@jasminewsun If you find anyone who's worked on the Microsoft data centers in Mount Pleasant WI: when is the second data center expected to be completed? What have the obstacles been during construction? Could also talk to people working on the Oak Creek and Paris natural gas projects in WI.
@ShanuMathew93@azeem@ExponentialView A lot of the chart looks wrong based on our research, and is missing a lot of data centers (not sure if it’s supposed to be comprehensive).
@TimDarcet@AymericRoucher The Epoch AI data only covers about 25% of total installed AI compute as of 2026 Q1. Also the lines shown are the largest single data center used by each company, not the total capacity used (which I think OP is estimating).
We're looking for a Researcher to join Epoch AI's data center tracking project! You'd help us monitor global frontier AI datacenter buildouts in detail—using satellite imagery, permits, and other sources to inform policymakers and the public.
Jokes aside, I think the coloring is from dust, and they seem to periodically clean or repaint the roofs to keep them reflective and cool. The top two buildings used to be as dark the bottom-left three now. You can see this by stepping through our satellite explorer.
I've seen people say Claude Mythos Preview was trained on Blackwell GPUs, but this seems unlikely to me given that Anthropic's biggest training cluster is all Trainium: https://t.co/zHClo4O0f5
Besides Stargate UAE, I haven't seen major progress on >100 MW data centers in the Gulf Region in the past year. But I've only had a brief look, so I wouldn't be very surprised to find another one.
After several months researching data centers mostly solo, I'm now leading a small team! I'm excited to improve and scale the research more rapidly at https://t.co/huLYo6mw8R
How far are autonomous robots from real-world deployment?
We reviewed robot capabilities across 14 concrete tasks. We find that while navigation is deployed at scale, manipulation mostly is not.
It's cool how Epoch datasets interact. We're using the Chip Sales dataset to estimate global AI data center capacity, and using the Frontier Data Centers dataset to inform chip numbers!
AI data centers can now use as much power as New York State uses on the hottest days of the years.
We find that data centers currently have a total capacity of around 30 GW.
A bit overlooked how well Amazon has executed its data center builds.
Stargate and Colossus have gotten more buzz, but Project Rainier Indiana is the most powerful in the world, despite using a less advanced chip (Trainium2 vs Blackwell) and Anthropic raising less $ than OAI
Increasingly, US hyperscalers are thinking about building data centers in the UAE
The conventional wisdom is that this is because the UAE has cost advantages over the US like cheaper energy and faster construction.
I wrote a deep-dive report on whether this is right. 1/
If 100k of the best US chips come online each year, then 1 GW looks achievable by 2028. Policy makes this uncertain, though. You can explore the data for yourself, with linked sources and calculations: https://t.co/tVk8UVIL2E
Getting 1 GW worth of *chips* before 2028 could also be a challenge. The US reportedly plans to approve up to 100k advanced AI chips per year for G42. Oracle might also be allowed to put their chips into Stargate UAE, but I'm not sure. https://t.co/6Z6wwSe3qQ
When will Stargate UAE reach 1 GW? We don't see clear signs beyond 200 MW yet. Optimistically, 8 more 100 MW buildings could start construction this December, and take 1.5 years like the first two, hitting 1 GW by Q3 2027. That’d still be over a year behind leading US projects.
OpenAI's Stargate Abilene campus keeps expanding. Based on new permit filings showing plans for a 9th and 10th building, we now estimate over 2GW of total facility power by Q2 2027. The land is already getting cleared. https://t.co/zgXzhNCQMv
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!
AI inference costs plummet 9-900x/year (o3, DeepSeek R1). @benmcottier warns that evaluations are expensive now, but we need them to understand frontier AI capabilities before they're cheap & widespread.👇