Compute Atlas
Where AI’s computing power lives.
Where the next wave is being built.
Source: Epoch AI, updated 28 Oct 2026. *Coverage estimate has a wide 90% interval of 30–80%. These are selected major AI sites, not a census of all data centres. An additional planned Australian site appears separately. Headline totals do not change with filters.
1 November review: Map, directory, totals and growth chart retain the 29 October capacity snapshot. Newer model estimates are discussed below; a target date passing is not independent confirmation of commissioning. No new sites or historical observations have been added.
Explore the buildout
Marker area scales with the selected current capacity (minimum size for visibility). Hollow markers have no operating capacity recorded. Location data: Epoch AI; S7 is approximate. Basemap: Natural Earth, public domain.
The project directory
Search and filters apply to the map and directory. Rows are ranked by the selected measure. Map region buttons change the view only. Unknown capacity is left blank, never inferred from power.
| Project / location | Hardware owner | Stage | H100 equivalents | IT power · MW |
|---|
The scale is changing
Estimates and forecasts are labelled separatelyCompute in the tracked sites
Millions of H100 equivalents · same 96-site cohort
Historical bars use the last dated capacity milestone on or before each date. October uses the current CSV estimate. 2027 is a projection from Epoch’s project milestones, carrying the last value forward. This is not a global growth rate or a promise of delivery.
The electricity backdrop
All data centres worldwide · TWh per year
Source: IEA, Key Questions on Energy and AI (2026). Approximately 1.96× growth. This includes conventional data centres as well as AI. Energy consumption and compute capacity are different measures.
The bulletin board
Selected developments · reviewed 1 Nov 2026A broader Australian buildout
Data Centres Australia and DC Byte report 1.6 GW of operational capacity across 163 facilities and forecast 3.9 GW by 2030, up from 3.2 GW in April. This covers the wider data-centre market, not just AI. The release does not explicitly define the power boundary; these figures are kept separate from the atlas’s IT-power totals.
Read the industry release ↗Lake Mariner: model versus commissioning
Epoch now estimates about 293,000 H100e and 210 MW IT for the Anthropic-associated buildings, versus this snapshot’s 58,565 H100e and 42 MW. Its 31 October milestone cites an earlier delivery target. TeraWulf’s August filing describes phased delivery; no newer commissioning confirmation was verified in this review. The increase is therefore not adopted as confirmed capacity here. The separate Core42 buildings are not added again.
Inspect Epoch’s estimate ↗ · Operator filing ↗A clearer picture of the physical buildout
Epoch’s 28 October dataset covers 96 major sites and estimates 15.2 million H100 equivalents of current capacity. It provides a research baseline for this atlas; its totals remain a partial view of the world.
Explore the source dataset ↗Google Cloud prepares Rubin availability
NVIDIA says Google Cloud A5X general availability is expected in November, with planned European expansion. The stated ambition is up to 960,000 Rubin GPUs across multiple sites—not a verified installed count today.
Read NVIDIA’s announcement ↗Sydney’s S7 remains a planned project
NEXTDC’s current site page lists 550+ MW of planned IT capacity at Eastern Creek. That is an infrastructure specification, not a disclosed GPU inventory. It stays outside the operating totals.
Read the operator’s current page ↗This is an editorial snapshot, not an automatically refreshing news feed. An announcement date, a commissioning target and a verified operating date are different events.
Australia in focus
Southgate Melbourne
46,488 H100e41 MW estimated IT capacity, with NVIDIA B300 hardware in Epoch’s model.
A single tracked site is not Australia’s total AI compute. Capacity estimate confidence: moderate.
S7, Eastern Creek
550+ MWPlanned IT capacity. H100-equivalent capacity is undisclosed and is not estimated here.
Confidence in the operator’s published plan: high. Confidence in eventual delivery and chip count: unknown.
What are we actually measuring?
H100 equivalents: the closest useful common currency
An H100 equivalent (H100e) expresses AI chip performance relative to one NVIDIA H100. Under Epoch’s convention, 1 H100e = 1.979 × 10¹⁵ theoretical 8-bit operations per second. Epoch uses dense theoretical 8-bit throughput, halving sparse specifications where necessary. Hardware mixes may be inferred with a one-quarter deployment lag. This is a model comparison, not measured workload performance or a physical H100 count. A B300, TPU or Trainium chip can contribute a different amount. xAI itself used “H100 GPU equivalents” in its January 2026 funding announcement. Confidence in this unit choice: high.
What about FLOPS, petaflops and exaflops?
FLOPS means floating-point operations per second: a rate of computation. A petaflop is 10¹⁵ FLOPS; an exaflop is 10¹⁸ FLOPS. Precision and sparsity matter: FP64, FP32, FP16 and FP8 figures are not interchangeable. This atlas keeps Epoch’s label “8-bit OP/s” for its normalised measure. Peak specifications do not predict actual training speed, useful tokens per second or model quality.
Why show megawatts and gigawatts too?
MW and GW measure power, not compute. IT power describes the servers and related IT equipment; total facility power also includes cooling and other overheads. TWh measures energy used over time. A gigawatt of older chips and a gigawatt of newer chips can produce very different compute. No fixed MW-to-H100e conversion is used here.
Operating, building and planned
Operating means Epoch identifies an operational site; a campus can still be expanding. Building uses Epoch’s non-operational projects, which its methodology limits to projects where construction has started. Planned is used for the additional S7 operator announcement; it is not evidence of construction. Announced ambitions without verified construction or sufficient location evidence are kept in the bulletin rather than counted as operating sites. Hollow markers denote no current capacity recorded, not zero future potential. Predictions are kept out of current capacity totals.
Sources, confidence and limitations
Core data and locations: Epoch AI, “AI data centers”, accessed 29 Oct 2026, reused under CC BY 4.0. Presentation, filtering, historical aggregation and rounding are adaptations by Compute Atlas. Company affiliations retain Epoch’s uncertainty labels. Compute estimates have moderate confidence and are not audited inventories. Data are biased towards large, recent US facilities. A hardware owner is not necessarily the building owner or the model developer using it. No global market-share ranking is inferred from this sample.
Australia planning source: NEXTDC S7 page, checked 29 Oct 2026. Energy chart: IEA (2026), “Key Questions on Energy and AI”, CC BY 4.0. Country outlines: Natural Earth public-domain 1:110m data, simplified for display. The map does not show campus boundaries or network connections.