Can OXMIQ’s OxPython turn CUDA against Nvidia?
The artificial intelligence industry has spent billions of dollars trying to build alternatives to Nvidia’s processors. Breaking the company’s hold on the market has proved much harder than designing another chip.
The reason is CUDA.
Introduced by Nvidia in 2006, CUDA is a computing platform that allows developers to use graphics processing units, or GPUs, for work beyond rendering images. Because GPUs can perform large numbers of calculations simultaneously, they have become essential to training and running artificial intelligence models.
Over two decades, Nvidia has built an extensive ecosystem around CUDA: programming tools, optimised libraries, compilers, debugging software and integrations with frameworks such as PyTorch. Researchers, universities and technology companies have simultaneously accumulated applications, expertise and working practices around it.
CUDA has consequently become one of Nvidia’s most important competitive advantages. It is not simply software supplied with its processors. It is an ecosystem that helps determine which processors customers buy.
The scale of Nvidia’s position is visible in its financial results. Their revenue reached $215.9 billion in its 2026 financial year, up 65% from the previous year, as demand for AI infrastructure continued to expand.
A competing accelerator may be cheaper or perform well on a particular workload. Moving to it can nevertheless require software to be adapted, dependencies replaced, engineers retrained and production systems tested again. Any hardware saving can quickly be outweighed by the cost and risk of leaving CUDA.
OXMIQ Labs believes there is another way: keep the CUDA-based software, but allow it to run on somebody else’s processor.
Nvidia’s deeper advantage
CUDA has created a self-reinforcing commercial cycle.
Developers use it because Nvidia processors are widely deployed. Organisations buy those processors because their developers and applications already use CUDA. That installed base encourages more companies to create CUDA-compatible tools and libraries, making the ecosystem still more useful — and more difficult to leave.
Nvidia’s success reflects years of investment and strong execution. The company recognised the potential of general-purpose GPU computing well before the current generative AI boom and built a complete platform rather than selling an isolated component.
The resulting concentration nevertheless presents companies and governments with strategic questions. Customers have limited alternatives if Nvidia processors are expensive or in short supply. Semiconductor companies face high barriers when introducing competing designs. Governments may finance domestic data centres and sovereign AI programmes, yet still depend on one American supplier for the underlying compute architecture and software environment.
The problem is therefore not simply whether another company can build a fast AI chip. It is whether customers can use it without abandoning the software on which their AI operations already depend.
That is the problem California-based OXMIQ is trying to solve.
How OxPython could separate CUDA from Nvidia hardware
OXMIQ is developing OxPython, a compatibility and execution layer that the company says allows AI software written for Nvidia’s CUDA environment to run on other companies’ processors.
To an application, OxPython presents the Nvidia interface it expects. Underneath, the software redirects the computational work to the processor that is actually present.
OXMIQ has demonstrated selected applications running through OxPython on processors developed by Tenstorrent, an AI chip company led by veteran architect Jim Keller. The examples include text generation with Meta’s Llama 3.2, mathematical reasoning using DeepSeek R1 Distill, EfficientNet image classification and video generation with CogVideoX.
The company says applications can run without being ported or rewritten and that OxPython supports commonly used inference tools including PyTorch, vLLM and Hugging Face.
If that capability works broadly and efficiently, customers would no longer have to choose between retaining CUDA compatibility and adopting a non-Nvidia processor. Nvidia’s software ecosystem could become an entry point to rival hardware.
That makes OxPython a potentially existential threat to Nvidia’s CUDA lock-in — although not yet to Nvidia itself.
The anti-CUDA alliance that still needs CUDA
OXMIQ emerged from stealth in August 2025 and is led by founder and chief executive Raja Koduri, a prominent GPU architect whose career has included senior roles at ATI, AMD, Apple and Intel.
Raja Koduri, Founder & CEO of OXMIQ (photo: OXMIQ)
On 1 July, the company announced a $35 million Series A financing round, bringing its total funding to $60 million. Fundomo and Samsung Catalyst Fund co-led the round, with MediaTek, AM Intelligence Labs, Pegatron Venture Capital, CDIB-TEN, Darwin Ventures and Morgan Creek Digital also participating.
There is no formally constituted anti-CUDA alliance, and OXMIQ has not said that these strategic investors will become customers. What is emerging is a group of companies with a shared interest in making AI compute less dependent on Nvidia hardware.
Samsung operates across memory, chip manufacturing and AI systems. MediaTek develops processors for smartphones, edge devices and other computing platforms. Pegatron manufactures electronics and computing systems. AM Intelligence Labs is planning large-scale renewable-powered AI infrastructure in India.
Keller has also joined OXMIQ’s board, strengthening its connection to Tenstorrent, the publicly identified hardware platform on which it has demonstrated OxPython.
The result resembles the beginnings of an anti-CUDA alliance, with an important paradox: its participants do not need developers to abandon CUDA. They stand to benefit if the investment already made in CUDA can be preserved while Nvidia’s exclusive claim on the resulting hardware demand is weakened.
The more software the industry builds around CUDA, the more valuable OxPython could become — if OXMIQ can make that software genuinely portable.
From CUDA compatibility to custom AI chips
OXMIQ’s plans extend beyond running CUDA applications on existing alternative processors. The company is also developing a licensable architecture called OxCore.
OxCore brings together three types of computing function: highly parallel GPU-style processing, tensor processing for AI calculations and an orchestration engine that coordinates workloads across the system.
OXMIQ argues that integrating these functions and moving computation closer to memory can reduce the movement of data around a processor. That matters because transferring information repeatedly between processors and memory consumes energy and creates delays, particularly as AI inference expands.
OxCore is currently running on a field-programmable gate array, or FPGA. This allows the architecture to be tested and demonstrated before it is manufactured as commercial silicon.
OxPython is intended to become the native software route into OxCore. Rather than only translating an application designed for Nvidia hardware, the software would schedule and optimise its execution across the architecture.
OXMIQ is also developing OxQuilt, which would allow customers to combine compute, memory and other chiplets in configurable packages. Semiconductor companies and AI infrastructure operators could select manufacturing processes, memory types, packaging technologies and foundries without designing an entire processor from the ground up.
Koduri has said OXMIQ wants to become “the Arm of this next era”. Arm licenses processor architecture and intellectual property to companies that incorporate it into their own chips. OXMIQ similarly wants to give organisations a shorter, less expensive route to customised AI silicon.
What OXMIQ has yet to prove
The central technical question is not whether OXMIQ can make selected applications run in a controlled demonstration. It is whether it can progress through four increasingly demanding levels of proof: selected models running, broad software compatibility, competitive performance and production reliability at data-centre scale.
So far, it has publicly demonstrated the first.
OXMIQ has not announced a customer that has licensed OxCore for a production chip. It has not provided a timetable for a customer tape-out, when a completed chip design is sent for manufacturing. Nor has it published comprehensive independent comparisons covering performance, energy efficiency or compatibility across the wider CUDA ecosystem.
“Runs unchanged” is an ambitious claim when applied to a software environment developed over 20 years. CUDA includes numerous libraries, custom kernels, framework versions and hardware-specific optimisations. Supporting selected AI models does not establish that every production application will behave predictably on alternative silicon.
Nvidia also develops processors, networking, systems and software together. Even if OxPython reproduces CUDA compatibility, competing hardware must still demonstrate that it can match Nvidia on performance, energy use, availability, reliability and support.
OxPython is therefore a credible challenge rather than a proven escape route.
But OXMIQ does not have to defeat Nvidia to alter the economics of AI compute. It only has to prove that choosing CUDA software no longer guarantees the purchase of an Nvidia processor.
If it succeeds, customers could preserve years of software investment while regaining meaningful control over their hardware decisions. Nvidia might remain the market leader, but it would have to win more purchases through the performance of each new generation of technology rather than inheriting them from code written years earlier.
The anti-CUDA alliance’s most effective weapon may turn out to be CUDA itself.
Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.