July 1, 2026, (Inside AI) — Startup Oxmiq has secured $35 million in fresh funding to develop a unified chip architecture that could slash the cost of building and running AI applications. The round was led by Samsung Catalyst Fund and Fudomo, with participation from MediaTek and Pegatron Venture Capital.
The Campbell, California-based company, founded by former Intel chief architect and AMD executive Raja Koduri, aims to collapse three distinct AI system components—graphics chips, central processors, and tensor engines—into a single licensable IP block. This approach promises to reduce the hundreds of millions of dollars and multi-year timelines typically required for cutting-edge AI chip development.
Koduri told Reuters the goal is to become the "Arm of this next era," referencing the UK firm whose designs power nearly every smartphone. Oxmiq plans to offer its integrated design as intellectual property that other companies can license, much like Arm's model, but tailored for AI workloads.
The funding brings Oxmiq's total raised to $60 million. The company will use the new capital to complete its first IP product and hire more engineers. Koduri said the initial batch of intellectual property is already in the works and will soon be available commercially.
Oxmiq's architecture also includes a computing fabric that packages chiplets—multiple specialized chips—and memory into a single unit. This tight integration could improve performance and energy efficiency, key factors as AI models grow larger and more expensive to run.
Beyond licensing, Oxmiq plans to enter the custom chip market, competing with established players like Broadcom, Marvell, and MediaTek. Koduri's deep experience at Intel and AMD gives the startup credibility, but it faces an uphill battle against incumbents with vast resources and existing customer relationships.
The Economics of AI Chip Design
Designing a state-of-the-art AI chip is a high-stakes gamble. Costs can exceed $500 million, and the process from architecture to tape-out often spans two to three years. Software ecosystems, including compilers and drivers, add further complexity. Oxmiq's bet is that a pre-integrated, licensable design can dramatically shorten time-to-market and lower barriers for companies wanting custom AI silicon.
Industry analysts note that the "Arm model" has succeeded in mobile because it abstracted away hardware complexity. Applying it to AI chips is logical but untested at scale. The challenge lies in balancing generality with the extreme performance demands of modern AI training and inference.
Competing viewpoints exist. Some experts argue that AI workloads are too diverse for a one-size-fits-all IP block. Hyperscalers like Google and Amazon already design their own fully custom chips, while Nvidia dominates with its GPU-centric approach. Oxmiq must prove its unified design can match or exceed these specialized solutions across a range of tasks.
The Koduri Factor and Market Realities
Raja Koduri's track record lends weight to Oxmiq's ambitions. At AMD, he revived the Radeon graphics division; at Intel, he spearheaded the company's discrete GPU efforts. His vision of merging CPU, GPU, and tensor capabilities echoes trends in chiplet-based designs, but executing on a licensable IP model requires a different skill set—one focused on ecosystem building rather than single-product excellence.
The custom chip market is crowded. Broadcom and Marvell already serve cloud giants with semi-custom solutions, while MediaTek's investment in Oxmiq suggests a potential partnership or customer relationship. Koduri said the startup will also target this space, but details on specific customers or timelines remain scarce.
Oxmiq's progress will be closely watched. If successful, it could democratize AI chip design, enabling smaller firms to deploy custom accelerators without the prohibitive upfront costs. Failure, however, would underscore the enduring dominance of vertically integrated giants and the difficulty of disrupting semiconductor incumbents.
