OXMIQ, a startup founded by GPU architect Raja Koduri, just closed a $35 million Series A funding round to license a custom GPU design that could reshape how companies build AI inference chips. The company's pitch is straightforward but ambitious: provide a blueprint for AI silicon that doesn't require companies to become chip manufacturers, and make it compatible with existing NVIDIA CUDA code so developers don't have to rewrite their applications.

Koduri's track record lends credibility to the effort. He architected the GPUs behind PlayStation and Xbox at AMD, and later delivered Ponte Vecchio, the 47-chiplet GPU powering the Aurora exascale supercomputer at Intel. His new company is taking a fundamentally different approach than competitors like Groq, Cerebras, and Etched, which all manufacture their own hardware. OXMIQ is selling the design itself.

Why Does NVIDIA's Software Moat Matter More Than Its Hardware?

The real innovation here isn't the chip architecture; it's the software layer. OXMIQ's OxPython compatibility tool claims to run existing Python-based CUDA applications on OxCore-powered hardware without requiring any code changes. PyTorch models port natively, meaning no refactoring, no HIPify rewrites, and no months of validation testing.

This directly targets NVIDIA's actual competitive advantage. NVIDIA holds 86% of AI datacenter revenue not because its hardware is technically superior, but because developers are locked into 20 years of CUDA investment baked into every production machine learning stack. Switching to a competitor means rewriting code that teams have spent years optimizing and debugging. OxPython's zero-change claim, if it holds up in practice, removes the biggest reason not to switch.

CUDA's problems are well-documented in the developer community. Version mismatches between the CUDA toolkit, driver, and PyTorch cause build failures routinely. Cutting-edge operations like FlashAttention-3 require dropping below CUDA into PTX, NVIDIA's lower-level assembly language, which is poorly documented and shifts between hardware generations. OxPython is a direct attempt to solve that friction.

What Does the OxCore Architecture Actually Do?

OxCore is a RISC-V-based GPU intellectual property core that integrates three compute engines in one design: scalar, vector, and tensor operations. It adds an orchestration layer for managing AI agents and coordinating workloads. The whole system is built for near-memory compute, a design philosophy that keeps data close to the processor to cut the latency and energy cost of moving information around, which is where inference workloads typically bleed efficiency.

The architecture scales through OxQuilt, a chiplet-based system-on-chip builder. A single OxCore handles edge deployments on devices. Thousands of cores cluster together for datacenter inference workloads. The design is SIMT-compatible, meaning it follows the same programming model as CUDA and OpenCL, so developers writing SIMT code today should find it familiar.

How to Evaluate OXMIQ's Competitive Position

• Investor Backing: Samsung Catalyst Fund and Fudomo co-led the round, with MediaTek and Pegatron Venture Capital also participating. These are not speculative investors; Samsung manufactures chips, MediaTek designs them for mobile and edge devices, and Pegatron builds the final products. Together they cover the full custom silicon supply chain from fabrication to final product, suggesting OXMIQ's first licensees may already be in the room.
• Board Credibility: Jim Keller, who currently runs Tenstorrent, a direct OXMIQ competitor, joined the board. His career spans the AMD Zen revival, Apple's A4 and A5 chips, and Tesla's Full Self-Driving processor. His involvement signals the architecture has passed some level of technical scrutiny.
• Market Timing: Inference costs are the defining infrastructure problem of 2026. Custom silicon demand is accelerating across the industry. The structural conditions for disruption exist, even if OXMIQ still faces execution risks.

The skeptical take is worth considering. There is no finished silicon yet. The RISC-V foundation is open-source, which raises long-term competitive moat questions. NVIDIA took decades to build its position in a fragmented mobile market before a dominant incumbent existed. Today's AI chip market is not fragmented; NVIDIA is the incumbent. OXMIQ is making large claims with no public benchmarks to back them up yet.

The software stack includes OxCapsule for high-level orchestration and claims day-zero support for new model architectures. The company's total funding now stands at $60 million after this Series A round.

"OXMIQ is renting out chip design rather than building chips, which is either the smartest move in AI infrastructure or an idea waiting for a hard market reality check," noted industry observers.

Industry Analysis, The Next Web

The real test will come with OXMIQ's first licensee announcement. That milestone will reveal whether the Arm licensing model, which worked in mobile, can translate to the AI chip market. If a major manufacturer licenses OxCore and ships silicon within 18 months, OXMIQ's bet on compatibility-first design could reshape how companies approach custom AI inference hardware. If the company struggles to convert investor enthusiasm into actual silicon, it will join a long list of well-credentialed teams that couldn't execute at scale.