Home
News
Tech Grid
Interviews
Anecdotes
Think Stack
Press Releases
Articles
  • Investment

Oxmiq Labs Unveils GPU IP and Software for AI and Multimodal Workloads


Oxmiq Labs Unveils GPU IP and Software for AI and Multimodal Workloads
  • by: Business Wire
  • |
  • August 5, 2025

Oxmiq Labs Inc., a Campbell, California-based GPU software and IP startup founded by renowned GPU architect Raja Koduri, emerged from stealth on August 5, 2025, after two years of intensive IP development. With a team boasting over 500 years of combined experience and hundreds of patents, Oxmiq aims to redefine GPU architecture for the multimodal AI era, securing $20 million in seed funding from investors including MediaTek.

Quick Intel

  • Funding: $20M seed round from MediaTek, angel investors, and strategic partners.

  • Leadership: Raja Koduri, former Intel and AMD GPU chief, as Founder and CEO.

  • Technology: Licensable GPU IP (OXCORE™) and software stack (OXCapsule™, OXPython™).

  • Innovation: Nano-agents, RISC-V cores, near/in-memory computing, and light transport.

  • Scalability: Supports edge AI to zettascale data centers via OXQUILT™ chiplet architecture.

  • Partnerships: Collaborates with Tenstorrent and supports Mihira Visual Labs.

Re-Architecting GPUs for Multimodal AI

Oxmiq is addressing the shift toward multimodal computing, where text, audio, video, images, and 3D environments converge. Unlike fixed-function AI accelerators, Oxmiq’s GPU IP offers general-purpose flexibility, integrating with mainstream operating systems via standardized APIs and unified memory models. Its licensable IP incorporates nano-agents using RISC-V cores, near-memory and in-memory computing, and light transport for enhanced performance in graphics and AI workloads.

The OXCORE™ GPU core combines scalar, vector, and tensor engines in a modular architecture, scalable from single-core edge devices to thousands of cores in data centers via the OXQUILT™ chiplet design. This enables customization for specific workloads, supporting nano-agents, native Python acceleration, and SIMD/CUDA compatibility.

Software-First Strategy

Oxmiq’s Software First approach prioritizes developer experience with OXCapsule™, a unified software ecosystem that abstracts hardware complexity for seamless deployment across diverse platforms. A key component, OXPython™, allows Python-based NVIDIA CUDA applications to run on non-NVIDIA hardware without code changes, launching on Tenstorrent’s Wormhole™ and Blackhole™ platforms in 2025. “OXPython’s ability to bring Python workloads for CUDA to AI platforms is great for developer portability,” said Jim Keller, CEO of Tenstorrent.

Capsule and OXPython will be available through partners by late 2025, with OXCORE and OXQUILT licensable in H1 2026.

Capital Efficiency and Market Traction

With $20M in seed funding from MediaTek and others, Oxmiq’s licensing-first model avoids the high costs of chip manufacturing, offering capital efficiency. The company has already recorded software revenue and holds a minority stake in Mihira Visual Labs, where Koduri serves as a strategic advisor. Mihira leverages Oxmiq’s GPU IP for its cinematic AI platform. “OXMIQ is building the deep-tech infrastructure that powers our next chapter,” said Shobu Yarlagadda, CEO of Mihira Visual Labs.

Strategic Vision and Industry Impact

“We want to be Arm for the next generation,” Koduri told Reuters, emphasizing Oxmiq’s goal to provide scalable GPU IP for AI and graphics. By focusing on licensable IP, Oxmiq bypasses the $500M+ costs of custom chip development, enabling partners to integrate its technology into mobile, automotive, and data center solutions. This aligns with global pushes, like India’s Semiconductor Mission, to foster AI innovation.

 

About OXMIQ Breakthrough

Beyond software, OXMIQ delivers a complete GPU hardware IP stack that powers silicon solutions scaling from Physical AI in edge devices and autonomous robots to enterprise edge infrastructure and zettascale data centers. Their scalable GPU core, OXCORE™, integrates scalar, vector, and tensor compute engines in a modular architecture customizable for specific workloads, enabling nano-agents, native Python acceleration, and compatibility with SIMD/CUDA paradigms.

OXCORE scales seamlessly from single core for compact Physical AI applications to thousands of cores for data centers via the OXQUILT™ chip-let architecture. Through OXQUILT, customers can configure optimal ratios of compute, memory and interconnect for their needs and achieve significant reduction in time to market, R&D and production costs v/s current industry standard methodologies.

  • Oxmiq LabsGPU InnovationRaja KoduriAI ComputingMultimodal AI
News Disclaimer
  • Share