Methods ledger · modeled values only
What does “H100-equivalent” actually mean?
This page exposes the arithmetic behind every conversion made by this project. Central estimates feed map sizes and totals; ranges and raw card counts remain first-class data. These are peak arithmetic comparisons—not promises of equal training time.
H100e =
site dense 8-bit OP/s
÷ 1.979 POP/s16-bit fallback: dense FP16/BF16 ÷ 0.9895 PFLOP/s
site dense 8-bit OP/s
÷ 1.979 POP/s16-bit fallback: dense FP16/BF16 ÷ 0.9895 PFLOP/s
One denominator, disclosed precision.
Epoch primarily normalizes dense FP8 or INT8 throughput to an H100 SXM's 1.979 POP/s dense FP8 baseline. When a chip lacks a trustworthy 8-bit specification, the atlas uses Epoch's fallback principle: dense FP16/BF16 divided by the H100's 989.5-TFLOPS dense 16-bit baseline. Vendor figures labeled sparse are halved first.
Each project-modeled row identifies its precision and evidence. Epoch-derived campus and cluster values retain Epoch's supplied normalization rather than being recalculated here.
External stock reference
How much AI compute has been shipped?
NVIDIA does not publish a clean quarterly GPU-unit table. Epoch estimates accelerator shipments from compute revenue, product mix and prices, then checks the model against NVIDIA's direct Hopper and Blackwell disclosures. This is a reference for the atlas's order of magnitude, not a site-allocation source.
Covered designers · 31 Dec 2025
20.34M H100e estimated cumulative shipments; approximate aggregate range 17.56M–23.29M. Covers Nvidia, Google, Amazon, AMD, Huawei and Cambricon.
NVIDIA · 31 Dec 2025
13.66M H100e estimated cumulative shipments (11.9M–15.55M). This is compute-equivalent capacity, not literal H100 units.
NVIDIA · 31 Mar 2026
17.05M H100e (14.86M–19.43M), represented by an estimated 11.82M physical accelerator packages.
Comparison boundary
Shipment estimates track dedicated data-center AI accelerators delivered and ready for installation, not necessarily installed, networked, online, or located. The atlas tracks source-attributed clusters and campuses and includes some post-reference-date capacity. The ratio is a scale cross-check, not a census-completeness estimate.
Epoch AI Chip Sales explorer · full shipment methodology. Epoch dataset updated 11 June 2026, retrieved 12 July 2026; values above are generated from data/global_compute_reference.json and recomputed from the archived download in data/epoch_ai_chip_sales/.
OSINT observability ladder
Can the world’s GPUs actually be tracked?
Only probabilistically. Public evidence can estimate production and identify many large deployments, but there is no public serial-number ledger joining a finished accelerator to a server rack. Each stage answers a different question.
| Stage | Useful public signals | What they establish | Hard limit |
|---|---|---|---|
| Foundry allocationTSMC / Samsung / SMIC | Foundry reports, node and end-market revenue, wafer capacity, customer concentration, earnings calls | Aggregate leading-edge capacity and direction of AI demand | TSMC does not publish wafer or CoWoS allocations by named chip/customer. A wafer is not a finished accelerator. |
| Packaging + HBMCoWoS, substrates, memory | Advanced-packaging capacity, HBM bit shipments, supplier capex and lead times | A strong cross-check on the ceiling and product-generation mix | Different packages consume different area and HBM; double counting across component chains is easy. |
| Finished shipmentsDesigner → OEM / cloud | Vendor revenue and direct unit disclosures, inventory, purchase commitments, OEM server orders | The best global stock estimate; this is what the Epoch denominator models | Sale or delivery does not show whether a chip is installed, networked, idle, in inventory, or retired. |
| Freightair, sea, road | Bills of lading, customs records, consignee names, weights, ports and routes; AIS/ADS-B only as corroboration | Occasionally links an OEM or integrator shipment to a country or consignee | HS 8542 combines many ICs; descriptions can be generic or confidential. High-value semiconductors often move by air, whose waybills are generally not public. AIS/ADS-B identifies the vehicle—not its cargo. |
| Installed clusterscampus attribution | Operator disclosures, permits, utility interconnects, procurement, export filings, satellite imagery and site photos | The strongest evidence that compute exists at a physical site | Secret sites, leased cloud capacity, unlocated company-wide totals and phased installations remain unresolved. |
Practical conclusion: track cohorts and confidence ranges, not individual chips. TSMC is primarily the wafer foundry and packaging provider; finished inventory then passes through chip designers, OSATs, board/server manufacturers, distributors and customers, so “TSMC-stocked ships” is usually the wrong model of the chain.
Two axes, kept separate
Hardware grade is not deployment scale.
Consumer / edge
Gaming GPUs and edge accelerators are dispersed, poorly attributable, and excluded from both this atlas and Epoch’s dedicated data-center AI denominator. Converting the global gaming stock to H100e would create false precision.
Enterprise · 1k–10k H100e
Material training or inference deployments below the atlas’s default frontier view. Lower the main map’s scale control to expose the 1,000-card / 1,000-H100e full-record inventory. Smaller processed-Epoch systems are retained only in aggregate statistics.
Frontier · 10k–100k H100e
The atlas’s default band: large enough to support frontier-scale work at its observation date, while still commonly deployed as a bounded cluster or campus phase.
Gigacluster · ≥100k H100e
A project convention for exceptionally concentrated present compute. It describes attributable deployment size, not networking quality, achieved utilization, or a vendor product tier.
China completeness control
The official 42 is not a public site list.
What is official
MIIT reports 42 completed 万卡级 intelligent-compute clusters at year-end 2025. The National Data Administration and National Energy Administration repeat the total, but publish neither a roster nor a counting definition.
What can be named
Public operator and government evidence supports roughly 18 strict cutoff-compatible logical deployments, or at most about 23 using looser correspondence. Approximately 19–24 official identities remain anonymous or ambiguous.
Why dots do not equal clusters
The official unit is a logical cluster. A physical campus can contain several fabrics, and a logical system can span multiple buildings. The atlas instead consolidates one record per physical campus and keeps distributed systems unlocated.
Accounting rule
Epoch’s anonymized Chinese records probably capture part of the unnamed tail. They are not added to named records without a source-backed match, and the residual is never represented with synthetic centroids.
Primary controls: Digital China Development Report 2025 and ODCC/CAICT summary. Detailed reconstruction: research/fragments/china-42-reconciliation-2026-07-13.md.
China quantity scenarios
Useful bubbles without invented precision.
Where a Chinese site discloses cards but not comparable FLOP/s, the map now assigns a deliberately wide accelerator-family scenario. Where only an AI-facility power envelope is public, it uses 100–900 H100e per facility MW with a 400 central display case. These values make otherwise invisible projects comparable in rough area, but are marked NON-ADDITIVE below and excluded from rankings, country shares and global totals.
Known family
Ranges distinguish 910B/910C, Zhenwu, MetaX, Enflame, Hygon, Kunlunxin, Sophgo, Moore Threads and undisclosed NVIDIA generations. Only 910C has a narrow first-party dense-FP16 anchor.
Unknown domestic card
The default 0.03–0.80 H100e/card family bracket is intentionally broad. Fully unidentified accelerators use 0.02–1.00. The midpoint is a display scenario, not the expected value of a statistical distribution.
Power estimate
Card-derived facility power uses model-specific envelopes where TDP/system data exist and 0.50–2.04 kW/card for unknown domestic hardware. The drawer shows low–high MW alongside the central bubble size.
Accounting boundary
Strong site-aggregate FP16 conversions and Epoch-supplied H100e remain additive. Generic unknown-chip and MW-to-compute scenarios never enter the pie chart or “share of global compute” denominator.
FLOP Map cross-check
725 displayed does not mean 725 clean campus records.
FLOP Map displayed 725 clusters at capture time; its downloaded broad Epoch mirror contains 786 rows and a separate 34-addition file. Epoch’s current recommended processed download contains 482 records and explicitly removes low-certainty rumors, possible duplicates and systems not known to be contiguous. This atlas keeps the processed file as its statistical input and snapshots FLOP Map separately under data/flopmap/ for leads and deduplication. For the anonymous-China crosshair, FLOP Map supplies the synthetic coordinate while Epoch’s current raw download supplies all 308 quantity rows. The marker responds to filters but remains non-additive because it can overlap named sites and processed Epoch records.
Live conversion ledger
Project-modeled sites
Generated from data/clusters.json on every build, so this table cannot silently drift away from the map.
| Site | Current H100e | Buildout H100e | Per card | Method and evidence |
|---|---|---|---|---|
| Jinko Zhongwei 1 GW Compute CenterChina | —not installed | ~400k100k–900k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] |
| Nscale Cedarvale Ward County AI CampusUnited States | —not installed | ~262.8k262.8k–262.8k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculationapproximately 104,000 GB300 GPUs × 2.5265 H100e/GPU [1] [2] [3] |
| IREN Childress AI CampusUnited States | —not installed | ~252.7k252.7k–252.7k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculationapproximately 100,000 B300 GPUs in the Microsoft contract × 2.5265 H100e/GPU; separate NVIDIA-contract hardware is not imputed [1] [2] [3] [4] [5] [6] |
| Argentum Boosteroid Bielsko-Biała AI Data CenterPoland | —not installed | ~118.7k118.7k–118.7k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation47,000 contracted GB300 GPUs × 2.5265 H100e/GPU [1] [2] [3] [4] |
| Firmus Southgate St Leonards AI FactoryAustralia | —not installed | ~93k93k–93k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation36,800 GB300 GPUs × 2.5265 H100e/GPU [1] [2] [3] [4] [5] |
| Aolani–ByteDance Malaysia Blackwell ClusterMalaysia | —not installed | ~81.9k81.9k–81.9k | 2.274×2.274–2.274× | vendor-spec dense BF16 calculation36,000 planned B200 GPUs multiplied by the atlas B200 factor; expansion remains planned and site-unresolved [1] [2] |
| Nscale Loughton AI CampusUnited Kingdom | —not installed | ~58.2k58.2k–58.2k | 2.527×2.527–2.527× | vendor-spec dense BF16 calculation23,040 GB300 GPUs × 2.5265 H100e/GPU; NVIDIA reports 180 PFLOPS dense BF16 per 72-GPU NVL72 [1] [2] [3] |
| IREN Prince George AI Cloud CampusCanada | ~53.2k50.3k–56k | ~53.2k50.3k–56k | 2.280×2.160–2.400× | family-mix dense 16-bit throughput estimateapproximately 23,300 delivered GPUs; range reflects undisclosed H100/H200 and B200/B300 splits plus MI350X comparability [1] [2] [3] |
| Yotta D2 Greater Noida AI CampusIndia | —not installed | ~52.4k52.4k–52.4k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation20,736 B300 GPUs × 2.5265 H100e/GPU [1] |
| China Unicom Wujiang Intelligent Computing CenterChina | —not installed | ~50k12.5k–112.5k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] |
| China Unicom Wuhu Intelligent Computing CenterChina | —not installed | ~48k12k–108k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] [3] |
| Firmus Southgate MelbourneAustralia | —not installed | ~46.5k46.5k–46.5k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation18,400 GB300 GPUs × 2.5265 H100e/GPU [1] [2] [3] [4] |
| Mistral Compute at EclairionFrance | —not installed | ~45.5k45.5k–45.5k | 2.527×2.527–2.527× | vendor-spec dense BF16 calculation18,000 GB200/GB300 GPUs × 2.5265 H100e/GPU; both official NVL72 specifications imply 2.5 PFLOPS dense BF16 per GPU [1] [2] [3] [4] |
| Anonymized Chinese System (Epoch record 189)China | ~36k4k–200k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] |
| Mistral Bruyères-le-Châtel AI CampusFrance | —not installed | ~34.9k34.9k–34.9k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation13,800 GB300 GPUs × 2.5265 H100e/GPU; NVIDIA reports 180 PFLOPS dense BF16 per 72-GPU NVL72 [1] [2] [3] |
| Northwest Zero-Carbon Qingyang AI Compute CenterChina | —not installed | ~31.7k7.9k–71.3k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] |
| 5C MEM01 Together AI FactoryUnited States | —not installed | ~27.8k27.8k–27.8k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation11,000 reported GB200/GB300 GPUs × 2.5265 H100e/GPU; both generations have the same official NVL72 dense BF16 throughput [1] [2] [3] [4] |
| Suzhou Municipal AI Computing CenterChina | —not installed | ~25.2k6.3k–56.7k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] [3] |
| Zhengzhou National Supercomputing Internet AI4S ClusterChina | ~25k10k–55k | ~25k10k–55k | 0.250×0.100–0.550× | heterogeneous-pool scenario bracket100,000-card pool with undisclosed accelerator mix; a 10,240-card ScaleX reference exceeds 5 EFLOPS but does not label precision and is not assumed representative of every installed card [1] [2] [3] [4] [5] |
| Global AI Endicott GB300 ClusterUnited States | ~23.3k23.3k–23.3k | ~23.3k23.3k–23.3k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation9,216 GB300 GPUs × 2.5265 H100e/GPU [1] |
| Sharon AI ESDS SPOCHub B300 ClusterAustralia | —not installed | ~20.7k20.7k–20.7k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation8,208 B300 GPUs × approximately 2.5265 H100e/GPU [1] |
| Alibaba Feitian East China Phase IIIChina | —not installed | ~20.7k5.2k–46.5k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] |
| TACC Horizon at Sabey Round RockUnited States | ~20.2k20.2k–20.2k | ~20.2k20.2k–20.2k | site totalaggregate conversion | system aggregate dense FP16 conversionTACC reports 20 EFLOPS BF16/FP16 divided by 0.9895 PFLOP/s per H100 SXM [1] [2] |
| Runze Pinghu International Information Port AI CenterChina | ~18k2k–100k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] [3] |
| EasyLink Zixing Intelligent Computing CenterChina | —not installed | ~17.7k4.4k–39.9k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] [3] |
| Taiwan Mobile Vantage Guishan AIDCTaiwan | —not installed | ~17.7k17.7k–17.7k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation7,000 GB300 GPUs × 2.5265 H100e/GPU [1] [2] [3] [4] |
| NHN Factory XSouth Korea | ~17.4k17.4k–17.4k | ~17.4k17.4k–17.4k | 2.274×2.274–2.274× | vendor-spec dense tensor calculation7,656 B200 GPUs multiplied by 2.274 H100e per B200 [1] |
| Eni Green Data Center HPC CampusItaly | ~17k17k–17k | ~17k17k–17k | site totalaggregate conversion | lineage-preserving mixed-system calculationExisting Epoch-derived 3,227 H100e plus 13,920 MI300A normalized at the Epoch MI300A factor, approximately 13,795 H100e [1] [2] [3] [4] [5] [6] [7] [8] [9] |
| Beijing Tongzhou AI Intelligent Computing CenterChina | —not installed | ~16k4k–36k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] [3] |
| Domyn Colosseum AI SupercomputerItaly | —not installed | ~14.6k14.6k–14.6k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation5,760 reported Grace Blackwell GPUs multiplied by the atlas GB200/GB300 dense-BF16 factor; exact physical host is undisclosed [1] [2] [3] [4] |
| Five Elephants Cloud Valley AI Computing CenterChina | ~14.4k3.6k–32.4k | —no modeled buildout | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] [3] |
| DataSection Inzai B300 AI Data CenterJapan | —not installed | ~12.8k12.8k–12.8k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation5,080 B300 GPUs × approximately 2.5265 H100e/GPU [1] [2] |
| TensorWave / TECfusions Tucson AMD AI ClusterUnited States | ~12.1k10.8k–12.1k | ~12.1k10.8k–12.1k | 1.321×1.321–1.321× | dense FP8 vendor-spec conversion plus Epoch lineage8,192 MI325X × (2.6149 PF dense FP8 / 1.979 PF per H100) = about 10,824 H100e, plus Epoch's 1,321-H100e MI300X tranche in the central case [1] [2] [3] [4] [5] [6] [7] |
| xAI Atlanta QTS ClusterUnited States | ~12.1k12.1k–12.1k | ~12.1k12.1k–12.1k | site totalaggregate conversion | mixed-fleet vendor-spec calculation12,000 H100 at 1.0 H100e plus 448 A100 at approximately 0.315 H100e [1] [2] |
| Wuhu Tier Immersion-Cooled AI BaseChina | ~11.7k2.9k–26.3k | —no modeled buildout | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] |
| Foxconn K-1 AI FactoryTaiwan | —not installed | ~11.6k11.6k–11.6k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation4,608 GB200 GPUs × 2.5265 H100e/GPU [1] |
| Nscale Verne Iceland Blackwell DeploymentIceland | —not installed | ~11.6k11.6k–11.6k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculationapproximately 4,600 Blackwell Ultra GPUs × 2.5265 H100e/GPU [1] |
| DataSection Bangkok B200 AI Data CenterThailand | —not installed | ~10.7k10.7k–10.7k | 2.274×2.274–2.274× | vendor-spec dense BF16 calculation4,696 B200 GPUs × approximately 2.274 H100e/GPU [1] |
| CoreWeave Barcelona AI CloudSpain | ~10.2k10.2k–10.2k | ~10.2k10.2k–10.2k | 1.000×1.000–1.000× | vendor-spec dense tensor calculation10,224 H200 GPUs multiplied by 1.0 H100e per H200; H200 changes memory, not peak tensor arithmetic [1] |
| Beijing Lianyun Fuzhi Edge and AI ParkChina | —not installed | ~10k2.5k–22.4k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] |
| ByteDance MegaScale ClusterChina | ~9.8k2.5k–31.1k | —no modeled buildout | 0.800×0.200–2.530× | China card-count / power-envelope scenario · NON-ADDITIVEundisclosed NVIDIA-generation scenario [1] |
| NAVER Corp SejongSouth Korea | ~9.8k9.1k–9.8k | ~9.8k9.1k–9.8k | site totalaggregate conversion | mixed-system lineage-preserving calculationExisting Epoch estimate of 706 H100e for 2,240 A100s plus 4,000 B200 GPUs at the atlas's 2.274 H100e per B200 convention; assumes the prior A100 system remains active [1] [2] [3] [4] [5] |
| Anonymized Chinese System (Epoch record 186)China | ~9k1k–50k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] |
| Shenzhen Ascend 910C Intelligent Computing ClusterChina | ~7.8k7.6k–8.1k | ~7.8k7.6k–8.1k | 0.785×0.760–0.809× | official-system dense FP16 conversionHuawei Atlas 900 A3 publishes 288.7–307.2 PFLOPS FP16 for 384 Ascend 910C cards, divided by 0.9895 PFLOP/s per H100 [1] [2] [3] [4] [5] |
| Beijing Fangshan High-Compute Vehicle Cloud CenterChina | —not installed | ~7.3k1.8k–16.5k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] |
| Beijing Yanqing AI Industry Empowerment CenterChina | —not installed | ~6.3k1.6k–14.2k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] |
| Beijing Seven Star Park AI Computing CenterChina | ~6.2k1.6k–13.9k | —no modeled buildout | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] [3] |
| Shijingshan Intelligent Computing CenterChina | —not installed | ~6.1k1.5k–13.7k | site totalaggregate conversion | China card-count / power-envelope scenario · NON-ADDITIVE100–900 H100e per disclosed facility MW; occupancy and hardware generation unknown [1] [2] [3] |
| China Telecom Zhongwei Computing HubChina | ~6k900–24k | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] [2] |
| Bell Merritt BUZZ Cohere GB200 ClusterCanada | —not installed | ~5.8k5.8k–5.8k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation2,304 GB200 GPUs × 2.5265 H100e/GPU [1] [2] [3] |
| Indosat Lintasarta Jakarta GB200 AI FactoryIndonesia | ~5.8k5.8k–5.8k | ~5.8k5.8k–5.8k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation2,304 GB200 GPUs × 2.5265 H100e/GPU [1] [2] |
| China Mobile Harbin Intelligent Computing CenterChina | ~5.2k3.5k–7k | ~5.2k3.5k–7k | 0.292×0.194–0.389× | site aggregate FP16 scenario midpointreported 6.6–6.93 EFLOPS half precision; low assumes sparse convention, high assumes dense, central is the scenario midpoint [1] [2] [3] [4] [5] |
| China Mobile Hohhot Intelligent Computing CenterChina | ~5.1k3.4k–6.8k | ~5.1k3.4k–6.8k | 0.254×0.169–0.339× | site aggregate FP16 scenario midpointreported 6.7 EFLOPS FP16; low assumes sparse convention, high assumes dense, central is the scenario midpoint [1] [2] [3] [4] |
| Kakao Data Center Ansan B200 ClusterSouth Korea | ~4.6k4.6k–4.6k | ~5.5k5.5k–5.5k | 2.274×2.274–2.274× | vendor-spec dense BF16 calculationB200 GPU count × approximately 2.274 H100e/GPU [1] [2] [3] |
| SDS-AI Sovereign AI CloudIsrael | ~4.6k4.6k–4.6k | ~4.6k4.6k–4.6k | 2.274×2.274–2.274× | vendor-spec dense BF16 calculation2,032 reported B200 GPUs multiplied by the atlas's approximately 2.274 H100e per B200 convention [1] [2] [3] |
| Cerebras / Scale Oklahoma City AI DatacenterUnited States | ~4.5k4.5k–4.5k | ~4.5k4.5k–4.5k | 15.160×15.160–15.160× | conservative dense-FP16 peak arithmetic conversionWSE-2 is 7.5 PFLOP/s dense FP16 and Cerebras states WSE-3 doubles WSE-2 performance, implying about 15 PFLOP/s dense FP16 per CS-3; divided by 0.9895 PFLOP/s dense BF16/FP16 per H100 SXM [1] [2] [3] [4] [5] |
| China Telecom–Huawei Shaoguan Ascend 10k ClusterChina | ~3.5k922–9.3k | —no modeled buildout | 0.300×0.080–0.810× | China card-count / power-envelope scenario · NON-ADDITIVEunknown Ascend-generation scenario [1] [2] [3] |
| Pengcheng CloudbrainChina | ~3.2k480–12.8k | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] [2] |
| FPT AI Factory JapanJapan | ~3k2k–6k | ~3k2k–6k | 1.000×1.000–1.000× | bounded interpretation of operator quantity languageFPT describes the Japan factory as powered by thousands of Hopper GPUs; Epoch models 3,000 and the same-size $200m project budget gives an approximate 6,000-GPU ceiling [1] [2] [3] [4] |
| FPT AI Factory VietnamVietnam | ~3k2k–6k | ~3k2k–6k | 1.000×1.000–1.000× | bounded interpretation of operator quantity languageFPT reports thousands of H100 GPUs at this factory; Epoch models 3,000 and a $200m project budget gives an approximate 6,000-GPU ceiling [1] [2] [3] [4] |
| KAUST Shaheen III GPU PartitionSaudi Arabia | ~2.8k2.8k–2.8k | ~2.8k2.8k–2.8k | 1.000×1.000–1.000× | architecture-class approximation2,800 GH200 accelerators treated as approximately 1 H100e each, consistent with the atlas's existing GH200 convention [1] [2] [3] |
| LillyPod Indianapolis B300 ClusterUnited States | ~2.6k2.6k–2.6k | ~2.6k2.6k–2.6k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation1,016 B300 GPUs × 2.5265 H100e/GPU [1] |
| China Unicom Zhongwei 10k ClusterChina | ~2.5k280–14k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] |
| BavariaAI Blue Swan at NHR@FAUGermany | —not installed | ~2.3k2.3k–2.3k | 2.274×2.274–2.274× | vendor-spec dense BF16 calculation1,024 B200 GPUs × approximately 2.274 H100e/GPU [1] [2] [3] |
| E2E Networks at L&T Vyoma ChennaiIndia | ~2.3k2.3k–2.3k | ~2.3k2.3k–2.3k | 2.274×2.274–2.274× | vendor-spec dense BF16 calculation1,024 B200 GPUs × approximately 2.274 H100e/GPU [1] [2] [3] [4] [5] |
| SiGTRON Tainan AI FactoryTaiwan | —not installed | ~2.3k2.3k–2.3k | 2.274×2.274–2.274× | vendor-spec dense BF16 calculation1,024 inferred B200 GPUs × approximately 2.274 H100e/GPU; RTX Pro nodes excluded from H100e conversion [1] |
| Sharon AI Supercluster at NEXTDC M3Australia | ~2.3k2.3k–2.3k | ~2.3k2.3k–2.3k | 2.274×2.274–2.274× | vendor-spec dense BF16 calculation1,000 physically attributed B200 GPUs multiplied by the atlas's approximately 2.274 H100e per B200 convention [1] [2] [3] |
| SK Telecom Haein Gasan B200 ClusterSouth Korea | ~2.3k2.3k–2.3k | ~2.3k2.3k–2.3k | 2.274×2.274–2.274× | vendor-spec dense BF16 calculation1,000-card disclosed lower bound × approximately 2.274 H100e/B200 [1] [2] [3] |
| Toli Green-Carbon Intelligent Computing ParkChina | ~2.2k250–12.5k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] |
| EuroHPC HammerHAIGermany | —not installed | ~2.1k2.1k–2.1k | 2.526×2.526–2.526× | vendor-spec dense BF16 calculation850-GPU disclosed lower bound × 2.5265 H100e per GB200 GPU [1] [2] |
| NCHC Nano 4Taiwan | ~2.1k2.1k–2.1k | ~2.1k2.1k–2.1k | site totalaggregate conversion | mixed-fleet vendor-spec calculation1,760 H200 at 1.0 H100e plus 144 GB200 at 2.5265 H100e [1] |
| Alibaba–China Telecom Shaoguan Zhenwu ClusterChina | ~2k500–4.5k | —no modeled buildout | 0.200×0.050–0.450× | China card-count / power-envelope scenario · NON-ADDITIVEZhenwu 810E order-of-magnitude scenario; dense throughput undisclosed [1] [2] [3] [4] |
| China Telecom Shanghai Lingang 10k ClusterChina | ~2k300–8k | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] |
| China Telecom Wuqing Domestic 10k ClusterChina | ~2k300–8k | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] [2] [3] |
| China Unicom Shanghai Lingang 10k ClusterChina | ~2k300–8k | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] [2] [3] [4] |
| China Unicom–Alibaba Sanjiangyuan ClusterChina | ~2k300–8k | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] [2] |
| Huaqing Qingyang 10k ClusterChina | ~2k300–8k | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] [2] |
| Moore Threads KUAE S5000 10k ClusterChina | ~2k500–5.5k | —no modeled buildout | 0.200×0.050–0.550× | China card-count / power-envelope scenario · NON-ADDITIVEMoore Threads generation scenario [1] [2] [3] |
| Wuxi MetaX Domestic GPU 10k ClusterChina | —not installed | ~2k500–5k | 0.200×0.050–0.500× | China card-count / power-envelope scenario · NON-ADDITIVEMetaX generation scenario; dense throughput undisclosed [1] [2] |
| iFlytek Feixing No. 1China | ~2k300–8k | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] [2] [3] |
| EuroHPC DAEDALUSGreece | —not installed | ~2k2k–2k | 1.000×1.000–1.000× | architecture-class approximationreported lower bound of 2,000 GH200 accelerators treated as approximately 1 H100e each [1] [2] [3] |
| Wuxi Huishan City Intelligent Compute CloudChina | ~2k220–11k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] |
| Chengfeng Erlai Yichang AI Compute CenterChina | —not installed | ~1.8k200–10k | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] [3] |
| China Mobile Guangdong–Hong Kong–Macao Guangzhou AI CenterChina | —not installed | ~1.8k200–10k | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] |
| China Mobile Yangtze River Delta Jiashan Intelligent Computing CenterChina | —not installed | ~1.8k200–10k | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] [3] |
| China Unicom Hohhot 10k Intelligent Computing CenterChina | ~1.8k200–10k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] |
| China Yacloud Shansi Kaiwu Ya'an Supercomputer CampusChina | —not installed | ~1.8k200–10k | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] [3] |
| SenseTime Lingang AIDCChina | ~1.8k200–10k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] |
| Shanghai Instrumentation Songjiang AI Compute CenterChina | ~1.8k200–10k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] |
| Xinjiang Jiangsuan 10k Hub Intelligent Computing CenterChina | —not installed | ~1.8k200–10k | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] |
| Huasao Data Valley Phase I GPU ClusterChina | ~1.7k190–9.5k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] |
| NSCC ASPIRE 2BSingapore | ~1.5k1.5k–1.5k | ~1.5k1.5k–1.5k | 1.000×1.000–1.000× | dense BF16 compute conventionH200 has the same dense BF16 compute throughput as H100 SXM; memory-capacity uplift is not counted [1] [2] [3] |
| EuroHPC ArrheniusSweden | ~1.5k1.5k–1.5k | ~2.5k2.5k–2.5k | site totalaggregate conversion | mixed-system vendor-spec calculation1,528 operating GH200 at 1.0 H100e plus 400 planned GB200 at 2.5265 H100e on the same Linköping campus [1] [2] [3] [4] |
| Alpha Compute ALPHA-02Sweden | —not installed | ~1.3k1.3k–1.3k | 2.274×2.274–2.274× | vendor-spec dense BF16 calculation576 B200 GPUs × approximately 2.274 H100e/GPU [1] [2] |
| China Mobile Qingyang / Yisuan Enflame ClusterChina | ~1.2k300–3.5k | —no modeled buildout | 0.120×0.030–0.350× | China card-count / power-envelope scenario · NON-ADDITIVEEnflame inference/training scenario; S60 dense throughput withheld [1] [2] [3] [4] |
| Sophgo Xiamen Torch Domestic 10k ClusterChina | —not installed | ~1.2k200–5k | 0.120×0.020–0.500× | China card-count / power-envelope scenario · NON-ADDITIVESophgo generation scenario [1] [2] |
| BUZZ Bell Manitoba B200 ClusterCanada | ~1.1k1.1k–1.1k | ~1.1k1.1k–1.1k | 2.274×2.274–2.274× | vendor-spec dense BF16 calculation504 B200 GPUs × approximately 2.274 H100e/GPU [1] [2] [3] [4] |
| Shanghai Waigaoqiao Domestic 3k ClusterChina | ~60090–2.4k | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] |
| China Mobile Guiyang Gui'an AI ClusterChina | ~54060–3k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] |
| Parallel Technology Qingdao 3k GPU Resource PoolChina | ~54060–3k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] |
| Anonymized Chinese System (Epoch record 177)China | ~36040–2k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] |
| WeRide Autonomous-Driving Training ClusterChina | ~36040–2k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] |
| China Unicom Huazhong Zhiyun Wuhan 1,024-card ClusterChina | ~317123–563 | —no modeled buildout | 0.310×0.120–0.550× | China card-count / power-envelope scenario · NON-ADDITIVE910B generation bracket anchored by disclosed site aggregates [1] [2] |
| SJTU Zhiyuan No. 1China | ~316316–316 | ~316316–316 | site totalaggregate conversion | site aggregate dense FP16 conversion313 PFLOPS reported FP16 divided by 0.9895 PFLOP/s dense FP16/BF16 per H100 SXM [1] [2] |
| China Mobile Digital Two-Asias Kunming ClusterChina | ~20030–800 | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] [2] |
| China Mobile Shandong Jinan Domestic 1k ClusterChina | ~20030–800 | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] [2] |
| China Telecom Nanjing Jishan Domestic 1k PoolChina | ~20030–800 | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] [2] |
| UCloud Shanghai Qingpu Domestic 1k ClusterChina | ~20030–800 | —no modeled buildout | 0.200×0.030–0.800× | China card-count / power-envelope scenario · NON-ADDITIVEunknown domestic-accelerator scenario [1] [2] |
| Ankang Intelligent Computing CenterChina | ~18020–1k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] |
| China Telecom Hangzhou Xiaoshan 1k PoolChina | ~18020–1k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] |
| China Unicom Shaoguan South-China Inference PoolChina | ~18020–1k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] [3] |
| GCL AI Baoshan Intelligent Computing CenterChina | ~18020–1k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] |
| Shenzhen Baiwangxin Zhonghe Huahui 1k ClusterChina | ~18020–1k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] [2] [3] |
| Shenzhen East Longgang Public AI ClusterChina | ~18020–1k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] |
| Starfire Zhongda Wuhan 1k Demonstration ClusterChina | ~18020–1k | —no modeled buildout | 0.180×0.020–1.000× | China card-count / power-envelope scenario · NON-ADDITIVEaccelerator model and generation undisclosed [1] |
| China Southern Power Grid Kunlun Domestic ClusterChina | ~15020–500 | —no modeled buildout | 0.150×0.020–0.500× | China card-count / power-envelope scenario · NON-ADDITIVEKunlunxin generation scenario [1] |
| Tongji University Domestic DCU Thousand-Card ClusterChina | ~12020–400 | —no modeled buildout | 0.120×0.020–0.400× | China card-count / power-envelope scenario · NON-ADDITIVEHygon DCU generation scenario [1] |
| Zhanjiang Domestic 1k Inference ClusterChina | —not installed | ~12030–350 | 0.120×0.030–0.350× | China card-count / power-envelope scenario · NON-ADDITIVEEnflame inference/training scenario; S60 dense throughput withheld [1] |
| Data Center Valley EkibastuzKazakhstan | ~00–0 | ~252.7k252.7k–252.7k | 2.526×2.526–2.526× | conservative disclosed-generation floor100,000 announced mixed GB300/Vera Rubin GPUs valued at the atlas's 2.5265 H100e GB300 factor; Rubin is not given a speculative uplift [1] [2] [3] |
Interpretation limits
Equivalent FLOPs are not equivalent systems.
Included
Peak dense 8-bit arithmetic where public, dense 16-bit fallback evidence where necessary, disclosed card quantities, explicitly published site totals, and uncertainty about specifications or fleet composition.
Not included
CUDA/CANN/software maturity, achieved utilization, collective-communication scaling, memory capacity and bandwidth, reliability, yield, workload mix, or wall-clock time-to-train.
Central estimates
The map needs one sortable value. Strong specification- or site-aggregate models are added to totals. Broad unknown-chip and power-envelope scenarios are still prefixed with “~” and sized on the map, but are explicitly non-additive.
Raw-card threshold
A site with ≥10,000 disclosed accelerators can remain in the default view even when its modeled H100e is lower. Conversely, roughly 3,958 GB200/GB300 GPUs exceed 10,000 H100e; the atlas therefore includes a 1,000-card inventory exposed through the scale control.
Aggregate-only sub-1k tail
Active, non-superseded Epoch systems below both full-record thresholds remain in regional, small/legacy and shipment-reference sums. They do not create map dots, ledger rows or site-count records, and their logical-cluster totals can overlap campus estimates.
Reference specifications
Primary anchors
- Epoch AI data-center methodology — dense FP8 H100e normalization, chip/power model, permits and satellite workflow.
- NVIDIA Hopper architecture — rounded 1,000/2,000 TFLOPS dense/sparse BF16 for H100 SXM.
- NVIDIA H100 product specifications — 1,979 TFLOPS BF16 with sparsity, implying 989.5 dense.
- Huawei Atlas 900 A3 disclosure — 384 Ascend 910C chips and up to 300 PFLOPS.
- Epoch AI chip-sales methodology — published 910B/910C specification ranges and conversion uncertainty.
- TSMC 2025 annual report — aggregate wafer shipments, process mix, customer and product counts; useful precisely because it shows the public data stop short of customer allocation.
- U.S. International Trade Administration, Taiwan semiconductors guide — HS 8542 scope includes processors, controllers, memory and other integrated circuits rather than an AI-GPU-only category.
- U.S. CBP bill-of-lading guide — cargo documents can contain shipper, consignee, route, weight and description, subject to access and confidentiality limits.
- U.S. BIS EAR §743 — some controlled advanced-IC transactions require quantities and designer information to be reported to government; those filings are not a public global ledger.
- NVIDIA GB200 NVL72 specifications — 360 PFLOPS sparse / 180 PFLOPS dense BF16 across 72 GPUs.