Table of Contents
Oxmiq’s CUDA-compatible architecture targets sovereign computing customers
Oxmiq Labs, the AI semiconductor startup founded by former Intel chief architect Raja Koduri, has raised $35 million in a Series A round co-led by Fundomo and Samsung Catalyst Fund. The round brings the company’s total capital to $60 million, including $20 million in earlier seed funding. Other backers include MediaTek, Pegatron Venture Capital, Darwin Ventures, and Morgan Creek Digital, while Intel Capital is participating as a strategic intellectual property partner rather than a traditional venture investor — a notable arrangement given Koduri’s history at the company.
Koduri is one of the more recognizable names in GPU architecture. Before serving as Executive Vice President and Chief Architect at Intel, he led AMD’s Radeon Technologies Group as Senior VP and Chief GPU Architect, giving him a hand in two of the three major GPU ecosystems of the past decade. After leaving Intel, he founded Mihira Visual Labs, a platform for content creators that he has described as the test environment that led directly to Oxmiq’s conception.
The company itself publicly launched in mid-2025 after roughly two years of stealth IP development, and now operates out of Campbell, California, with an additional hardware and engineering site in Hyderabad, India. Oxmiq says the new funding will go toward finishing its first batch of IP and integrating with early customers.
Licensing model and market strategy
Oxmiq’s business model is worth pausing on, because it’s fundamentally different from what Nvidia and AMD sell. The company doesn’t manufacture chips at all. Instead, it licenses GPU and AI chip designs — effectively renting out a customizable architecture that customers integrate into their own silicon, rather than buying finished accelerators off the shelf. It’s closer to the Arm playbook than the Nvidia one.
Data-center GPUs from Nvidia and AMD remain expensive and supply-constrained, with the most advanced parts booked far in advance. Oxmiq is pitching its licensing model as a lower-cost, decentralized alternative for organizations that would rather own their compute than queue for someone else’s. The target list includes semiconductor companies that lack internal GPU design teams, cloud data-center operators, and AI system builders.
The most interesting customer segment, though, is what Oxmiq calls “sovereign computing” — governments and regional cloud providers that want domestic alternatives to US-based GPU vendors and local control over their AI infrastructure. That pitch taps into a genuine industry anxiety about how concentrated AI compute has become. Oxmiq also frames its timing around what it calls the “Age of Inference,” arguing that as AI models move from training into large-scale deployment, efficient inference compute becomes the primary bottleneck. That’s a reasonable read of where the market is heading, even if it’s also a convenient one for a company selling inference-oriented IP.
Hardware architecture and specifications
The flagship product is OxCore, a licensable architecture that combines a CUDA-compatible GPU engine, a tensor processing engine, and a CPU-based orchestration engine in a single scalable core. Koduri has described it as a new kind of computing core — a building block that encapsulates CPU, GPU, and tensor functions and can be replicated across scales, from edge devices to large data centers. For now, OxCore runs on FPGAs for live prototype demonstrations. Customers can see working silicon, but there’s no mass-produced ASIC yet, which is an important caveat when weighing the company’s performance claims.
The second platform is OxQuilt, a chiplet integration architecture for heterogeneous compute. The idea is to let customers mix GPU cores, tensor engines, CPUs, and memory in a single package, with flexibility across foundries, memory types, interconnect standards, and packaging options. That aligns with where the broader industry is already heading — AMD’s Instinct line and the hyperscalers’ custom ASICs have all leaned into chiplet-based designs — but Oxmiq is offering the approach as licensable IP rather than a finished product.
Underneath it all, Oxmiq is building its hardware and software GPU IP on the open RISC-V instruction-set architecture. The company says the design is parameterized to scale from ultra-low-power “physical AI” applications like robotics up to zettascale data centers. That’s an enormous range, and it remains a claim rather than a demonstrated capability.
Ecosystem issues
The smartest decision Oxmiq has made is arguably CUDA-ecosystem compatibility. Nvidia’s software moat is at least as formidable as its hardware, and any alternative silicon that can’t run existing CUDA workloads starts from a deep hole. By letting customers bring their existing software to Oxmiq-derived chips, the company lowers one of the biggest barriers to adoption. It’s also layering on its own software — what it calls “electron-to-token machines,” or ETMs, aimed at optimizing the full path from physical compute resources to AI model outputs — alongside cloud software for AI factories and autonomous agents.
But the friction runs the other way, too. Buying an Nvidia or AMD accelerator gets you guaranteed, validated hardware. Licensing IP and building custom silicon from it means assuming the cost, complexity, and multi-year risk of a chip design program — a heavy lift for organizations without deep silicon experience, which is precisely the audience Oxmiq is courting. And CUDA compatibility on paper is not the same as CUDA compatibility in production. Real-world performance parity against Nvidia’s deeply entrenched ecosystem has tripped up plenty of well-funded challengers before.
That said, the backing from MediaTek, Pegatron, and Samsung’s venture arm suggests serious hardware players see something worth betting on. Whether that translates into shipping silicon is the question the next couple of years will answer.
