Microsoft to use OpenAI chip designs in AI hardware push

Microsoft will use access to OpenAI’s custom AI chip development to support its own chip initiatives, CEO Satya Nadella said on a podcast released November 12.

The company, which has worked on in-house chips but trailed rivals like Google, can access OpenAI’s AI models through 2032 and research until 2030 or until a panel deems artificial general intelligence has been reached.

OpenAI is developing custom chips and networking hardware with Broadcom.

Nadella said Microsoft will use designs from both OpenAI and its internal team, citing its IP rights under a revised agreement.

The comments clarify how Microsoft’s partnership with OpenAI will shape its hardware and AI strategy in the coming years.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Microsoft access to OpenAI chip designs faces uncertain production timing

Microsoft can use OpenAI chip designs under the revised IP deal 1. OpenAI’s first custom part may not enter mass production until 2026 using Taiwan Semiconductor Manufacturing Co. (TSMC) 3-nanometer process with a systolic array architecture (a grid-like layout that efficiently performs repeated matrix operations for AI) 1.
The chip targets inference (running AI models to generate outputs), not training (teaching models using large datasets) 1. Early rollout looks limited, so Microsoft may not get near-term cost relief or reduced Nvidia reliance in Azure.
OpenAI staffs a smaller chip group than Google or Amazon 1. Each new version can run $500 million or more including software and peripherals 1, so Microsoft’s access gives it leverage in supplier talks, not a quick substitute for Nvidia GPUs.

Third-party developers can build portability tools across Microsoft’s mix of AI accelerators

Azure could host OpenAI silicon next to Maia accelerators. That mix invites independent software vendors to build hardware-agnostic compilers, debugging tools, and performance optimization layers that span both types.
Microsoft backs Open Neural Network Exchange (ONNX) Runtime, which enables cross-platform inference across multiple frameworks and hardware accelerators 2. Developers can extend it to cover new custom silicon and create portability add-ons.
System integrators (firms that combine hardware and software into turnkey solutions) can help enterprises migrate and optimize AI workloads across a mixed set of Azure accelerators. That work can span Nvidia, AMD, and Maia hardware, with OpenAI chips if integrated.