OpenAI’s Stargate Gets Bigger: The Compute Play Behind AGI

OpenAI just announced they’re expanding Stargate, their sprawling compute infrastructure project, to build out more data center capacity. The goal is straightforward: they need more hardware to run the models that are supposed to lead to AGI.

This isn’t a surprise to anyone watching the industry. Training and inference at scale require stupendous amounts of compute. I’ve seen estimates that a single frontier model training run can cost hundreds of millions of dollars in GPU time alone. And that’s before you factor in the power, cooling, and networking.

What’s interesting here is the scale. Stargate isn’t just another data center cluster. It’s a multi-billion dollar bet that the path to AGI is paved with silicon. OpenAI is essentially saying: we can’t get there without building our own infrastructure, because cloud providers can’t or won’t give us enough.

I’ve been skeptical about the “just throw more GPUs at it” approach for a while. There’s a diminishing returns problem — each new model generation needs exponentially more compute for what feels like linear gains in capability. But OpenAI seems to think the curve is still worth riding.

The expansion adds capacity in phases, with new data centers coming online in the US and potentially other regions. No hard numbers on total compute or cost, but given that the original Stargate project was pegged at around $100 billion over several years, this is a serious escalation.

One thing that stands out: OpenAI is also talking about custom chips and optimized networking. They’re not just buying off-the-shelf NVIDIA gear (though plenty of that too). They’re designing their own infrastructure stack. This is what hyperscalers like Google and AWS have done for years. Now OpenAI is joining that club.

Is this necessary? Probably. The demand for inference is exploding — ChatGPT alone serves hundreds of millions of users. And training GPT-5 or whatever comes next will dwarf previous runs. If OpenAI doesn’t build this, they’ll hit a wall.

But there’s a risk here too. Tying up that much capital in hardware means less flexibility. If the AI landscape shifts — say, if smaller models become more capable or if a new architecture changes the compute requirements — OpenAI could be left with a lot of expensive, purpose-built infrastructure.

Still, I respect the conviction. OpenAI is betting that compute is the moat, and they’re digging deep. Whether AGI arrives on schedule or not, they’ll have a hell of a lot of computing power to show for it.