OpenAI builds custom chips and software to reduce Nvidia dependency
In brief
- OpenAI creates software enabling AI workloads to run on multiple hardware manufacturers, not just Nvidia
- AMD and Broadcom partnerships signed in October 2025 for accelerator co-development and GPU deployment
- Custom inference chips designed with TSMC targeted for 2026 production
- Per-query cost reductions could reshape consumer AI pricing economics
Partnerships and Custom Silicon
OpenAI formalized a multi-year partnership with Broadcom in October 2025 focused on co-developing accelerator and networking systems. That same month, the company signed a deal with AMD to deploy up to 6 gigawatts of AMD Instinct GPUs, with deployment starting in the second half of 2026.
Beyond partnerships, OpenAI's Hardware organization is designing its own silicon solutions, working with TSMC on chiplet architectures and advanced packaging technologies. Custom inference chips are slated for production by 2026, giving OpenAI another option beyond buying off-the-shelf hardware.
The Software Layer
The core of OpenAI's strategy is deceptively simple. The effort centers on creating low-level platform primitives, essentially the foundational code that tells different chips how to handle the same AI tasks. This abstraction layer means OpenAI can swap hardware without rewriting its entire software stack.
Competitors have already walked this path. Google has its TPU chips. Amazon has Trainium and Inferentia. Microsoft has its own custom AI accelerator, Maia. OpenAI is now following suit, but with a twist: it's building the software to orchestrate multiple vendors at once.
The Stargate Backdrop
The Stargate initiative, OpenAI's massive data center scaling effort, provides the physical backdrop for all of this. The company's hiring tells the story. OpenAI's Compute Infrastructure team is explicitly hiring for roles focused on building automatable, repeatable systems for heterogeneous clusters.
The stakes are real. If OpenAI achieves similar results by 2026, it would reduce its per-query costs substantially, potentially changing the economics of offering AI services at consumer price points. That shift would ripple across the industry, forcing Nvidia to compete on more than just brand dominance and supply.
