Raja Koduri’s latest venture, Oxmiq Labs, has just closed a $35 million Series A round. The startup doesn’t manufacture chips; it licenses a GPU architecture — OxCore — aimed at companies that want to integrate custom AI accelerators without enduring a full, multi-year design cycle.

The model echoes ARM’s approach in the CPU world: an intellectual property block that chipmakers can take, customize, and send to fabrication without redesigning compute pipelines, memory controllers, and interconnects from scratch. OxCore targets neural network inference and training workloads, meaning Large Language Models and generative AI. In an ecosystem dominated by NVIDIA, this approach opens a door for alternative architectures potentially closer to the needs of organizations that deploy on their own infrastructure.

The round brings the startup’s total funding to $60 million. The plan is to expand the engineering team and bring OxCore to a maturity level suitable for integration by industrial partners. Koduri is no stranger to bold bets: he led GPU development for AMD’s Vega and Navi lines and tried to build a discrete GPU division at Intel with the Xe project. With Oxmiq, the focus shifts from hardware manufacturing to intellectual property, a sector with high margins but the challenge of convincing enough licensees to bet on a new architecture.

For those evaluating on-premise deployment of LLMs, this initiative points in an interesting direction. Today, hardware choices are often constrained to high-performance NVIDIA GPUs, with acquisition costs and energy consumption that weigh heavily on Total Cost of Ownership. Custom architectures built from blocks like OxCore could someday yield chips optimized for specific enterprise workloads, offering greater control over latency, throughput, and energy efficiency. This is especially relevant where data sovereignty and GDPR compliance push toward self-hosted infrastructure, away from public clouds.

Of course, turning a licensable blueprint into a working production chip remains complex. Fabrication partners, physical design, validation, and unit volumes that justify the investment are all necessary. Oxmiq’s bet is that demand for specialized AI accelerators will grow enough to sustain a “licensed chip” ecosystem, much like what happened for mobile SoCs. If it works, the on-premise AI hardware landscape could become far more fragmented and competitive.