OpenAI is building the software plumbing to run AI workloads across chips from multiple providers, a deliberate effort to loosen the grip that Nvidia’s CUDA ecosystem has held over the entire AI industry for years.

The initiative involves developing low-level runtime systems and compiler infrastructure that allow AI models to operate efficiently on hardware from different vendors.

The multi-vendor blueprint

In January 2026, OpenAI struck a deal with Cerebras to provide up to 750 MW of compute capacity dedicated to AI inference over a three-year period.

In October 2025, OpenAI announced a multiyear strategic collaboration with Broadcom focused on developing custom AI accelerators. The ambition there is staggering: a target scale of 10 GW, with mass production anticipated in 2026.

AMD is in the picture too. OpenAI has reportedly signed multi-gigawatt deals for MI450-class accelerators, adding yet another hardware vendor to a growing roster. And the company is exploring integration with Google TPUs and AWS Trainium, which would mean OpenAI’s infrastructure touches nearly every major chip architecture in the market.

Why software is the real story

Most AI workloads are deeply entangled with Nvidia’s CUDA, a proprietary software platform that makes it easy to program Nvidia GPUs and extremely painful to use anything else.

OpenAI’s approach is to build software infrastructure, including compiler work for cross-architecture portability, that abstracts away the differences between chip vendors.

The company is actively recruiting engineers for compute roles with job postings emphasizing the creation of low-level platform structures. These roles focus on building systems that facilitate the operation and automation of diverse hardware environments. The postings have been ongoing as of 2026, suggesting this is a sustained buildout rather than a one-off project.

What this means for investors

For Nvidia, this changes the negotiating dynamics. When your biggest customer can credibly threaten to shift workloads to AMD or Broadcom or Cerebras, your pricing power erodes.

For AMD and Broadcom, this is the opportunity they’ve been positioning for. AMD’s MI450-class accelerators get a much more compelling market story if OpenAI’s software layer can deliver competitive performance without requiring developers to abandon familiar programming models. Broadcom’s custom accelerator work with OpenAI could establish the company as a major player in AI compute.

Cerebras stands to benefit significantly from the validation alone. A commitment of up to 750 MW from the world’s most prominent AI company is the kind of endorsement that can attract additional customers and investors.

The risk to watch is execution complexity. Running AI workloads efficiently across heterogeneous hardware is one of the hardest problems in systems engineering. Performance gaps between native optimization and abstracted runtimes can be significant. If OpenAI’s cross-architecture software adds even a modest latency or throughput penalty, the cost savings from hardware diversification could be eaten up by efficiency losses.