Announcing our Frontier Data Centers Hub! The world is about to see multiple 1 GW+ AI data centers. We mapped their construction using satellite imagery, permits & public sources — releasing everything for free, including commissioned satellite images. Highlights in thread!

Explore the interactive insight on our website: https://t.co/4lFnZTyAEZ Or dig into our Frontier Data Centers hub yourself:

Our compute estimates for these data centers are based on satellite imagery of the cooling infrastructure, permitting documents, and company disclosures.

Using data from our Frontier Data Centers and AI Models hubs, we calculated the number of GPT-4-scale training runs achievable on the largest data center in a single month. To keep things comparable with GPT-4, we assume training is done at 16-bit precision.

Microsoft's Fairwater Atlanta (today's largest data center) could likely train over 20 models the size of GPT-4 in the course of a month. This computational power will enable AI companies to increase the number and scale of both experiments and training runs.

Using data from our Frontier Data Centers and AI Models hubs, we calculated the number of GPT-4-scale training runs achievable in a single month. To keep things comparable with GPT-4, we assume training is done at 16-bit precision.

Learn more about Alibaba Zhangbei and other data centers using our free Satellite Explorer!

In total, we estimate between 200–500 MW of total facility power is operational for Alibaba in Zhangbei today. This power capacity is close to leading US data centers, but the compute capacity is significantly lower, due to the older buildings with slower chips.

We also noticed a mid-flight change to the design of 6 buildings from 2024 to 2025. The updated design puts many more chillers on the roof, roughly doubling the power density. This is an unusual development, and suggests that more advanced chips are being installed.

How do we know the power capacity of these buildings? Cooling equipment gives it away. For example, each building shown below has 40 air-cooled chillers with 12 fans. Plugging this into our model, we estimate 38 MW of cooling capacity, which is closely related to IT capacity.

We first found out about an Alibaba facility located in Zhangbei from a web search about Chinese data centers. We then purchased high-resolution satellite images of the region and identified 20 data center buildings, many more than we expected.

We’ve identified an Alibaba data center in Zhangbei, China with an estimated 200–500 MW capacity. Roughly half of that predates the AI boom and likely hosts little modern compute. But several newer buildings show a high power density consistent with advanced AI chips. 🧵

You can find more details about this research here:

This project was funded by Google DeepMind. It was done in collaboration with researchers from @EpochAIResearch and @GoogleDeepMind — @ansonwhho, @js_denain, @DJAtanasov, @SamuelAlbanie, and @rohinmshah.

And it also allows us to study accelerations in AI capabilities. We ran simulations with synthetic benchmark data, where capabilities accelerated two-fold in 2027. Using our framework, we’re able to detect this speed-up within 2-3 months.

It lets us study improvements in AI software, where better algorithms and data need less training compute to reach the same estimated capability.

This opens several promising directions of research. For example, it lets us project future improvements in AI capabilities:

Using real-world data from our benchmark hub, we estimate these parameters to get capability scores for different models. So we can compare models even if they’re not evaluated on the same benchmarks! And we also see a clear trend in capabilities over time:

We then relate these parameters to benchmark scores with an S-curve, with 3 regimes: 1. Capability << Difficulty: performance is close to random 2. Capability ≈ Difficulty: capability grows linearly with performance 3. Capability >> Difficulty: the benchmark is saturated

The core idea is similar to Item Response Theory: assume each model has a latent capability and each benchmark has a latent difficulty This is like how chess players and puzzles both have Elo scores Finally, each benchmark has a latent slope that tells us how fast it saturates.

Most AI benchmarks share a common flaw: they saturate too quickly to study long-run trends. Our solution: “stitch” many benchmarks together. This lets us compare models across a wide range of capabilities on a single unified scale. Here’s how this works.🧵