a serious founder (ex-Intel chief GPU architect) with named quotes and technical detail confirmed across Reuters and EE Times reporting. Notable for ML infrastructure practitioners evaluating non-Nvidia hardware, though the software-compatibility claims remain unproven at silicon scale.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
