OpenAI and Broadcom team up to deploy 10GW of custom AI chips

Sam Altman has spent months touring the globe talking about massive compute demands. Now, OpenAI is finally putting its own silicon on the map.

OpenAI just inked a multi-year deal with Broadcom to co-develop and deploy a staggering 10 gigawatts of custom AI accelerators. The rollout is scheduled to kick off in the second half of 2026 and wrap up by the end of 2029.

Instead of just picking chips off the shelf, OpenAI will design the accelerators and systems in-house. The idea is to bake lessons learned from training frontier models directly into the silicon, while Broadcom handles the development, optical connectivity, PCIe, and Ethernet networking.

A massive bet on Ethernet and custom silicon

The deal goes all-in on Broadcom networking gear, using standard Ethernet for both scale-up and scale-out connections across OpenAI facilities and partner data centers. That is a direct challenge to the proprietary InfiniBand setups that have dominated high-end AI clusters.

While official financial terms were kept under wraps, the Financial Times estimates the total cost could land between $350 billion and $500 billion.

OpenAI needs every watt it can get. The company revealed it now boasts over 800 million weekly active users alongside surging enterprise demand. To feed that appetite, OpenAI has been on an unprecedented infrastructure shopping spree. Just last week, it bought 6 gigawatts of chips from AMD. In September, Nvidia pledged a $100 billion investment alongside a 10-gigawatt letter of intent, and reports surfaced of an unconfirmed $300 billion cloud pact with Oracle.

Why it matters

We think this move shows OpenAI is refusing to remain a hostage to Nvidia pricing and supply constraints. By pairing up with Broadcom, OpenAI joins Google and Meta in the custom ASIC playbook, tailoring silicon specifically for its own architecture while backing Ethernet to break networking lock-in.

For developers and enterprise customers, this massive capacity buildout promises cheaper inference and more reliable access once the racks go live. But that relief is still years away. Deliveries do not even start until late 2026, meaning teams building on OpenAI today will not feel the benefits anytime soon.

The real risk here is financial overextension. Committing hundreds of billions across Broadcom, AMD, and Nvidia essentially bets that generative AI revenue will balloon fast enough to cover eye-watering depreciation cycles. If enterprise monetization hits even a temporary plateau, these gargantuan power contracts could quickly turn from strategic assets into suffocating liabilities.