Public Kempax report

Daily Update Global Data Center and GPU capacity · 2026-09-04

Kempax Research

Loop report · 2026-09-04

Global data-center and GPU capacity — 4 September 2026

As of 4 September 2026 UTC, the accepted baseline and prior dated updates are retained. This run adds five source-qualified site or project records, a US development-cost benchmark, five accelerator catalog checkpoints, and like-for-like IT-load and hosted-rate comparisons. The new evidence covers Hut 8’s 704 MW contracted IT campus against a 1,000 MW utility envelope Hut 8 , Chevron’s 2.67 GW generation plan with a 20-year Microsoft PPA Chevron , an exploratory Egyptian AI-hub discussion MCIT , Macquarie’s under-construction 47 MW IT facility Macquarie , and RT-One’s 100 MW first phase with expansion up to 400 MW Uberlândia . The figures use different products, states, economic objects, and capacity boundaries. No global data-center, electricity, installed-GPU, investment, or comparable-price total is asserted or summed.

Current-run update — 2 September 2026

This additive update keeps the accepted baseline intact and records the fresh evidence packet below. The nine data-center and power records and two accelerator catalog records are selected public signals, not a global census. ND means not disclosed or not verified in the reviewed evidence, not zero. IT load, facility or campus power, grid interconnection, generation PPA, cloud orderability, installed GPU inventory, and economic objects remain separate; no unlike or overlapping measures are summed.

Data-center plans, construction, power, and ownership

Record location Source date and state Capacity and electricity basis GPU inventory specification Company, developer, operator, investor or financing role Price economic object Direct evidence --- --- --- --- --- --- --- DC-01 · Cerebras Compute Nordic — Mikkeli, Finland 1 Sep 2026 release dateline; initial 50 MW phase reported under construction; contract not reported energized 165 MW contracted IT capacity delivered in phases of 50 MW, 80 MW, and 165 MW; seven-year service orders; grid interconnection, generation PPA, and annual-energy basis ND No GPU count or model; source describes Cerebras wafer-scale AI compute, not a GPU inventory Cerebras is long-term customer platform; Compute Nordic Finland is development partner; project equity, lender, and investor roles ND €1.0–1.7B indicative regional investment at full scale; estimate, not stated committed capex, rent, or GPU price Cerebras IR release — page index labels 31 Aug while release dateline says 1 Sep; discrepancy retained DC-02 · Bitdeer AI A102 — Malaysia 31 Aug page label 1 Sep release dateline; sold out under offtake commitments ahead of Q1 2027 energization ; not reported operating Approximately 9.5 MW data-center capacity ; IT boundary not stated. Separate Bitdeer target of up to 350 MW AI-cloud capacity by Q1 2028 is not additive; grid interconnection, generation PPA, and tariff ND Planned rack-scale NVIDIA GB300 NVL72 deployment; installed, delivered, ordered, and committed count ND Bitdeer AI, part of Bitdeer Technologies Group, is AI-cloud provider developer; customers unnamed. More than 50% of associated capex is expected from customer prepayments; other financing names ND More than US$800M total expected revenue over five-year offtake commitments; contract service revenue expected from Q1 2027, not capex, rent, or hardware price Bitdeer A102 release DC-03 · Bitdeer new parcel Rockdale — Milam County, Texas, US Fee-simple greenfield acquisition completed and announced 1 Sep 2026; no construction or energization stated 200 acres acquired. Company context cites about 563 MW interconnected grid capacity and plans expected to scale to 742 MW; post-transaction statement cites about 255 acres and 742 MW existing pipeline power in the county. These are portfolio context, not capacity assigned to the new parcel No GPU count, delivery state, or model; AI HPC development intent only Bitdeer is owner, developer, and infrastructure operator; paid in cash; external investor, lender, and project-finance roles ND Approximately US$100M cash land-acquisition consideration ; land transaction value, not data-center capex, lease value, tariff, or GPU price Bitdeer land-acquisition release DC-04 · Digital Vault Sulcis — Nuraxi Figus, Sardinia, Italy National-strategic designation announced 1 Sep 2026; accelerated administrative pathway; not a permit, construction, energization, or operating claim Initial Edge AI phase up to approximately 30 MW over 24 months , scaling toward 50–60 MW IT and up to 100 MW total site power ; renewable generation and storage integrated; utility import limit, generation nameplate, PPA, and annual energy ND GPU count, model, and allocation ND Energy Vault is developer and Build, Own & Operate platform; Italian government designation and Autonomous Region of Sardinia support; project equity, lender, and financing roles ND Project capex and financing ND; no price inferred from the designation BusinessWire-syndicated release · Energy Vault newsroom — direct BusinessWire page unavailable in the evidence packet DC-05 · Fervo Energy Google Cape Station — Utah, US PPA announced 1 Sep 2026; Cape Station EGS GeoCluster expected online in 2028; potential data-center plans remain subject to approvals and commercial conditions 396 MW geothermal PPA is generation offtake capacity, not data-center IT load. Google has an option for approximately 600 MW additional offtake, nearly 1 GW by June 2030; option is not committed current capacity. Grid interconnection and site IT load ND No GPU count or model; no operating data-center inventory claimed Fervo is geothermal developer generator; Google is offtaker and potential data-center anchor; investor, lender, and project-finance roles ND PPA price, project capex, tariff, and colocation cloud price ND Fervo IR release DC-06 · FuelCell Energy unnamed operator — Texas, US First capacity-reservation agreement reported 2 Sep 2026; planned project; no construction, energization, operator identity, or delivery date stated 75 MW planned project ; source does not specify whether this is IT load, generation, or total facility capacity. Grid interconnection, PPA, and tariff ND GPU count and model ND FuelCell Energy is supplier; major data-center operator is unnamed; investor, lender, and project-finance roles ND Upfront reservation payment exists, but amount and terms are ND; reservation signal, not project capex or GPU price FuelCell Energy Q3 2026 release DC-07 · FuelCell Energy Fit Energy — US data-center power supply Agreement reported 2 Sep 2026; four potential phases; not operating or energized 30 MW committed fuel-cell generation plus an option for up to 350 MW additional generation ; source describes up to 380 MW across both. Generation is not IT load; site allocation, grid, and PPA ND No GPU count or model; end data-center operator unnamed FuelCell Energy is fuel-cell system supplier; Fit Energy USA LP is purchaser option holder; lender and equity-investor roles ND US$1.3B committed backlog includes the 30 MW commitment; US$2.4B awarded-capacity backlog relates to the 350 MW option. Backlog is not project capex, rent, or hardware price FuelCell Energy Q3 2026 release DC-08 · Hyperscale Data Michigan facility — US Bitcoin mining ceased effective 1 Sep 2026 ; facility is in AI engineering and equipment-procurement preparation; AI energization and operation not reported Initial MSA provides 20 MW critical AI compute capacity and a right to an additional 32 MW . Company describes approximately 340 MW eventual facility power capacity ; this is not current IT load and is not added to the MSA figures GPU count, model, delivery state, and allocation ND; California neocloud customer unnamed Hyperscale Data is facility owner operator and AI-conversion developer; California neocloud is customer; investor, lender, and equipment-finance roles ND More than US$1.2B expected revenue for the initial 10-year MSA plus options; more than US$3.0B at maximum term and option scope. Expected service lease revenue, not capex or GPU price Hyperscale Data release DC-09 · KIDZ AI HUADING national cloud data center — Phnom Penh, Cambodia Company-distributed release dated 2 Sep 2026 reports the Tier III facility operational and energized; KIDZ transaction remains subject to customary closing conditions Pending transaction includes a three-year right of first refusal to lease up to 1 MW of data-center load ; this is contractual access, not current leased or installed load. Exact grid MW, MVA, generation, and PPA ND AI GPU deployments described generically; GPU count, model, and installed state ND KIDZ AI is proposed 2.0% minority equity holder and strategic partner; HUADING DEVELOPMENT CAMBODIA CO., LTD. is owner operator; Cambodian government national-cloud designation; consideration, lender, and other investor roles ND Acquisition consideration, lease rent, capex, tariff, and GPU price ND KIDZ AI company-distributed release

GPU inventory, model specifications, and current price checkpoint

Record Quantity state Model specification Company, vendor, investor or financing role Possible price with original unit, type, and date Direct evidence --- --- --- --- --- --- GPU-01 · NVIDIA GB300 NVL72 NVIDIA lists the platform as Available Now ; Bitdeer A102 identifies planned rack-scale deployments. No installed, delivered, ordered, committed, or allocated site count is disclosed Fully liquid-cooled rack-scale platform with 72 NVIDIA Blackwell Ultra GPUs and 36 Arm-based NVIDIA Grace CPUs ; NVIDIA also describes 800 Gb s networking per GPU. Vendor specification is not a project inventory count NVIDIA is OEM platform provider; Bitdeer A102 is intended deployment context; purchaser, investor, and financing roles ND OEM purchase price and rental price ND; no unit-price date verified NVIDIA GB300 NVL72 · Bitdeer A102 GPU-02 · Lambda on-demand catalog Current catalog offers 1x, 2x, 4x, and 8x GPU instances; 1-Click Clusters offer 16–2,000+ interconnected B200 or H100 GPUs . This is orderability offer capacity, not a deployed fleet count NVIDIA B200 SXM6: 180 GB VRAM GPU ; NVIDIA H100 SXM: 80 GB VRAM GPU ; catalog also lists A100 and V100 Lambda is cloud operator; NVIDIA is OEM; investor, lender, and project-allocation roles ND US$6.69 per B200 GPU-hour and US$3.99 per H100 GPU-hour ; hosted rental list rates, catalog captured 2 Sep 2026, no page publication date; applicable sales tax VAT GST may apply Lambda GPU instances · NVIDIA B200 · NVIDIA H100

The current packet does not verify an OEM purchase price or executed hardware transaction price for GB300, B200, H100, H200, Rubin, or MI350X. Cloud rates are not hardware prices, installed-inventory counts, electricity tariffs, or project capex.

Economic and price signals

Record Value currency Unit Economic type Date Boundary and non-additivity control Direct evidence --- --- --- --- --- --- --- Cerebras Compute Nordic Mikkeli €1.0–1.7B Total regional investment estimate at full scale Indicative investment estimate 1 Sep 2026 release dateline; captured 2 Sep Not stated committed project capex, rent, or GPU price Cerebras IR release Bitdeer A102 Malaysia More than US$800M Total expected revenue over five years Contracted AI-cloud offtake service revenue 1 Sep 2026 release dateline; revenue expected from Q1 2027 Not capex, colocation rent, tariff, or hardware price; not added to the separate 350 MW target Bitdeer A102 release Bitdeer Rockdale land acquisition Approximately US$100M Cash land-acquisition consideration Completed land transaction 1 Sep 2026 announcement Not data-center construction capex or GPU price; parcel-specific capacity ND Bitdeer land-acquisition release FuelCell Energy Fit Energy commitment US$1.3B Committed backlog associated with 30 MW generation Company backlog economic signal 2 Sep 2026 Q3 release Backlog is not project capex, rent, tariff, or hardware price FuelCell Energy Q3 2026 release FuelCell Energy Fit Energy option US$2.4B Awarded-capacity backlog associated with up to 350 MW optional generation Company backlog economic signal 2 Sep 2026 Q3 release Option is not committed generation and backlog is not project capex or GPU price FuelCell Energy Q3 2026 release Hyperscale Data Michigan MSA More than US$1.2B Expected revenue over 10-year MSA plus options Expected AI data-center service lease revenue 1 Sep 2026 release Not current revenue, facility capex, or GPU price; 32 MW option and 340 MW eventual power are not summed Hyperscale Data release Hyperscale Data maximum MSA scope More than US$3.0B Expected revenue including 32 MW option and maximum term Expected AI data-center service lease revenue 1 Sep 2026 release Not capex, realized revenue, or GPU price; it is an expanded contractual scenario Hyperscale Data release Lambda B200 catalog US$6.69 Per GPU-hour Hosted rental list rate Catalog capture 2 Sep 2026; page publication date ND Dynamic provider rate; taxes may apply; not hardware purchase price or project economics Lambda GPU instances Lambda H100 catalog US$3.99 Per GPU-hour Hosted rental list rate Catalog capture 2 Sep 2026; page publication date ND Dynamic provider rate; taxes may apply; not hardware purchase price or project economics Lambda GPU instances Digital Vault Sulcis ND Project capex and financing No accepted price verified 1 Sep 2026 designation No price inferred from an approval pathway BusinessWire-syndicated release Fervo Energy Google Cape Station ND PPA price, project capex, and tariff No accepted price verified 1 Sep 2026 PPA announcement 396 MW is generation offtake capacity, not an IT-load price basis Fervo IR release

Methodology and figure discipline

The immutable accepted report baseline at generation 414 and commit e8de9024815ca6ac46576a5d56ab2ae04135358e was read before editing. Its complete kempax-ui block remains byte-for-byte unchanged; the current-run material is appended outside it.
The current evidence packet was captured on 2 September 2026 UTC. Direct company, investor-relations, vendor, and distributed-release URLs are kept beside the relevant claims. The Energy Vault row identifies its syndicated-source limitation; the KIDZ row identifies its company-distributed release.
Capacity is kept in the source-defined boundary: contracted IT, critical AI compute, data-center capacity, facility site power, grid interconnection, generation, and PPA are not substituted. No MVA-to-MW or energy-to-power conversion is made.
GPU states remain distinct: available-now product, planned deployment, catalog orderability, installed, delivered, ordered, committed, allocated, and potential are not interchangeable. A rack specification is not a site inventory count.
Economic objects remain distinct: investment estimate, land consideration, backlog, expected service revenue, reservation payment, PPA price, cloud rental, and hardware purchase price are not combined. Original currency, unit, type, and source date are retained; no FX conversion is made.
Parent targets, child sites, options, and portfolio context are not added. Examples include Bitdeer’s 350 MW target, Milam County’s 742 MW context, Fervo’s additional 600 MW option, Hyperscale’s 340 MW eventual facility potential, and Fit Energy’s optional generation.
ND means not disclosed or not verified in the reviewed evidence; it does not mean zero. Company claims remain source-qualified and are not promoted to independent audits.

Uncertainty and unresolved risks

Risk Retained treatment Evidence and implication --- --- --- Date-boundary conflict Cerebras and Bitdeer A102 retain both page index and release-dateline dates Cerebras and Bitdeer A102; do not silently backdate or forward-date milestones Status risk Announced, contracted, land-acquired, designated, PPA, and AI-preparation states are not upgraded to construction, energization, operation, or delivery Energy Vault, Fervo Google, and Hyperscale Data show why status is material Boundary risk 75 MW FuelCell project basis and 9.5 MW Bitdeer A102 IT boundary remain ND; generation PPA is not IT load FuelCell Energy, Bitdeer A102, and Fervo Google support separate treatment GPU inventory risk GB300 rack specifications and Lambda cluster offers do not establish installed, delivered, ordered, or globally available site stock NVIDIA GB300 and Lambda; no current project GPU count is manufactured Ownership and financing risk Customers, project investors, lenders, and counterparties remain unnamed where the source does not name them; supplier, government, and offtaker roles are not relabeled as equity ownership Bitdeer A102, FuelCell Energy, and KIDZ AI Source-access risk The Energy Vault claim uses a BusinessWire-labeled syndication because the direct page was unavailable; KIDZ uses a company-distributed Newswire release Syndicated Energy Vault release and KIDZ release; neither is treated as an independent facility audit Price risk Lambda rates are provider-listed rental snapshots; no comparable OEM hardware quote, project electricity tariff, PPA price, or GPU-cost-per-MW estimate is verified Lambda, Fervo Google, and NVIDIA GB300 Census risk The current-run records are a bounded selection and do not support a global total The report retains the selected rows and baseline context without adding unlike or overlapping measures

Milestone watchlist

Watch item Verification target Direct evidence --- --- --- Finland and Malaysia commissioning Confirm construction progress, energization, first service, and the date discrepancy for the 165 MW Finland phases and 9.5 MW A102 site Cerebras Compute Nordic · Bitdeer A102 Rockdale parcel Obtain parcel-specific permits, interconnection allocation, and construction milestone; do not reuse Milam County portfolio power as parcel capacity Bitdeer land acquisition Sulcis and Cape Station Check authorization, site-load definition, PPA economics, and actual generation data-center coupling Energy Vault · Fervo Google FuelCell agreements Verify the 75 MW boundary, named operator, reservation-payment terms, Fit Energy phase schedule, and delivered generation FuelCell Energy Q3 2026 release Michigan AI conversion Confirm MSA load actually energized, customer identity when disclosed, equipment delivery, and whether any of the 340 MW eventual potential is contracted Hyperscale Data release Cambodia transaction Check closing of the 2.0% minority stake and whether any portion of the 1 MW lease right is executed or energized KIDZ AI release GPU and price validation Recheck GB300 deployment counts and Lambda rates; obtain a named OEM or executed procurement quote before treating any price as hardware cost NVIDIA GB300 · Lambda

Current-run update - 3 September 2026

This additive update preserves the accepted baseline and appends the latest evidence packet. Four new data-center, power, or manufacturing records, nine accelerator and inventory records, normalized product specifications, and current provider-rate checkpoints are shown below. ND means not disclosed or not verified in the reviewed evidence, not zero. IT load, facility or campus power, grid or interconnection, generation or PPA, manufacturing throughput, cloud orderability, installed inventory, and economic objects remain separate. No unlike or overlapping measures are summed.

New data-center, power, and manufacturing signals

Record location Source date and state Capacity and electricity boundary Company, developer, operator, investor or financing role Price economic object Direct evidence --- --- --- --- --- --- DC-12 · Linea Energy Google Duffy Solar - Matagorda County, Texas, US Fifteen-year PPA announced May 2026; solar construction scheduled for Q3 2026; co-located battery storage reported under construction Site IT load ND. 500 MW solar PPA and 235 MWac Duffy BESS are generation and storage values, not site IT load or interconnection capacity Linea develops, owns, and operates clean-energy infrastructure; EnCap Investments and EnCap Energy Transition sponsor Linea; Google is offtaker; project financing ND PPA price and project capex ND Linea Google Duffy PPA DC-16 · Desert Technologies Swiss CTS Group CTS-DT - Saudi Arabia Joint venture announced 2 Sep 2026 at LEAP; no construction, energization, or operating site evidence 300 MW year initial manufacturing throughput for advanced AI data-center solutions, scaling toward 1.2 GW year . This is manufacturing output, not installed data-center capacity. Site IT, grid, generation, PPA, and tariff ND Desert Technologies and Swiss CTS Group establish CTS-DT; JV ownership split, investors, lenders, and customer or site allocation ND Capex, financing, contract value, and price ND MENAFN Mid-East Info · CTS-DT DC-17 · Kasi Cloud LOS1 Lekki campus - Lagos, Nigeria Commissioning and operational-readiness ceremony reported 19 May 2026; staged campus; full current operating state ND Approximately 100 MW critical IT capacity at full development ; LOS1 phase load ND. Grid, interconnection, generation, and PPA MW ND Kasi Cloud Datacenters develops and operates; government and NSIA participation is reported around commissioning; equity and lender allocation ND Project capex, financing, lease, tariff, and GPU price ND. The release’s US$850M annual offshore-cloud-spend figure is market context, not project capital Kasi commissioning release DC-18 · iXAfrica NBOX1 campus - Nairobi, Kenya NBOX1.1 reported live; NBOX1.2 is the planned full-site phase; page captured 3 Sep 2026 4.5 MW live IT power and 22.5 MW full-site IT power are nested. Initial feeders are 2 x 4 MVA; planned substation is 2 x 40 MVA; MVA is retained without conversion iXAfrica operates; Future-tech provided design expertise; Helios Investment Partners supplied earlier growth capital; site financing allocation ND US$50M growth-capital signal from 2022; not current site capex or GPU price iXAfrica campus specifications

Accelerator inventory, state, and company roles

Record Quantity and state Capacity or electricity boundary Company, operator, OEM, investor or financing role Price signal and control Direct evidence --- --- --- --- --- --- GI-01 · xAI Colossus - United States 200,000 H100 GPUs in one interconnected cluster; operating, first-party claim captured 3 Sep 2026. A one-million-GPU roadmap is future context, not installed stock MW and IT load ND on the current page xAI SpaceXAI operator; NVIDIA OEM; investor and financier roles ND on the cited page Price ND; do not treat the company claim as an independent audit xAI Colossus GI-02 · AWS Project Rainier Nearly 500,000 Trainium2 chips in service; operating, first-party AWS claim. Trainium2 is a custom accelerator, not an NVIDIA GPU Site power ND AWS Amazon operator and builder; Anthropic collaborator and user; investor and financier roles ND AWS cites 30-40% better price-performance than certain GPU instances, but no comparable dollar rate; exclude from NVIDIA GPU totals AWS Project Rainier GI-03 · Yotta D2 - Greater Noida, India 20,736 NVIDIA B300 GPUs ; committed or planned. Later release targets delivery and completion by 30 Sep 2026 ; completion was not verified at capture D2 60 MW , scalable to 250 MW ; separate Navi Mumbai scale signal is not assigned to D2 Yotta operator; NVIDIA DGX Cloud relationship; Gorilla commercial and infrastructure partner; Hiranandani Group backing or leadership relationship indicated by Yotta; equity split and financing allocation ND Approx. US$2.8B aggregate commercial framework, not a card price or proof of completion Yotta Gorilla GI-04 · Noetra national AI infrastructure - Japan 27,500 Rubin GPUs plus 13,750 Vera CPUs ; announced and planned, not verified installed 140 MW data-center capacity ; not stated as per-GPU IT load Noetra operator and establisher; NVIDIA OEM and architect; METI support; investor and financing amount ND OEM price ND; architecture quantity is not live inventory NVIDIA Japan GI-05 · DataSection AI data center - Thailand 4,696 B200 GPUs in 587 server systems ; acquired or contracted, with planned Jul-Aug 2026 delivery; live status ND Site power and capacity ND DataSection operator and buyer; ASRock Rack server OEM; unnamed US customer; customer advances, scheduled loans, and share-subscription rights are disclosed; external investor allocation not verified Estimated fixed-asset acquisition US$257.4M approximately ¥41.13B for server systems, not card-only price DataSection count · JapanIR disclosure mirror · BigGo disclosure mirror GI-06 · National supercomputer access program - Israel 1,000 NVIDIA B200 accelerators over coming years; allocated program capacity, not necessarily cards installed at one site Site power and capacity ND Nebius provider and operator selected by competitive tender; Israel Innovation Authority and Telem government program; investor role ND Discounted versus market is stated; dollar price ND Israel Innovation Authority GI-07 · Stargate Norway - Nscale Aker 100,000 NVIDIA GPUs targeted by end-2026 ; a separate agreement adds more than 30,000 Rubin GPUs for Microsoft in 2027. Targets and commitments are not added or treated as installed 230 MW initial planned capacity, with ambition for another 290 MW ; not labelled as IT load Nscale designer and builder; planned 50 50 Nscale-Aker JV subject to closing; OpenAI initial offtaker; Microsoft later customer; NVIDIA OEM Approximately US$1B initial 20 MW commitment including more than US$250M equity; project financing, not GPU price Nscale Norway · Nscale Microsoft GI-08 · Cassava AI Factories - South Africa and planned African expansion NVIDIA-powered AI factories confirmed in South Africa with planned expansion to Nigeria, Kenya, Egypt, and Morocco; model, count, delivery state, and site allocation ND MW and site capacity ND in the current official release Cassava Technologies operator; NVIDIA platform and OEM relationship; project investor and financing allocation ND Price ND; secondary quantity or investment claims are not promoted without first-party confirmation Cassava AI factory announcement GI-09 · Stargate UAE - G42 OpenAI Oracle NVIDIA Grace Blackwell GB300 systems; GPU count ND; announced and planned 1 GW cluster , with 200 MW expected live in 2026 ; source boundary is not an installed GPU or IT-load count G42 builds; OpenAI and Oracle operate; NVIDIA supplies systems; Cisco and SoftBank are named partners; financing amount ND Price ND; official OpenAI page body was unavailable in the current capture, so no GPU count is inferred G42 announcement · OpenAI Stargate UAE

Normalized accelerator and system specifications

Product specifications describe model or system capability. They do not establish how many units an operator owns, has installed, or can deliver to a named site.

Vendor model Published specification Form factor, power, or interconnect Availability and role Price signal Direct evidence --- --- --- --- --- --- NVIDIA H100 SXM PCIe Hopper; 80 GB HBM per GPU. SXM 3 TB s memory bandwidth; PCIe 2 TB s SXM max TDP 700 W ; PCIe max TDP 350 W . SXM NVLink 900 GB s ; PCIe Gen5 NVIDIA OEM; product availability is not physical inventory; operator and investor ND at product level OEM MSRP ND; hosted rental rates appear below H100 product page · H100 datasheet NVIDIA H200 Hopper; 141 GB HBM3e ; 4.8 TB s memory bandwidth; datasheet cites 4 PFLOPS FP8 NVIDIA describes the same power profile as H100; exact variant TDP ND in the retrieved source NVIDIA says available; availability is not a site inventory count OEM MSRP ND; RunPod rental signal below H200 product page · H200 datasheet NVIDIA DGX B200 Eight Blackwell GPUs; 1,440 GB total GPU memory; 64 TB s HBM3e bandwidth; 72 PFLOPS training and 144 PFLOPS inference Approximately 14.3 kW maximum system power; 14.4 TB s aggregate NVLink; 10 RU. These are system values NVIDIA OEM and system vendor; no physical deployment count OEM system price ND; cloud rental signals below DGX B200 · DGX B200 datasheet NVIDIA B300 GB300 Blackwell Ultra B300 GPU: 279 GB HBM3E , up to 20 PFLOPS FP4 sparse inference. GB300 NVL72: 72 GPUs plus 36 Grace CPUs , up to 37 TB high-speed memory and up to 1.44 exaFLOPS Fully liquid-cooled rack; 72-GPU NVLink domain; 800 Gb s networking per GPU. DGX B300 is 8 GPUs, 2.1 TB total memory, approximately 14 kW NVIDIA pages describe availability or shipping; not installed inventory OEM price ND; RunPod and project values below are different economic objects GB300 NVL72 · Blackwell Ultra datasheet · DGX B300 NVIDIA Vera Rubin DGX Vera Rubin NVL72 72 Rubin GPUs and 36 Vera CPUs , 75 TB total fast memory, up to 3,600 PFLOPS NVFP4 inference and 2,520 PFLOPS NVFP4 training 100% liquid-cooled rack; sixth-generation NVLink; 72 GPUs in one NVLink domain; per-GPU HBM and TDP ND NVIDIA production-ramp signal; Japan quantity is a planned deployment OEM price ND Vera Rubin product page · Vera Rubin datasheet AMD Instinct MI300X CDNA3; 304 CUs , 192 GB HBM3 , 5.3 TB s peak theoretical bandwidth OAM accelerator; exact board power ND on current page AMD OEM; no public physical-inventory count; operator and investor ND OEM MSRP ND AMD MI300 family AMD Instinct MI325X CDNA3; 304 CUs , 256 GB HBM3E , 6 TB s . Eight-module platform: 2 TB total HBM3E and 48 TB s aggregate bandwidth OAM platform specification; exact board TDP ND AMD product availability is not a deployment count; operator and investor ND OEM MSRP ND AMD MI300 family AMD Instinct MI350X CDNA4; 288 GB HBM3E , 8 TB s ; AMD cites 9.2 PFLOPS FP8 with sparsity and MXFP6 MXFP4 support OAM; exact board TDP ND AMD OEM; no public physical-inventory count; operator and investor ND OEM MSRP ND AMD MI350X AWS Trainium2 Trn2 UltraServer Custom AWS accelerator, not a GPU. Trn2 instance has 16 chips ; UltraServer has 64 chips , up to 6 TB HBM , 83.2 PFLOPS , and 185 TB s bandwidth NeuronLink across four Trn2 instances; AWS says UltraServers are available now AWS Amazon operator and designer; Anthropic is Project Rainier collaborator and user AWS cites relative price-performance, not a comparable dollar quote Trn2 documentation · Project Rainier Google Ironwood TPU7x Custom Google ASIC, not a GPU. Per chip: 2,307 TFLOPS BF16 , 4,614 TFLOPS FP8 , 192 GiB HBM , 7,380 GB s HBM bandwidth, 1,200 GB s bidirectional ICI Up to 9,216 chips per pod configuration; 4-chip VM Google operator and designer; documentation exposes configurations, not installed stock Cloud price ND TPU7x documentation · Ironwood architecture

Current price signals with original units and dates

Record Original value and unit Economic type Date boundary Direct evidence --- --- --- --- --- Lambda rate card H100 SXM 80 GB US$3.99 GPU-hour ; B200 SXM6 180 GB US$6.69 GPU-hour Hosted provider rental list rate Capture 3 Sep 2026; taxes may apply; not hardware purchase price Lambda pricing RunPod Pods rate card H100 PCIe US$2.89 GPU-hour ; H100 SXM US$3.29 GPU-hour ; H200 US$4.59 GPU-hour ; B200 US$6.79 GPU-hour ; B300 US$7.89 GPU-hour Hosted provider rental; community and secure-cloud plans differ Capture 3 Sep 2026; region, plan, and availability can change; not card purchase price RunPod pricing CoreWeave North America rate card HGX H100 US$49.24 , H200 US$50.44 , B200 US$68.80 per 8-GPU instance-hour Hosted provider rental in instance-hour units Capture 3 Sep 2026; not directly comparable with per-GPU-hour rows without explicit normalization CoreWeave pricing DataSection Thailand acquisition Estimated US$257.4M approximately ¥41.13B for 587 server systems containing 4,696 B200 GPUs Fixed-asset acquisition estimate for server systems 3-4 Jun 2026 disclosure; not B200 card-only price and live deployment ND DataSection · TDnet mirror Yotta Gorilla India framework Approximately US$2.8B Aggregate commercial framework or offtake value 30 Apr 2026; 20,736 B300 delivery target, not GPU price or proof of completion Yotta Gorilla OEM price check ND for H100, H200, B200, B300, GB300, Rubin, AMD, AWS, and Google products in the reviewed sources No accepted OEM MSRP or list price Capture 3 Sep 2026; do not back-solve from rental rates, financing, or project values NVIDIA · AMD

Evidence risks retained

Risk Retained treatment Evidence and implication --- --- --- Global census risk The current packet supports selected source-qualified records only. No global MW, electricity, installed-GPU, investment, or comparable-price total is asserted The quantity chart mixes operating claims, targets, allocations, and contracted or acquired records; bars are not additive State and timing risk xAI’s 200,000 H100 value is a current first-party claim; Yotta’s 20,736 B300 target remains unverified as complete at the 3 Sep capture; Nscale’s 100,000 figure remains a target xAI · Yotta · Nscale Boundary risk Manufacturing throughput, generation, storage, MVA, grid power, facility power, IT load, and custom-accelerator chip counts remain separate CTS-DT · Linea Google · iXAfrica Source-access risk CTS-DT capacity is retained as secondary-source evidence because the official landing page did not expose the announcement body. DataSection monetary terms use disclosure mirrors while its count is anchored to the official press page. UAE GPU count remains ND because the official OpenAI page body was unavailable CTS-DT · DataSection · OpenAI Stargate UAE Price risk Provider rates are dynamic hosted rentals; DataSection and Yotta values cover systems or commercial frameworks; no OEM hardware price was verified Lambda · RunPod · CoreWeave Custom accelerator risk AWS Trainium2 and Google Ironwood are included for context but excluded from NVIDIA GPU quantities and totals AWS Trn2 · Google TPU7x

Current-run update — 4 September 2026

This additive update preserves every preceding section and the established report presentation. It adds only the current-run evidence below. Source-defined figures are non-additive unless explicitly comparable: parent programs, phases, options, IT load, facility or campus power, grid interconnection, generation PPA, MVA, manufacturing output, financing, rental rates, and expected revenue are kept separate. ND means not disclosed or not verified in the reviewed source, not zero. A rack specification is not site inventory; cloud rental, backlog, transaction value, PPA, tariff, and expected service revenue are not hardware price or project capex.

Global data-center plans and construction register — current-run delta

The five current-run records are organized by region and stage. The fields intentionally distinguish possible or announced capacity, permitted status, construction, operation, IT load, facility or campus power, grid interconnection, generation PPA, and MVA.

Region record Stage and status date Possible announced capacity Permitted approved Under construction Operating energized IT load Facility campus power Grid interconnection Generation PPA MVA Developer, operator, company roles Explicit investor financing role Capital possible price GPU inventory Direct evidence --- --- --- --- --- --- --- --- --- --- --- --- --- --- --- --- North America · DC-10 Hut 8 Beacon Point — Nueces County, Texas, US Second phase commercialized by lease in a 20 Jul 2026 release; Phase 1 construction commenced and was financed with investment-grade debt; Phase 2 data-hall delivery expected Q2 2028; no operating or energized state claimed in the July release 352 MW IT second phase; 704 MW contracted IT after Phase 1 and 2; 1,000 MW utility envelope . These are separate, non-additive boundaries ND as a separate permit record Phase 1 construction commenced per company; Phase 2 construction status ND in the July release ND; first-phase release stated initial energization target Q1 2027, with no current confirmation read 352 MW second phase and 704 MW contracted campus IT 1,000 MW campus utility capacity; second AI factory supported by about 500 MW utility capacity; not IT load 1,000 MW utility capacity secured under an interconnection agreement with AEP Texas ; current delivery confirmation ND ND; utility capacity is not generation ND Hut 8 developer operator and infrastructure platform; confidential high-investment-grade tenant; AEP Texas utility interconnection role; NVIDIA DSX design reference only Phase 1 investment-grade debt reported; lender, equity investor, and project-finance allocation ND Company-reported $9.8B 15-year lease for the second phase; $19.6B campus base-term value and potential $50.2B renewal value. Lease contract value, not capex, tariff, or GPU price Count, model allocation, delivery, and operation ND; NVIDIA DSX is architecture intent, not inventory Hut 8 second-phase release — source 20 Jul 2026, captured 4 Sep 2026; first-phase release — source 6 May 2026, captured 4 Sep 2026 North America · DC-11 Chevron Project Kilby — West Texas, US Planned co-located power facility for a Microsoft-operated data center; 22 Jun 2026 release describes a milestone toward FID expected by end-2026, subject to conditions, and first power anticipated in 2028; no construction or energization claim Approximately 2.67 GW generation capacity in phased modular delivery; data-center capacity and campus envelope ND FID not yet reported; permit and approval status ND No construction claim No operating or energized claim; first power anticipated 2028 ND Data-center facility and campus power ND; 2.67 GW is not relabeled as facility or IT power Interconnection agreement, grid import limit, and utility delivery capacity ND; co-location is intended to reduce regional-grid impact 20-year PPA between Energy Forge One LLC, a wholly owned Chevron subsidiary, and Microsoft for dedicated electricity; PPA price ND; GE Vernova and Solar Turbines supply generation equipment ND Energy Forge One Chevron power developer and supplier; Microsoft data-center operator and offtaker; Chevron and Engine No. 1 collaborators; GE Vernova and Solar Turbines equipment suppliers No named lender, equity investor, or project-finance structure; Chevron’s targeted mid-teen return is a company economic claim, not a financing term Mid-teen return target; capex, tariff, lease, and PPA price ND Count, model, state, and allocation ND Chevron Project Kilby release — source 22 Jun 2026, captured 4 Sep 2026 Africa · DC-12 Heka Data US consortium AI hub — Egypt Official Ministry of Communications and Information Technology release dated 3 Sep 2026 reports discussion of initial scope, scale, and implementation phases; further studies and procedures remain; exploratory proposal, not a build Initial scope, scale, phases, site, and possible capacity ND No permit or approval claim No construction claim No energization or operating claim ND IT, facility, and campus power ND MCIT references integrating digital infrastructure with needed power; grid connection, import limit, utility, and interconnection capacity ND Generation, PPA, storage, and tariff ND ND Heka Data named as developer; US consortium unnamed; MCIT, NTRA, and ITIDA officials are government regulatory participants; operator and customer ND No named investor, lender, equity commitment, or project-finance role; the US consortium is not treated as an investor absent disclosure Investment, capex, lease, service revenue, tariff, and hardware price ND Count, model, state, and allocation ND Egyptian MCIT press release — source 3 Sep 2026, captured 4 Sep 2026 Oceania · DC-13 Macquarie Data Centres IC3 Super West — Sydney, Australia Current official pages describe the facility as under construction , with end-state power secured and service advertised for 2026; pages are undated and captured 4 Sep 2026 Project page describes 65 MW potential campus IT load ; facility page body states 47 MW IT load while its header displays “45MW”; discrepancy retained State Significant Development Application in hand; construction certificates listed; approval evidence is not grid delivery Yes , per current project facility pages No operating or energized claim 47 MW IT load per body text; 45 MW header discrepancy; 65 MW potential campus IT load Entire end-state power secured, but MW, MVA, and annual-energy basis ND Utility, grid connection, and interconnection agreement ND; development approval and certificates do not establish power delivery Generation, PPA, storage, and tariff ND ND; page does not provide a convertible electrical rating Macquarie Data Centres developer operator; customers and offtakers ND; design supports hyperscale, cloud, government, enterprise, and AI use Investor, lender, and project-finance role ND on directly read pages Capex, land, lease, service price, tariff, and GPU price ND Count, model, delivery, and allocation ND; cooling modes are described but do not establish inventory IC3 Super West facility page and project page — current pages captured 4 Sep 2026 Latin America · DC-14 RT-One AI data center — Uberlândia, Minas Gerais, Brazil Municipality update dated 23 Feb 2026 says architectural review was the final local step before works and later operation; environmental and other licenses may remain; RT-One company page dated 11 Oct 2025; not reported constructed, energized, or operating 100 MW first phase , expansion up to 400 MW ; source boundary is not defined as IT, facility, campus, or generation Municipal architectural zoning review described; remaining licensing steps possible; no complete permit record verified No construction claim No operating or energized claim ND; 100 400 MW boundary not defined ND; the announced capacity is not relabeled as facility power Power discussions with Cemig ; no interconnection agreement, grid capacity, import limit, or tariff disclosed Company claims 100% renewable energy ; generation nameplate, PPA, storage, and annual energy ND ND RT-One developer; Uberlândia municipality planning zoning authority; Cemig power-discussion counterparty; DMAE water feasibility of 2.77 L s; operator, customer, and vendors ND No named external investor, lender, or project-finance structure More than R$6B announced project investment, original BRL, company claim; not a funding close, tariff, lease, or GPU price Count, model, state, and allocation ND RT-One project page — source 11 Oct 2025; Uberlândia municipal update — source 23 Feb 2026; both captured 4 Sep 2026

Current-run development-cost benchmark

Benchmark Status source date Possible or operating scope IT load power boundary Capital or price signal Control and direct evidence --- --- --- --- --- --- COST-01 · Cushman & Wakefield 2026 US Data Center Development Cost Guide 2026 guide; release date 3 Sep 2026 surfaced in the current evidence; official page captured 4 Sep 2026 US market benchmark, not a named project; pipeline approaching $2.3T , including $492B under construction and $1.8T planned precommitted No project-specific IT, facility, campus, grid, generation, PPA, or MVA value Construction costs up 21% per MW from the prior report; average greenfield cost $17.6M MW for the most modern facilities; original USD, per MW, 2026 market-average development-cost benchmark Do not apply the benchmark to a row, convert it to GPU capex, or treat it as an electricity tariff. Cushman & Wakefield guide

Accelerator, inventory, specification, company, investor, and price register

Quantity state remains separate from catalog orderability or product specification. No row below establishes a global installed-GPU total. Hosted rates are provider rental signals, not hardware purchase prices.

Record Quantity and state Model key specification Vendor, provider, and company role Explicit investor financing role Possible price — original unit, type, date Direct evidence --- --- --- --- --- --- --- GPU-01 · NVIDIA GB300 NVL72 Site-deployed quantity ND ; NVIDIA lists the platform Available Now catalog orderability ; Bitdeer A102 is planned rack-scale context; installed, delivered, ordered, committed, and allocated count ND Fully liquid-cooled rack with 72 NVIDIA Blackwell Ultra GPUs and 36 Arm-based Grace CPUs ; 800 Gb s network connectivity per GPU; rack specification is not site inventory NVIDIA OEM platform provider; Bitdeer intended deployment context; purchaser and operator allocation ND Investor, lender, and financing role ND OEM purchase and rental price ND ; no price inferred from site revenue or cloud rates NVIDIA GB300 NVL72 — vendor page captured 4 Sep 2026; related Bitdeer A102 release GPU-02 · Lambda on-demand catalog Catalog orderability: 1x, 2x, 4x, and 8x GPU instances; 1-Click Clusters offer 16–2,000+ interconnected B200 or H100 GPUs ; no deployed fleet count B200 SXM6 180 GB VRAM GPU ; H100 SXM 80 GB VRAM GPU ; A100 and V100 also listed Lambda cloud provider operator; NVIDIA OEM; inventory-finance and project-allocation roles ND Investor, lender, and inventory-finance role ND B200 $6.69 GPU-hour; H100 $3.99 GPU-hour , USD, pay-by-minute on-demand catalog rates, captured 4 Sep 2026; taxes may apply; cloud rental, not hardware price Lambda GPU instances — current catalog captured 4 Sep 2026 GPU-03 · Nebius NVIDIA HGX B300 Provider-deployed quantity ND ; operating provider-deployed claim because Nebius says HGX B300 is deployed in production on Nebius AI Cloud; no count disclosed 8× Blackwell Ultra GPUs; 270 GB HBM3e GPU , 2.1 TB total HBM3e , 62 TB s total memory bandwidth, 14.4 TB s total NVLink bandwidth, 1.6 Tb s networking; projected performance labels retained Nebius cloud provider operator; NVIDIA platform OEM Investor, lender, and equipment-finance role ND $4.30 GPU-hour preemptible; $7.85 GPU-hour on-demand , USD, provider rental rates captured 4 Sep 2026; not hardware price Nebius HGX B300 and Nebius GPU pricing — captured 4 Sep 2026 GPU-04 · Nebius current GPU catalog — B200, H200, H100 Provider fleet count ND ; catalog orderability ; VM offers are not installed inventory B200 180 GB GPU ; H200 141 GB HBM3e GPU, 4.8 TB s ; H100 80 GB GPU ; specifications do not imply inventory Nebius cloud provider operator; NVIDIA OEM platform provider Investor, lender, and financing roles ND Nebius page captured 4 Sep 2026: B200 $3.95 preemptible $7.15 on-demand; H200 $2.45 $4.50; H100 $2.15 $3.85 per GPU-hour , USD, hosted rental Nebius GPU pricing — current rates captured 4 Sep 2026; NVIDIA H200 — vendor specification captured 4 Sep 2026; NVIDIA DGX B200 and Lambda instances — platform specification references GPU-05 · Nebius NVIDIA GB300 NVL72 Provider-deployed racks GPUs ND ; operating provider-deployed claim because Nebius says the platform is deployed in production on Nebius AI Cloud; no fleet count or customer allocation Liquid-cooled rack with 72 Blackwell Ultra GPUs and 36 Grace CPUs , 37 TB total fast memory, 1,440 PFLOPS sparse FP4 performance, 576 TB s total memory bandwidth as listed by Nebius Nebius cloud provider operator; NVIDIA OEM platform provider Investor, lender, and financing role ND Nebius pricing page says Contact us ; numerical rental and purchase price ND , captured 4 Sep 2026 Nebius GB300 NVL72 and Nebius GPU pricing — provider pages captured 4 Sep 2026; related NVIDIA GB300 NVL72

Current-run analytical charts

The charts use only same-unit or record-count comparisons. Nested phases and different economic objects are shown with explicit non-additivity notes.

Current-run evidence controls

Risk Treatment in this update Direct evidence --- --- --- Boundary and non-additivity IT load, facility campus power, utility interconnection, generation PPA, MVA, manufacturing output, and project capital remain separate. No global total is asserted. Hut 8 · Chevron · Macquarie Status and timing Announced, discussion-stage, permitted, under-construction, operating, and energized states are not substituted. Hut 8 current energization, Chevron FID, Heka site scale, Macquarie 45 47 MW discrepancy, and RT-One licensing remain open. MCIT · Uberlândia GPU inventory and specification NVIDIA Nebius rack specifications, provider claims, and catalog orderability are not treated as deployed counts. No global installed-GPU total is inferred. NVIDIA GB300 · Nebius B300 Price type Provider rental rates are preserved in original USD GPU-hour units and dates. C&W is a market-average development-cost benchmark. No OEM hardware price was verified. Lambda · Nebius pricing · C&W guide

Report details

Run date: 4 September 2026 UTC · Evidence capture: 4 September 2026 UTC · Status: draft candidate awaiting evaluation; not published · Owner: General Codex Agent · Run: looprun-5e7802d8-64de-4361-99c1-012dc56b4b9b · Session: loopsession-bf17750f-0c33-4ae9-bdd6-5f3200747c37 · Closure: report candidate updated and dashboard upsert performed; no publication, schedule change, external communication, or production-data change performed.