ai race map plots AI organisations by HQ and aggregates all-time AI funding by country. Each pin shows a single org; each country tier (S–D) is a function of the total funding of orgs with HQ inside it.
An organisation qualifies for the map if it satisfies any of:
The bar is uniform worldwide — a Berlin or Shanghai org clears or fails the same gate a San Francisco one does. Country-vs-country comparisons would be meaningless otherwise. New candidates surfaced by research, submissions, or rebuilds are evaluated against these three legs; orgs that fall below all three are excluded until they cross one.
What we exclude. Even when a company appears on a major AI list, we cut it if its core product isn't AI. Concretely:
ai_infra carve-out. Inference clouds and GPU-as-a-service
(Baseten, Fireworks AI, FriendliAI, Together AI, Fal, Crusoe) stay on the map but don't
count toward country tier totals — their funding reflects infra spend, not training spend.
Datacenter operators are similarly excluded from tiers.
Every org has a type: ai_native (private AI-focused
company that raises VC), ai_infra (inference cloud or
GPU-as-a-service whose product is AI compute, but who doesn't train models),
oss_lab (nonprofit / research collective),
sovereign (government-funded national lab), big_tech
(diversified company with an AI arm), or datacenter (physical infra).
Countries rank independently per metric (funding, compute, power). The choropleth uses three buckets: the rank-1 country pops in saturated color, the top-25%-by-rank tail glows in medium tone, the rest fade to a low-saturation fill. Each metric has its own palette — green for funding, amber for compute, cyan for power — so the active metric reads at a glance even in a static screenshot.
Each org is merged from multiple datasets. When two sources disagree on a field (HQ, funding total, latest round), the higher-ranked source wins. Source rank:
HQ city/country comes from primary-source pages or, for orgs that lacked it,
a separate manual backfill (hq-backfill-2026-05-08).
Coordinates are looked up from a hand-maintained city table — orgs in a city
not yet in the table show up in the data probe but not on the map.
Datacenters are a separate map layer alongside labs (toggle in the bottom-right).
Each DC pin shows current power draw (MW) and derived FP16 compute (EFLOPS) —
EFLOPS is computed as H100-equivalents × 989 TFLOP/s (NVIDIA H100 SXM
dense FP16, no sparsity). The H100-equivalent count comes directly from Epoch.
A datacenter qualifies for the map if it satisfies any of:
Same uniform-bar logic as the org gate: a Narvik or Riyadh site clears or fails the same gate an Abilene one does. Sites below all three legs (generic colocation, crypto-mining campuses, sub-frontier inference POPs, "intent" announcements without site selection or named developer) are excluded.
Status. operational = at least partially live and
consuming AI compute today. planned = announced or under construction
with a site, developer, and groundbreaking confirmed but no current power draw.
Coverage caveat. Epoch's frontier list is explicitly US-focused
— they cover the two-to-three largest sites for each major US frontier lab and
estimate ~15% of global delivered AI compute. We layer non-US frontier sites on
top of Epoch from a hand-curated global-data-centers source with
primary-source URLs (press releases, EuroHPC pages, government announcements),
so the same uniform gate applies to every country. Compute and MW figures from
non-Epoch sources are self-reported; Epoch's satellite + permit derivation is
treated as more authoritative when both cover the same site.
Every number you see in the side panel — total funding, latest round equity, valuation — carries a superscript that links to the source the figure came from. Click the number to read the original announcement; hover to see which snapshot the data was pulled from.
Funding totals are best-effort: undisclosed rounds, secondary-only sales, and non-equity capital (debt, compute commitments, partnership-style "investments" that aren't equity) are inconsistent across sources. We treat Epoch's totals as canonical for frontier labs and document anything else inline.
Country ranks shift whenever a single very large round closes (an OpenAI or xAI round can shift the US funding total by tens of billions). Compute and power ranks are similarly sensitive to a handful of frontier datacenters coming online.