OpenAI bets on its own silicon: 10 GW deal with Broadcom starts 2026

Sam Altman has been telegraphing this move for years, and now it’s official. OpenAI is breaking its near-total reliance on Nvidia with a mammoth, multi-year deal to co-develop custom AI accelerators with Broadcom. We’re not talking about a small pilot program. The term sheet targets the deployment of 10 gigawatts of AI infrastructure — racks of accelerators and networking gear — with the first systems slated to arrive in the second half of 2026 and full deployment stretching to the end of 2029.

The decision is a fundamental shift in how OpenAI builds its physical footprint. By designing its own chips, the company can hard-wire the lessons it has learned from training frontier models like those powering ChatGPT directly into the silicon. Greg Brockman framed it bluntly, stating that embedding what they’ve learned from creating frontier models and products directly into the hardware will unlock “new levels of capability and intelligence.” It’s a tacit admission that off-the-shelf solutions, however powerful, leave performance on the table for a company running software as demanding as theirs.

For Broadcom, the win is strategic. This isn’t just about shipping chips. The entire cluster is scaled using Broadcom’s Ethernet, PCIe, and optical connectivity portfolio. Charlie Kawwas, Broadcom’s semiconductor president, emphasized that these custom accelerators are paired with standards-based Ethernet for both scale-up and scale-out networking, a direct counterweight to Nvidia’s proprietary InfiniBand ecosystem. The message to the industry is clear: the path to massive AI clusters doesn’t have to run exclusively through a single vendor’s proprietary interconnect.

The timing is aggressive and the scale is staggering. 10 gigawatts is utility-scale power consumption, more than many large cities use. It signals that OpenAI sees its compute needs growing at a rate that simply cannot be met by waiting in line for the next generation of GPUs. With over 800 million weekly active users and deep enterprise penetration, the pressure to control its own destiny — from the model architecture down to the silicon — has become an existential requirement, not a luxury.