Microsoft and OpenAI are getting ready to build a $100B supercomputer

Microsoft and OpenAI are secretly plotting one of the most ambitious computing projects ever conceived—a $100 billion AI supercomputer called "Stargate".

What's going on here?

Stargate is part of a five-phase project to create the infrastructure for advanced AI models.

What does this mean?

The proposed Stargate supercomputer could cost over $100 billion and require 5 gigawatts of power by 2030—the equivalent of several large data centres. It's part of a 5-phase plan, with Stargate as the final phase 5.

In the near term, Microsoft is working on a $10 billion "phase 4" supercomputer for OpenAI using Nvidia AI chips, targeted for the 2026 launch. We're currently in phase 3, with Microsoft providing hundreds of thousands of GPUs to OpenAI this year and next.

The total multi-phase plan could exceed $115 billion in costs for Microsoft - over 3x of their capital expenditure last year. Key goals are packing way more GPUs per server rack and using non-Nvidia networking tech like Ethernet instead of InfiniBand.

Why should I care?

It’s hard to imagine what OpenAI and Microsoft are plotting. Microsoft keeps investing in OpenAI rivals, but then you see numbers like this which make a few hundred million seem like chump change.

Is Stargate a part of Sam Altman’s plans of raising $7T for building AI infrastructure? Sam is vocal about the need for more powerful computers for the AI’s bidding i.e. training the models and serving them to the public.

Building this supercomputer isn’t just about money though. It also relies on OpenAI delivering quality upgrades in its models way beyond GPT-4. Both Microsoft and OpenAI are betting that stupidly high compute will unlock self-improving AI models that can learn from synthetic data.