Nine AI Factories, One Map: Which Bets Are Winning?
Over nine case studies we visited the largest AI factories on Earth, from a converted Memphis factory to a Norwegian fjord. Each made one big bet: speed, outside capital, a network of sites, its own chips, any building that holds a GPU, a decade-long head start, debt, brute force, or national strategy. This final part puts them side by side on one timeline and one scoreboard, and pulls out the patterns that matter more than any single site.
1. The nine bets
2. Master timeline
Every case study on one chart
Each bar runs from the first concrete step (land, financing or announcement) to today, with a dot at first compute and a dashed extension to the next planned milestone. Switch to days to first compute to line the sites up from a common start. Tap a row for details.
3. How big, really?
Part 1 showed three honest ways to count a gigawatt. Across the series, the headline figure was almost always the largest of them: a nameplate, a target or a campus plan. Estimates of power actually in use, mostly from Epoch AI, are smaller.
Headline vs estimated now
Megawatts, log scale. Grey is the headline figure; colour is the best estimate of capacity in use today (hatched where we found none and show the first tranche instead).
Google and China are left out: Google does not publish site sizes and runs a fleet, and China's capacity is spread across many state and company sites.
4. Scoreboard
Rank the bets your way
Each site is scored 1โ5 on six questions. Untick the questions you do not care about and the ranking updates. Scores are editorial judgements based on Parts 1โ9; tap a site to see the reasons. With all six questions ticked, Project Rainier ranks first; since this site is written partly with Anthropic's Claude, which trains on Rainier, we show every score and reason so you can check and re-weight them.
5. Five patterns
What the nine have in common
Tap a pattern to see which sites show it most strongly.
6. Who carries the risk
If AI revenue grows more slowly than AI spending, who pays? The series shows the answer is spreading. At Colossus, Fairwater, Rainier and Google, the owner carries most of it on its own balance sheet. At Abilene and Hyperion, investors and bondholders own most of the buildings. Neoclouds put it on lenders secured by GPUs. Utilities building gas plants for data centers put some of it, unless regulators prevent it, on ratepayers.
The risk chain
Tap a link. Arrows show who depends on whom being paid.
7. What to watch in 2027
8. Series map
This series sits alongside three others in "Build the Machine": the AI Compute Stack (the layers from power to chips), Producing Intelligence (one model through one factory) and Who Runs the AI Factory (the people who run each layer).
9. Reading the numbers
- Sources
- Every figure comes from Parts 1โ9, which cite filings, company announcements, regulators, Epoch AI and the press, and flag secondary claims with โ ๏ธ.
- Timeline
- Start points differ by site (land purchase, financing, announcement or construction start), so "days to first compute" compares like with like only roughly; each row states its start.
- Sizes
- "Now" values are Epoch AI estimates (Colossus 2, Abilene, Fairwater Atlanta, Rainier), CoreWeave's reported active power across 51 data centers, or the first tranche where nothing is running yet.
- Scoreboard
- Editorial 1โ5 scores; higher always means better from that question's point of view (for example, a higher "neighbours" score means fewer local harms reported).