OpenAI builds heterogeneous AI stack

OpenAI is building software to run AI workloads across chips from multiple vendors, according to a CryptoBriefing report published on June 1. The effort is aimed at reducing dependence on Nvidia’s CUDA ecosystem by letting OpenAI place and schedule jobs across different kinds of accelerators, the report said. CryptoBriefing said the work involves low-level runtime systems and compiler infrastructure, and named AMD and Broadcom among the hardware partners involved. OpenAI did not immediately publish a corresponding announcement on Tuesday. ### Why would OpenAI want software that spans multiple chip vendors? Nvidia’s CUDA stack has become a standard layer for much of the AI industry, which means customers that build around it can face switching costs when they try to use other hardware. CryptoBriefing reported that OpenAI’s project is designed to reduce that dependence by abstracting workloads away from a single vendor’s software environment. (cryptobriefing.com) That would let the company use a broader mix of chips as it expands training and inference capacity. AMD already has a formal infrastructure relationship with OpenAI. Data Center Frontier reported that AMD and OpenAI announced a multi-year, multi-generation partnership to deploy up to 6 gigawatts of GPU compute, beginning with a 1 gigawatt phase in 2026 using AMD’s Instinct MI450 accelerators. That arrangement gives OpenAI a concrete reason to build software that can manage more than one hardware architecture. (cryptobriefing.com) ### Where do AMD and Broadcom fit in? Broadcom is part of OpenAI’s custom-silicon push. OpenAI said on October 13, 2025 that it had entered a multi-year partnership with Broadcom to deploy 10 gigawatts of OpenAI-designed AI accelerators by 2029, alongside co-developed network systems. OpenAI said the first deployments would start in 2026. (datacenterfrontier.com) CryptoBriefing said the new software effort sits alongside those hardware relationships, giving OpenAI a way to run workloads across Nvidia, AMD, Broadcom and other chip providers rather than treating each platform as a separate island. CryptoBriefing also named Cerebras in a related report about the software layer. (openai.com) ### What does “heterogeneous” mean in practice? A heterogeneous stack means the same company can use different processors for different jobs. In practice, that can mean sending some workloads to general-purpose GPUs, others to custom accelerators, and others to systems optimized for networking or inference, depending on cost, availability and performance. CryptoBriefing reported that OpenAI is building the runtime and compiler tools needed to make that possible. (cryptobriefing.com) The key technical problem is not only buying chips. The harder task is making software decide where work should run, and then keeping models portable enough to move between systems with different architectures and toolchains. That inference is supported by CryptoBriefing’s description of “low-level runtime systems and compiler infrastructure,” and by OpenAI’s existing mix of GPU and custom-chip partnerships. (cryptobriefing.com) ### Does this mean OpenAI is abandoning Nvidia? OpenAI’s reported move does not mean Nvidia is out. CryptoBriefing framed the effort as a way to reduce reliance on Nvidia, not replace it altogether. Other reports on OpenAI’s chip strategy have described AMD and Broadcom as additions to, rather than substitutes for, Nvidia capacity. (cryptobriefing.com) OpenAI’s next visible milestones are tied to hardware rollouts already on the calendar. Broadcom and OpenAI said their first OpenAI-designed accelerator deployments would begin in 2026, while the AMD partnership is set to start with a 1 gigawatt phase in 2026, according to company and industry reports. (openai.com) (cryptobriefing.com)