Oxmiq Labs Raises $35M for Open GPU Architecture to Challenge Nvidia Dominance

Oxmiq Labs, a Campbell, California-based startup founded by former Intel, AMD and Apple chip architect Raja Koduri, announced on July 1, 2026 that it has raised $35 million in a Series A funding round. The company is developing a licensable GPU core called OxCore that it intends to license rather than manufacture itself.

The round was co-led by Fundomo and Samsung Catalyst Fund, with participation from MediaTek, AM Intelligence Labs, Pegatron Venture Capital, CDIB-TEN, Darwin Ventures, Morgan Creek Digital and Intel Capital. Oxmiq's approach could provide an alternative to Nvidia's dominant position in AI acceleration. Legendary chip architect Jim Keller, known for his work on AMD's Zen architecture, Apple's A4/A5 chips and Tesla's self-driving hardware, sits on Oxmiq's board and said: "This is more important than ever. OXMIQ's open, configurable foundation is exactly where compute should be heading."

What Happened: Oxmiq's $35M GPU Architecture Bet

Oxmiq Labs announced its $35 million Series A, bringing total funding to approximately $60 million. The company's strategy is to design a competitive GPU architecture and license it to data center and PC vendors, rather than manufacturing chips itself. Koduri framed the goal in terms of accessibility: "Bring that cost down, and you widen who gets to build with it. I believe AI is a force for good when it is a tool everyone can pick up and use."

Raja Koduri's involvement brings significant credibility. He has held senior chip-architecture leadership roles at Intel, AMD and Apple, giving him deep experience across GPU, graphics and custom-silicon design.

Key Details

OxCore is a licensable GPU core that integrates three compute engines in one design: a CUDA-compatible GPU engine, a tensor processing engine, and an orchestration engine that functions as a CPU. It is built for near-memory compute to minimize data movement and improve efficiency, and Oxmiq says the design scales from a single core up to full datacenter deployments. As of the funding announcement, OxCore runs on FPGA hardware rather than fabricated silicon, an important distinction for a company positioning itself as a Nvidia alternative. By licensing the architecture rather than building chips, Oxmiq aims to attract OEMs and system vendors who want differentiated products without the massive investment of developing their own GPUs from scratch. Alongside OxCore itself, Oxmiq is building OxPython, a compatibility layer designed to run existing CUDA and PyTorch code on OxCore without modification, directly targeting the software-portability problem that has historically been the hardest part of displacing Nvidia, since even a competitive chip is a hard sell if it forces developers to rewrite their existing AI codebases.

The investor mix includes Samsung Catalyst Fund and Intel Capital, suggesting potential foundry and manufacturing partnerships. MediaTek's participation indicates interest from companies that could integrate OxCore into their own product lines. Fundomo's Rajeev Surati described the economics of the licensing model this way: "OXMIQ does the opposite, and that flips a cost center into leverage."

Running on FPGA hardware rather than fabricated silicon is a meaningful caveat to Oxmiq's pitch, not a minor technical footnote. A field-programmable gate array is reconfigurable chip hardware that can be programmed to emulate a custom circuit design, which makes it useful for testing and validating an architecture before committing to the enormously expensive process of actually fabricating custom silicon, but FPGAs run meaningfully slower and less power-efficient than a purpose-built chip manufactured on a modern process node. That means every performance claim about OxCore today is a projection based on FPGA-validated architecture, not a measurement of how a fabricated OxCore chip would actually perform against Nvidia's silicon in the real world, a gap that won't close until Oxmiq or a licensee actually tapes out and manufactures a physical chip.

Why It Matters

Nvidia's dominance in AI GPUs has created supply constraints and pricing power that many in the industry find concerning. An open, licensable alternative could reduce dependence on a single vendor and potentially lower costs for AI infrastructure.

However, designing a competitive GPU is enormously difficult. Nvidia's advantage comes not just from hardware but from its CUDA software ecosystem, which has taken years to build. Oxmiq will need to address both hardware and software to be viable.

CUDA's moat is specifically a switching-cost problem, not a raw performance one. Millions of developers and thousands of companies have built AI training and inference pipelines directly against CUDA's programming model over more than a decade, and rewriting that code for a different architecture is expensive and risky even if the alternative hardware is competitive or cheaper. OxPython, Oxmiq's CUDA-and-PyTorch compatibility layer, is a direct attempt to neutralize that switching cost by letting existing code run unmodified, following the same playbook other Nvidia challengers, AMD's ROCm platform among them, have attempted with mixed success. Whether OxPython's compatibility is close enough in practice that developers trust it with production workloads, rather than treating it as a partial solution that still requires costly validation and rewrites, is the real test of whether Oxmiq's software story closes the gap its hardware story alone can't.

Industry Context

Multiple companies are attempting to challenge Nvidia's AI GPU dominance. AMD's MI series, Intel's Gaudi accelerators, and various startup approaches all compete in this space. None has yet achieved significant market share against Nvidia's entrenched position.

The licensing model is relatively unusual in GPUs. Most successful GPU companies manufacture their own chips. Oxmiq's approach is more similar to ARM's CPU licensing model, which has been highly successful in mobile and is expanding into servers.

What It Means for Users and the Industry

For the industry, Oxmiq represents another attempt to diversify the AI hardware ecosystem. Success would create more options for system vendors and potentially reduce costs. Failure would confirm Nvidia's dominance for the foreseeable future.

For investors, the $35 million round is a relatively modest bet on a potentially transformative opportunity. The payoff could be enormous if OxCore achieves meaningful market share.

What Happens Next

Oxmiq will use the funding to accelerate chip design work and build partnerships with potential licensees. The company needs to demonstrate that OxCore can deliver competitive performance and power efficiency compared to established alternatives.

Final Takeaway

Oxmiq Labs represents a bold attempt to challenge Nvidia through an open licensing model. The company's success depends on whether it can deliver competitive hardware and attract the software ecosystem support needed for adoption.

Key Points

The GPU Licensing Model

Oxmiq Labs' decision to license rather than manufacture GPU architecture is relatively unusual in the semiconductor industry. Most GPU companies, including Nvidia, AMD, and Intel, design and manufacture their own chips. ARM's successful CPU licensing model provides a precedent, but GPU licensing has not achieved comparable scale.

The licensing approach offers several advantages. It avoids the massive capital requirements of building fabrication facilities. It allows rapid scaling through multiple licensees rather than single-company production. And it reduces risk by sharing development costs across the licensee ecosystem.

However, the licensing model also faces challenges. Licensees may demand extensive customization, fragmenting the architecture and complicating software support. Nvidia's CUDA ecosystem creates powerful lock-in that licensed architectures must overcome. And licensees may be reluctant to depend on a startup for critical components.

FAQs