Raja Koduri's OXMIQ raises $35M for licensable AI GPU

Key Takeaways

• OXMIQ raised $35M Series A co-led by Samsung Catalyst Fund and Fundomo, bringing total funding to $60M
• The startup aims to license GPU IP through its OxCore architecture, similar to how Arm licenses CPU designs
• Raja Koduri argues AI compute costs must fall 50-100x for mass adoption in markets like India

Raja Koduri, who spent over two decades leading GPU development at AMD and Intel, has raised $35 million in Series A funding for his startup OXMIQ. The company is building a licensable AI GPU architecture that lets companies and governments design custom AI chips without starting from zero.

Samsung Catalyst Fund and Fundomo co-led the round. Existing investors MediaTek, Pegatron Venture Capital, Intel Capital, and others participated. Total funding now stands at $60 million.

What is OXMIQ building?

OXMIQ calls itself "Arm for AI GPUs." Arm Holdings licenses CPU designs to companies like Apple, Qualcomm, and Samsung. They build the architecture; customers build the chips. OXMIQ wants to do the same for AI accelerators through its flagship product, OxCore.

The pitch is straightforward. Developing AI chips from scratch costs billions. Nvidia dominates the market. Most companies and governments can't compete. OXMIQ offers a middle path: license proven GPU IP, customize it for specific workloads, and manufacture it at a foundry.

Koduri plans to complete OxCore by late 2025 or early 2026. After that, customer deployments and commercial tape-outs begin. The company launched a public beta of its software stack in November 2025. Twenty companies and ten universities are now testing it across nearly 300 GPUs.

Why sovereign AI matters to this bet

"India cannot play in AI without taking control of the cost of it," Koduri told the Economic Times. "AI compute costs would need to fall by 50-100 times to drive mass adoption." He argues that chip design captures 92% of margins. If countries want affordable AI infrastructure, they need to own critical chip IP.

This is the sovereign AI thesis. Governments don't want their AI capabilities dependent on a handful of American chip suppliers. China faces export controls. India, Southeast Asia, and the Middle East are building massive data centers. They want alternatives.

OXMIQ is targeting semiconductor firms, neocloud providers, AI infrastructure builders, and government programs. Initial commercial focus: India and Southeast Asia, including Malaysia and Indonesia.

The India play is already live

OXMIQ has engineering teams in Hyderabad and Bengaluru. Earlier this year, it partnered with AM Intelligence Labs, part of AM Green Group, to architect a 2 GW renewable-powered AI compute platform in Uttar Pradesh. The first 1 GW phase is expected to go live by end of 2027.

That's a significant anchor project. It signals the kind of customer OXMIQ wants: large-scale infrastructure builders who need custom silicon, not off-the-shelf Nvidia hardware.

Can a licensable GPU model work?

Arm proved the model works for CPUs. But GPUs are different. AI workloads evolve rapidly. Nvidia's dominance comes from CUDA, its software ecosystem, not just silicon. A licensable GPU architecture needs a software stack that developers will actually adopt.

OXMIQ's beta program suggests they understand this. Getting 20 companies and 10 universities testing the software is a start. But the real test comes when customers tape out commercial chips and run production workloads.

Koduri has credibility. He led AMD's Radeon Technologies Group, then joined Intel in 2017 as Chief Architect to build their discrete graphics division, Intel Arc. He departed Intel in early 2023. If anyone can design a licensable GPU architecture, he has the track record.

The competitive landscape

Nvidia holds over 70% of the AI training chip market. AMD is a distant second. Intel struggles. Google, Amazon, and Microsoft build custom AI chips for their own clouds but don't license them.

Startups like Groq, Cerebras, and Tenstorrent take different approaches. Groq and Cerebras build their own chips. Tenstorrent, led by Jim Keller, offers licensable AI IP, making it OXMIQ's closest competitor.

OXMIQ's angle is cost. If sovereign AI programs and emerging market data centers can't afford Nvidia's prices, and can't invest billions in R&D, a licensable architecture might be the only practical option.

What comes next

The $35 million goes toward completing OxCore. Customer deployments and commercial tape-outs follow. The 2 GW Uttar Pradesh project gives OXMIQ a reference customer. If that goes live in 2027 running OXMIQ-designed chips, it validates the entire model.

Koduri isn't promising to replace Nvidia overnight. He's offering an alternative for customers who can't or won't pay Nvidia prices, and for governments that want chip independence. Whether that market is big enough to support a standalone company remains the open question.

Frequently Asked Questions

What is OXMIQ?

OXMIQ is an AI chip architecture startup founded by former Intel graphics chief Raja Koduri. It licenses GPU intellectual property through its OxCore product, allowing companies and governments to build custom AI chips without developing silicon from scratch.

How much funding has OXMIQ raised?

OXMIQ raised $35 million in Series A funding co-led by Samsung Catalyst Fund and Fundomo. Total funding is now $60 million, including participation from MediaTek, Intel Capital, and other investors.

What is sovereign AI?

Sovereign AI refers to a country's ability to develop and control its own AI infrastructure, including chips, data centers, and models, without depending on foreign suppliers. Several nations are pursuing sovereign AI programs to reduce reliance on US chip companies.

Who are OXMIQ's competitors?

OXMIQ competes with Nvidia and AMD in the broader AI chip market. For licensable AI IP specifically, Tenstorrent, led by Jim Keller, is the closest competitor. Groq and Cerebras build their own chips rather than licensing designs.

When will OXMIQ's chips be available?

OXMIQ expects to complete OxCore architecture by late 2025 or early 2026. Customer deployments and commercial chip tape-outs will follow. The company's partnership with AM Intelligence Labs targets first silicon deployment by end of 2027.

Source: Tech-Economic Times / ET