- OXMIQ, founded by Raja Koduri, has raised a $35 million in Series A funding, bringing its total funding to $60 million
- The round lands as Groq, once Nvidia’s sharpest inference rival, rebuilds after licensing its core technology to Nvidia and losing its founder in the deal
- OXMIQ never builds a finished chip at all: it licenses GPU architecture to whoever else wants to
Groq spent a decade building a chip good enough that Nvidia wanted it more than Groq’s independence. In December 2025, the company licensed its core inference technology to Nvidia in a deal reportedly worth around $17 billion, and its founder, Jonathan Ross, left to join Nvidia along with roughly 90% of Groq’s engineering team.
Now another AI chip startup founded by a semiconductor exec is betting it can dodge that outcome by never building a finished chip Nvidia would want in the first place.
A different kind of chip company
OXMIQ, based in Campbell, California, has closed a $35 million Series A, bringing its total funding to $60 million. Founded by Raja Koduri, who spent years as Intel’s chief architect and led the Radeon Technologies Group at AMD before launching OXMIQ from stealth in August 2025.
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. It follows a $20 million seed round backed by Tenstorrent and MediaTek.
“A licensable core with an open architecture means design teams everywhere can build the custom AI silicon they need,” says Koduri.
Unlike Groq, Cerebras, or Etched, OXMIQ does not manufacture its own chip. It licenses GPU architecture instead. OxCore, its core product, combines a CUDA-compatible GPU engine, tensor processing and orchestration into a single configurable design that other companies build their own silicon around.
A companion tool, OxQuilt, lets customers combine chiplets, memory, and packaging without locking into a single foundry, while OxPython runs existing CUDA and PyTorch code without modification. For example, a semiconductor firm without the capital for a full chip programme could license OxCore rather than spend years and billions of dollars designing a processor from scratch.
Why Groq’s exit is the story to watch here
The AI chip market is projected to grow from around $100 billion in 2026 to more than $400 billion by 2030, and Nvidia still controls an estimated 80% of AI training chip demand. That dominance has increasingly drawn challengers into Nvidia’s orbit rather than pushing them further away.
Groq’s licensing deal is the clearest recent example: Nvidia absorbed the technology and the talent, and Groq, under interim leadership, had to raise $650 million just to rebuild a company around the inference market it had originally pioneered. Cerebras took the opposite route, going public in May 2026 at a valuation above $56 billion, though its IPO filing also disclosed heavy revenue concentration in a small number of customers.
Both remain hardware companies with finished products that are either good enough for Nvidia to buy or risky enough to rely on a handful of buyers. OXMIQ’s answer is that neither outcome applies to a company that never finishes a chip of its own: its revenue comes from licensing fees paid by whoever else is doing the building, not from a product Nvidia would ever need to acquire.
Backers who are also customers
Samsung Catalyst Fund and MediaTek are reinvesting from OXMIQ’s seed round. David Goldschmidt, senior vice president and managing director of Samsung Catalyst Fund, says, “OXMIQ’s novel AI core and software platform enable heterogeneous compute for efficient, custom inference solutions serving large-scale agentic workloads.”
Rajeev Surati, a partner at Fundomo, adds, “Raja has built silicon at every layer of the stack, and he knows exactly where the constraints sit. Most compute IP makes the customer bend their memory, packaging, and foundry around the chip. OXMIQ does the opposite, and that flips a cost centre into leverage. We backed this team because they will define how AI computing gets built this decade.”
The round also drew Pegatron Venture Capital, the investment arm of one of the world’s largest electronics manufacturers, and Taiwan-based CDIB-TEN and Darwin Ventures. These investors with direct ties to the chip foundries and packaging houses OXMIQ needs as licensees, not just as capital sources.
Jim Keller, chief executive of Tenstorrent and previously a chip architect at Apple, AMD and Tesla, has joined OXMIQ’s board; Tenstorrent was also a seed investor and is the first partner running OxPython on its Wormhole and Blackhole platforms.
OXMIQ is wagering that being the infrastructure on which other companies build is safer than building a finished chip of its own. Whether that model generates enough revenue to outlast the hardware rivals it is trying to avoid becoming is the question its next round will need to answer.
