Where AI compute is going — and how it gets powered.
AI compute is splitting two ways: giant centralized campuses chasing the lowest cost per token, and distributed sites placed closer to load, users and available power. Different shapes — one shared constraint: firm power, and the years-long wait to get it from the grid.
Centralized = a few enormous gigawatt campuses optimized for training at scale. Distributed = many smaller sites near metros, load and users, increasingly for inference. Both are growing fast — and both hit the same wall: the grid can’t energize them in time. The fix is the same too — firm power, built on site.
Distributed AI
- Scale: tens to ~100+ MW per metro or regional campus — many sites, not one
- Where: near metros, users, fiber and available power — spread across markets
- Mission: low-latency inference, regional capacity, resilience, proximity to data
- Why it’s growing: inference is exploding, latency matters, and giant single sites keep getting blocked
- The constraint: firm power per site — metro grid queues stall these builds for years
Centralized AI
- Scale: 500 MW to multi-gigawatt single mega-campuses
- Where: wherever land, water and power can be assembled cheaply — often remote
- Mission: large-scale model training — the lowest possible cost per token
- Why it’s growing: training at frontier scale rewards concentrating compute and power
- The constraint: gigawatts in one place — transmission and generation take 5–7+ years
| Dimension | Distributed AI | Centralized AI |
|---|---|---|
| What it is | Many smaller sites placed near load and users | A few enormous single campuses |
| Typical scale | Tens to ~100+ MW per site, across many markets | 500 MW to multi-gigawatt in one location |
| Where it sits | Near metros, fiber and available power | Where cheap land, water and power can be assembled |
| Primary workloads | Inference, regional capacity, low-latency services | Frontier-scale model training |
| Why it’s growing | Inference demand, latency, and blocked mega-sites | Training scale rewards concentration |
| The power problem | Firm power per site — metro queues stall builds | Gigawatts in one place — 5–7+ yr transmission |
| Relationship | Both, not either/or — the same operators build both, and both run into the same wall: the grid can’t deliver firm power fast enough. On-site generation is the unlock for either shape. | |
Distributed or centralized, the missing layer is the same: firm power, on site.
Whichever shape AI compute takes, the bottleneck is power — and metro grid connections take 5–7+ years. AnchorPower closes that gap with one modular architecture that scales from a single distributed campus to a centralized gigawatt factory.
Fully off-grid, behind-the-meter firm power under a 15-year fixed-price PPA. One modular platform scales from a 50–100 MW campus to 500 MW+ and gigawatt-scale — no interconnection queue, first power in 18–20 months.
AnchorPower systems are modular, containerized and mostly pre-fabricated — the same factory-built approach that makes modern compute fast to deploy. Power arrives the way the racks do: in modules, ready to energize.