Facilities tracked
346
Each with at least one cited source
About & Methodology
How the AI Data Center Index Works
The AI Data Center Index is a source-backed public database tracking the physical infrastructure that powers large-scale AI. It exists to make the global compute buildout legible — not just to industry insiders, but to anyone trying to understand where AI infrastructure actually lives.
Countries covered
64
Across 6 continents
Operators indexed
225
Hyperscalers, sovereigns, and specialists
Tracked capacity
117.2 GW
Selected MW where published
Bitter Research Nexus
Mapping the compute and software frontiers.
AI Data Center Index is maintained by Bitter as an open auto-research instrument for the physical compute frontier: where AI infrastructure is being built, who controls it, and how the market is scaling.
Bitter starts from the Bitter Lesson: general methods that scale with compute, including search and learning, tend to win. Understanding the compute substrate is therefore strategic work, not trivia.
The index is built to remove human bottlenecks: public sources become normalized records, coverage checks, exports, and reports instead of one-off research notes.
Agent software frontierBitter FrontierSource-backed field notes on coding agents, harnesses, operating policy, permissions, evidence, and what serious builders should test next.Operating layerBitter.shThe agent-ready operating base for building and improving software systems with receipts, deploys, checkout, memory, and governed work loops.
Sourcing Standard
Every entry requires at least one verifiable public source before it is published. Sources are linked on each facility page so readers can evaluate the evidence themselves.
What counts as a source
Official press releases, SEC and regulatory filings, planning applications, power purchase agreements, and reporting from credible industry publications (Data Center Dynamics, Bloomberg, Reuters, The Information).
What does not count
Rumors, anonymous tips, social media posts, or single-source blog speculation. If the only evidence is a LinkedIn post or an unattributed claim, the facility is not included until better evidence appears.
Uncertainty is explicit
Where capacity or status data is approximate, the record says so. The coverage report shows which fields are missing across the full dataset.
What Each Record Contains
Identity & location
Facility name, operating company, country, region, city, and geographic coordinates for map placement.
Power capacity
Selected public MW capacity, scoped where known as IT load, system power, facility power, program capacity, or another disclosed basis. Blank when no credible public capacity is available.
Status & timeline
Operational, under construction, planned, or announced — plus dated milestones where available (announcement year, construction start, operational date).
Energy & AI focus
Primary energy source (renewable, nuclear, gas, mixed) and AI workload type (training, inference, mixed) when disclosed.
Sources
Every entry links to the original announcements, filings, or reporting that substantiate it.
Methodology
Inclusion criteria
A facility is included if it is primarily designed for or significantly allocated to AI workloads (training, inference, or both) and has at least one verifiable public source. Traditional colocation and enterprise data centers are excluded unless they have a documented AI-specific allocation.
Capacity figures
Megawatt values reflect the best publicly available number for each facility. Some operators report total site power, others report IT load, system power, grid connection, or program capacity. The index stores the selected figure with its capacity scope where known. Treat MW figures as directionally useful, not perfectly comparable across operators.
Energy classification
"Renewable" means the operator has a specific clean energy commitment (PPA, on-site generation, or 100% renewable matching) for that site. "Nuclear" means a direct nuclear power supply or signed nuclear PPA. "Mixed" means the facility draws from the local grid without a specific clean energy commitment. "Gas" means natural gas is the disclosed primary source.
Status definitions
Operational: facility is live and serving compute workloads.
Under construction: ground has been broken and construction is documented.
Planned: site acquired, permits filed, or detailed plans published.
Announced: operator has publicly stated intent but no construction evidence yet.
Update frequency
The index is updated continuously as new sources become available. There is no fixed publication schedule — when a new facility is announced or an existing entry has new information, it is added as soon as it can be sourced and verified.
Known limitations
The index depends on public disclosure. Facilities that have not been publicly announced or reported are not included. Coverage is strongest in North America and Europe; Southeast Asia, Africa, and Latin America are known gaps. The coverage transparency pages show exactly which fields and regions are incomplete.
Data Source Categories
Company filings
Government records
Industry publications
Research & policy
All sources are cited inline on each facility page. Readers can evaluate source quality directly — the index does not ask for trust without evidence.
Frequently Asked Questions
- What is an AI data center?
- A facility designed for the high-density compute requirements of AI training and inference. Unlike traditional data centers optimized for web serving or storage, AI facilities are built around GPU clusters and require significantly more power per rack — often 50–100 kW versus 5–10 kW for conventional racks.
- Where are the largest AI data centers located?
- The largest concentrations are in the United States (Virginia, Texas, Iowa), with major facilities in Europe (Ireland, Netherlands, Denmark) and Asia (Japan, South Korea, Singapore). Multi-gigawatt campuses are under development in the Middle East. See the largest facilities ranking and the interactive map.
- Which companies operate AI data centers?
- The major hyperscalers — Microsoft, Google, Amazon (AWS), and Meta — operate the largest fleets. OpenAI partners with Microsoft for compute. Oracle, CoreWeave, and Lambda Labs operate large GPU-focused facilities. National AI initiatives in India, the UAE, Saudi Arabia, and France are also tracked. See the operators index.
- How much power do AI data centers use?
- Individual facilities range from a few megawatts for smaller inference clusters to multi-gigawatt campuses. Microsoft's Stargate initiative targets 5 GW of new AI data center capacity. The full global AI compute fleet draws tens of gigawatts — a figure growing rapidly as frontier model training scales.
- Is the data free?
- Browsing is free. Every facility page, operator profile, and geographic view is public HTML. Bulk dataset exports (CSV, JSON, GeoJSON) and structured intelligence reports are available through the data access page.
- How do I submit a new facility?
- Use the submission form. Include the facility name, location, capacity in MW, operating companies, and at least one public source. Submissions are reviewed before publishing.