CUDA lock-in has been the most expensive tax on AI infrastructure for nearly two decades. Once your training scripts are written in CUDA, you are writing them for NVIDIA hardware — period. Switching to any alternative means rewriting everything, revalidating on new silicon, and accepting months of lost time. That is the wall every NVIDIA challenger has run into: the hardware gets built, but the software moat holds. Raja Koduri, who architected GPU generations at both AMD and Intel, thinks he has found the door through that wall. His startup Oxmiq just raised $35 million to prove it.
OxPython: Zero Rewrites, Same Code #
The headline from Oxmiq’s $35M Series A is not the money — it is OxPython. This is a software portability layer that runs existing Python-based CUDA and PyTorch code on non-NVIDIA hardware without any modification. Not a recompile. Not a library swap. Not a migration guide. The same scripts you have running on your H100s today, Oxmiq claims, should run on OxCore-powered silicon unchanged.
That claim would be easy to dismiss as vaporware — except Oxmiq has already demonstrated OxPython on Tenstorrent’s Wormhole and Blackhole AI accelerators, which are production third-party hardware. This is not a lab demo on their own chips. It is a live validation on someone else’s silicon. Jim Keller, CEO of Tenstorrent and the architect behind AMD Zen, the Apple A4, and Tesla’s FSD chip, has joined Oxmiq’s board. That is not a PR move. Keller has been saying the quiet part loud for years: “CUDA’s a swamp, not a moat. x86 was a swamp too.”
What OxCore Actually Is #
OxCore is Oxmiq’s licensable GPU IP — built on RISC-V and designed to be embedded in third-party silicon, not sold as a finished chip. The architecture tightly couples three compute engines into one block: a CUDA-compatible GPU engine for parallel workloads, a tensor processing engine for matrix math, and a RISC-V-based orchestration engine that coordinates everything. The near-memory compute design reduces data movement, which is where most AI workload energy and cost actually goes.
Above that sits OxQuilt, a chiplet architecture that scales the same OxCore design from edge devices all the way to data center configurations. According to the HPCwire announcement, OxCore licensing is available now. Semiconductor companies and neoclouds can license the IP, build their own silicon around it, and ship hardware that OxPython already knows how to run.
The Arm Parallel #
Koduri’s framing of the business model is deliberately clear: “We would want to be the Arm of this next era.” Arm does not sell phones. It licenses CPU designs, and every major phone chip company builds on top of those designs. Oxmiq does not want to compete with NVIDIA on the shelf. It wants to be the underlying architecture that makes a dozen other chips CUDA-compatible.
The investors signal that the strategy is credible. Samsung Catalyst Fund, MediaTek, and Intel Capital all participated in the Series A. These are not financial bets — these are strategic semiconductor players who have skin in the game of building NVIDIA-alternative silicon. MediaTek was also in Oxmiq’s seed round. When the companies that actually fabricate chips are writing checks, the licensing model has more than theoretical legs.
Where This Stands Today #
Be honest about the timeline. OxPython runs on Tenstorrent hardware today. OxCore silicon is not yet in production. The value chain here is: Oxmiq licenses IP → chipmakers build silicon → products reach developers. That is an 18-to-24-month runway before most developers encounter an OxCore-powered chip in the wild. Jon Peddie Research notes that the Arm model took years to gain traction before it became inevitable — and then it upended everything.
The comparison to prior CUDA challengers is instructive. AMD ROCm is legitimate but still requires container swaps and some porting work. Intel’s oneAPI is standards-based but C++-heavy and creates real friction for Python-first AI teams. OpenAI’s Triton is powerful but requires rewriting kernels in a different language. OxPython’s bet — that Python-level compatibility is sufficient to unlock adoption — is a different wager than any of them. Tom’s Hardware covers the technical architecture in depth if you want to go deeper.
What is worth watching: which neoclouds license OxCore first, whether OxPython performance benchmarks hold up against real workloads, and what the Qualcomm-Tenstorrent deal (reportedly up to $10 billion) means for the board member relationship Keller has just formalized. The ecosystem around OxCore is assembling fast. The CUDA moat has looked invincible before — and so did x86.