By ByteBot
Raja Koduri has designed GPUs for AMD, Apple, and Intel — including Ponte Vecchio, Intel’s first peta-scale GPU assembled from 47 chiplets. He knows better than most what it costs to build serious silicon from scratch. So when Koduri’s new startup, Oxmiq, raised $35 million on July 1, 2026, to scale its OxCore architecture, the approach was deliberately not to build another chip.
Oxmiq licenses GPU architecture. OxCore — a unified, CUDA-compatible IP block combining a GPU engine, a tensor processing engine, and a CPU orchestration layer — is designed to be embedded in other companies’ silicon. Think ARM, not AMD.
What OxCore Actually Is
Most chips handling AI workloads run three conceptually separate compute engines: a CPU for scheduling and orchestration, a GPU for parallel compute, and a tensor unit for matrix math. OxCore fuses all three into a single licensable IP block. Semiconductor companies license it, customize it for their target workload, and tape out their own chips. Oxmiq collects royalties on every chip produced.
The architecture was described in the Series A announcement as “purpose-built for near-memory compute, minimizing data movement to enhance compute and energy efficiency.” That’s a direct shot at the bandwidth bottleneck that plagues today’s GPU-to-HBM pipelines. OxCore is running on FPGA today with live demonstrations available; production silicon depends on licensees taping out their own chips.
The Series A brings Oxmiq’s total funding to $60 million. Investors are not passive — the round was co-led by Samsung Catalyst Fund, with MediaTek and AM Intelligence Labs as strategic participants. Samsung spent $73 billion on semiconductor R&D and CapEx in 2026. MediaTek ships hundreds of millions of chips a year. These are companies that want to build custom AI silicon and need GPU IP to do it. Their participation in Oxmiq’s round is a statement of intent, not just a financial bet.
OxPython: The Part That Matters for Developers
Hardware licensing is an infrastructure story. OxPython is the developer story.
The practical obstacle for every non-NVIDIA AI chip isn’t raw compute — it’s that production ML code is written against CUDA. PyTorch assumes torch.cuda. HuggingFace pipelines call NVIDIA-specific libraries. Rewriting that code to run on different hardware is a multi-month engineering project, which is why most teams never bother evaluating alternatives in the first place.
OxPython addresses this directly: existing CUDA and PyTorch code runs on OxCore-powered hardware without code changes. According to Tom’s Hardware, Tenstorrent hardware is the first OxPython target. If that compatibility holds under production workloads, switching from NVIDIA becomes a procurement conversation instead of a re-engineering project.
The CUDA Moat Is Real — OxCore Routes Around It
CUDA has 4 million registered developers and 40,000+ organizations running CUDA-dependent applications. The real lock-in isn’t the GPU — it’s the decade of optimizations, library assumptions, and kernel fusions baked into production code. AMD’s ROCm has matured; OpenAI’s Triton makes AMD and Intel more accessible. But for large-scale training and inference, the switching cost is still measured in engineering months.
Oxmiq is not trying to out-CUDA CUDA. OxCore targets the companies building chips, not the engineers writing model training code. OxPython sits above the hardware and below the framework — a compatibility shim that makes the CUDA software stack portable. It’s a fundamentally different attack vector than every other company that has tried to dethrone NVIDIA by shipping a competing chip.
That said, FPGA demonstrations and production-scale cuDNN kernel performance are different things. The CUDA ecosystem’s depth — cuDNN, cuBLAS, NCCL, mixed-precision behaviors tuned to NVIDIA’s math libraries — will test OxPython’s compatibility layer in ways a demo cannot. The promise is credible; the proof requires real chips.
What to Watch
Oxmiq’s validation moment is the first OxCore-based chip entering production from a licensee. If Samsung or MediaTek announce a product built on OxCore architecture — whether for edge inference, mobile AI, or datacenter compute — that is the ARM-for-AI-GPUs story moving from theory to hardware. SiliconANGLE noted this round dramatically compresses the design cost for companies that have historically avoided custom silicon.
The broader AI chip market is clearly splintering. NVIDIA dominates training and frontier inference. Cerebras owns throughput benchmarks. Tenstorrent is the open-architecture bet. Oxmiq is playing a different layer entirely — the architecture substrate — with a business model that doesn’t require beating NVIDIA to win. The ARM analogy is apt. ARM didn’t beat Intel by building a faster desktop CPU. It built a licensing framework that quietly ran the world from the inside.
OxCore is early. But so was ARM in 1993 when Texas Instruments first licensed it.
