OpenAI’s Stargate Gets Bigger: The Real Cost of AGI Infrastructure

OpenAI just announced another expansion of Stargate, their sprawling compute infrastructure project. More data centers, more capacity, more everything. The stated goal: building the physical foundation for the Intelligence Age, aka AGI.

I’ve been watching this project since the rumors first leaked, and honestly, the scale keeps surprising me. We’re not talking about a few server racks here. Stargate is already one of the largest capital projects in tech history, and this latest round adds enough compute to train models that would have seemed like science fiction five years ago.

Let’s be real about what this means. Training frontier models isn’t cheap. Every time OpenAI announces a new data center, they’re effectively saying “we believe the next generation of AI will require ten times the compute of the last one.” That’s a massive bet on scaling laws holding up, and on the returns from more compute continuing to justify the exponential costs.

But here’s the thing I don’t see enough people talking about: this kind of infrastructure buildout has real consequences. The energy requirements alone are staggering. Stargate’s power draw will rival small cities. And while OpenAI talks about renewable energy offsets—and I believe they’re serious about it—the immediate reality is that AI training is becoming one of the most energy-intensive activities on the planet.

There’s also the geopolitical angle. Control over compute is becoming a form of power in itself. Countries that can build and maintain these facilities will have a massive advantage in the AI race. Stargate isn’t just a technical project; it’s a strategic asset.

I also wonder about the diminishing returns. At some point, throwing more GPUs at a problem yields less and less improvement. We saw hints of this with GPT-4—incredible model, but the jump from GPT-3 wasn’t as dramatic as the jump from GPT-2. Scaling laws aren’t infinite, and OpenAI is betting billions that they’re not nearing the ceiling yet.

That said, I have to respect the audacity. Building infrastructure at this scale is hard. Really hard. Supply chain issues, construction delays, talent shortages—the list of potential failure points is long. If OpenAI pulls this off, they’ll have built something genuinely unprecedented.

The article mentions “meeting growing AI demand,” which is corporate-speak for “everyone wants access to the best models, and we need to serve them.” But I think there’s more to it. This isn’t just about serving ChatGPT queries faster. It’s about building the platform for whatever comes after AGI. They’re laying tracks for a train that hasn’t been built yet.

I’ll be watching closely to see how this plays out. The technical challenges are fascinating, but the real story might be about who gets left behind when compute becomes the new oil.