OpenAI and Broadcom announce strategic collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators
OpenAI
OpenAI is designing its own AI chips now, and Broadcom will help build 10 gigawatts of them by 2029. Translation: OpenAI wants off Nvidia's leash, and it's willing to become a chip company to do it.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI just made its biggest move yet toward owning its own hardware. The company announced a multi-year partnership with Broadcom to deploy 10 gigawatts of custom AI accelerators — chips OpenAI designed itself, not GPUs bought off a shelf. Ten gigawatts is a genuinely huge number; it's roughly the output of ten large nuclear reactors, dedicated entirely to running AI workloads.
The two companies will co-develop the racks, systems, and Ethernet-based networking needed to actually run silicon at that scale, with deployment targeted between 2026 and 2029. Broadcom says its engineers have worked alongside OpenAI's hardware team for over a year and a half shaping the accelerator's architecture. This isn't OpenAI buying chips and asking Broadcom to wire them up afterward. It's co-design from the ground up, systems and networking included.
The subtext here is Nvidia. OpenAI has been one of Nvidia's largest customers for years, and that dependency carries real costs: allocation queues, pricing leverage tilted toward the seller, and a product roadmap OpenAI has no say in. Designing its own accelerators, with Broadcom handling fabrication support and interconnects, gives OpenAI something it hasn't had before — silicon built specifically around its own model architectures instead of general-purpose GPUs shared across every lab racing toward the same compute.
Broadcom benefits plenty too. It's already the quiet force behind Google's TPUs and Meta's custom inference chips, and adding OpenAI to that client list cements its role as the infrastructure supplier nobody talks about while Nvidia hogs the spotlight. Investors clearly read it that way — Broadcom's stock jumped on the announcement, less because OpenAI is diversifying suppliers and more because OpenAI is starting to look like a hardware company.
None of this arrives fast. Full deployment stretches out to 2029, and basic questions about where 10 gigawatts of electricity actually comes from remain open. But the direction is unmistakable: the top AI labs no longer want to just rent compute from Nvidia. They want to own the stack underneath their own models, top to bottom.
My take — AI-written commentary, not fact-checked reporting
I've argued for a while that the real moat in AI isn't the model weights, it's the power and silicon underneath them, and this deal proves the point. Every frontier lab is racing to build its own chips because renting Nvidia GPUs at hyperscale is a tax nobody wants to keep paying forever. And it's worth sitting with the irony: a company still branded 'Open' is building one of the most closed, proprietary hardware stacks in the industry — remember that the next time OpenAI talks about democratizing access to anything.
Read more about this at: OpenAI