7 segments · 62 companies · as of Sep 18, 2026
Hyperscale Cloud
The cloud operators and captive builders buying AI compute at scale
Where it sits
Fed by: Servers & Systems · Data Center InfrastructureFeeds: Models & Platforms
Cloud operators and large AI developers turn chips, power and data centers into usable compute. This theme separates merchant clouds from captive buyers and regional providers. Spending figures identify their period and accounting basis; corporate capex is not a measure of AI spending alone.
Why now
July and August guides raised several US spending plans, while Oracle updated its fiscal-year outlook in September. Regional clouds are buying Blackwell clusters, and Chinese providers face accelerator import limits. Contracted and planned capacity still exceeds energized supply.
US-serving hyperscale clouds
Large merchant clouds serving US AI demand range from broad public-cloud platforms to specialist GPU suppliers. US headquarters are not required. Spending guides use different periods and lease treatments; corporate capex is not AI-only spending.
AWS GPU/Trainium build; July raises CY2026 cash capex to ~$220B (est.), including non-AWS spending
leaderAzure GPU/Maia build; July CY2026 capex ~$175B (est.) after lease reclassification, not an investment cut
leaderGoogle Cloud GPU/TPU build; July raises CY2026 capex guide to $195-205B (est.)
majorOCI/OpenAI campuses; September FY2027 capex guide $90-95B (est.), net cash outlay up to $70B (est.)
majorBuilds US GPU regions for frontier labs; August CY2026 capex guide $35-39B (est.)
challengerBuilds US Microsoft/Meta GPU clusters; August CY2026 capex guide $20-25B (est.), versus Q2 spending of $5.7B (est.)
challengerCloud Blackwell Ultra instances supply US enterprise training and inference alongside managed watsonx services
niche
Sources: microsoft.com, ir.aboutamazon.com, abc.xyz, srgresearch.com, unwire.pro, assets.nebius.com
Captive and frontier compute buyers
Model developers and consumer platforms reserve cloud capacity or build their own clusters. Corporate capex, multi-year purchase commitments and investment rounds fund different assets and are shown separately rather than treated as equivalent budgets.
Plans 5 GW Hyperion and >$50B regional investment; reported phase 1 targets 1.5 GW by late 2027, not live
leaderAWS pact expands by $100B/2 GW Trainium (est.); Stargate buyer; standalone 2026 capex undisclosed
leaderReported ~$517B/14.8 GW compute commitments in 11 months; AWS/Google ~11 GW/>$300B over ~10 years, not live
majorOwns xAI/Colossus after February merger; Q2 AI capex $15.8B (est.); also leases capacity to Anthropic
majorBuilds Cortex for autonomy and robotics; July CY2026 group capex guide >$25B (est.), not all AI
majorBuilds Apple-silicon Private Cloud Compute and rents third-party cloud; separate 2026 AI capex undisclosed
niche
Sources: investor.atmeta.com, anthropic.com, openai.com, en.sedaily.com, about.fb.com
Chinese cloud infrastructure
Chinese clouds combine merchant GPUs, domestic accelerators and state-backed networks. Import restrictions affect procurement. Disclosed quarterly capex is retained as quarterly data where a comparable full-year guide is unavailable.
AliCloud GPU and domestic-chip build; June-quarter capex RMB67.7B (est.), above its old three-year plan pace
leaderExpands AI cloud and captive models; Q2 capex RMB52.8B (est.); domestic accelerators supplement imported GPUs
majorBuilds Ascend-based Huawei Cloud capacity under import constraints; separate 2026 cloud capex undisclosed
majorVolcano Engine/Doubao buyer; reported CY2026 capex discussions up to $70B (est.), not company guidance
majorAI Cloud and Kunlunxin capacity; Q2 capex RMB11.4B (est.); no comparable full-year AI guide
majorMobile Cloud build; 2026 group capex RMB136.6B (est.); compute RMB37.8B (est.) is only part of the total
majorTianyi Cloud expansion; 2026 capex plan RMB73B with RMB25.5B in compute (est.), distinct from H1 spending
majorUnicom Cloud and AI clusters; 2026 group capex plan RMB50B (est.), not all AI
majorGPU cloud for Xiaomi/Kingsoft; H1 capex plus leased assets RMB6.2B (est.), not a full-year guide
challengerSenseCore Token Factory supports SenseNova and sells AI compute externally; August reports 48,000 PFLOPS deployed (est.)
challengerReported JD Cloud/Moore Threads 100k-GPU cluster plan (est.) targets merchant training and inference; not delivered
challengerBlueCloud operates Azure China; NC/NV GPU instances run in Chinese regions under local operating arrangements
niche
Sources: alibabagroup.com, tencent.com, ir.baidu.com, unwire.pro, chinaunicom.com.hk, bloomberg.com
Enterprise and European clouds
Enterprise and European providers sell controlled infrastructure, local data residency and managed AI services. Their smaller spending bases and sovereign requirements support differentiated clouds without making them peers of the US hyperscalers.
Power Virtual Server integrates watsonx with customers' hybrid Power estates; private-cloud availability expands in 2026
majorEU AI Cloud and Delos sovereign infrastructure; EUR20B long-term investment plan (est.), not 2026 capex
majorOwns European cloud/GPU infrastructure; FY2026 recurring capex guide 33-35% of revenue (est.)
majorScaleway GPU cloud and OpCore partner; H1 group capex EUR804M (est.); separate AI capex undisclosed
majorSovereign cloud and GPU servers; August CY2026 group capex guide EUR75-85M (est.)
challengerSTACKIT cloud and planned Luebbenau campus; EUR11B project budget (est.), not annual 2026 capex
challengerAdds RTX PRO GPU dedicated servers in 2026; private operator does not disclose a cloud capex guide
nicheT-Systems supplies Industrial AI Cloud compute to SOOFI and enterprise customers; Munich service opened in February
majorSovereign Cloud launches in July with EU-engineered infrastructure for regulated enterprise and public-sector workloads
nicheAI Stack offers on-demand Kubernetes GPUs, dedicated L40S servers and private GPU clouds from European infrastructure
nicheExoscale's dedicated AI inference reaches general availability in June; customer models run on dedicated European GPUs
niche
Sources: srgresearch.com, news.sap.com, corporate.ovhcloud.com, atos.net, arubacloud.com, changelog.exoscale.com
Developer and edge clouds
These clouds sell smaller GPU clusters, developer infrastructure or distributed inference. GPU availability, software simplicity and enterprise operating requirements matter more than the scale of their general cloud revenue.
Builds GPU cloud; H1 cash equipment capex $81.6M (est.); 155 MW committed site capacity (est.), mostly future
majorBuilds GPU cloud for signed AI customers; August 2026 capex guide ~40% of revenue (est.)
majorWorkers AI runs distributed inference; 2026 network capex guide 14-15% of revenue (est.), not all GPUs
majorOhio AMD MI355X cluster targets 50 MW and 24k GPUs (est.); >$1B project budget (est.), not annual guidance
challengerSeptember Blackwell/Palantir sovereign-pod launch; operated private clouds, with 2026 AI capex undisclosed
nicheCompute edge platform and AI Accelerator cache model responses for OpenAI/Gemini APIs; acceleration, not owned GPU cloud
nicheAI Gateway routes and bills third-party model inference for applications deployed with the Vercel developer platform
challengerPython-native serverless GPU functions scale inference and training jobs on B200/H200 hardware with usage-based billing
challenger
Sources: fastly.com, vercel.com, modal.com, investors.digitalocean.com, rackspace.com, srgresearch.com
Korean cloud infrastructure
Korean clouds serve national GPU programs, researchers and enterprise customers with local data residency. Existing services are distinguished from national-center construction and longer-term GPU allocations that have not all been delivered.
Plans ~$10B Gak Sejong DSX AI factory with Nvidia/Brookfield; 55 MW initial build expands to 200 MW by 2028
leaderSCP B300 GPUaaS live since March; leads national AI-center build; separate 2026 capex undisclosed
majorVia NHN Cloud, operates Factory X; 7,656 B200 GPUs (est., September); separate AI capex undisclosed
majorVia KT Cloud, expands campuses; September plan KRW6T over five years (est.), not a 2026 annual capex guide
majorAWS Ulsan plans 41 MW in Nov 2027 and 103 MW in Feb 2029; expansion toward 1 GW remains a plan
majorAI Box modular GPU infrastructure; 1.2 MW per module (est.); no separate 2026 AI capex guide
challengerVia Kakao Enterprise, operates public cloud and government B200 GPU services; broad proprietary-fleet expansion is constrained
niche
Sources: samsungsds.com, nvidianews.nvidia.com, digitaltoday.co.kr, kedglobal.com, msit.go.kr, datacenterdynamics.com
Japanese and Indian regional clouds
Regional cloud operators combine domestic enterprise relationships with dedicated AI infrastructure. Multi-year facility investments and group telecom capex should not be read as current GPU purchases or as fully contracted AI revenue.
Via NTT DATA, adds global AI-ready campuses; FY2026 new-capacity target 410 MW (est.); AI capex not separated
leaderOsaka Sakai Blackwell cloud operating in 2026; AI-specific annual capex undisclosed
majorPlans Sakai/Tomakomai AI campuses; August Sakai opening target FY2027; AI-only 2026 capex undisclosed
majorJio/Jamnagar AI infrastructure; Meta signs 168 MW (est.) for phased delivery; separate 2026 AI capex undisclosed
majorExpands Koukaryoku H200/B200 cloud; July Rubin investment plan JPY26B (est.), not all spent in 2026
challengerEnterprise cloud and Singapore GPU build; August FY2026 group capex guide JPY38B (est.)
nicheVayu GPU Cloud supplies H100/H200/L40S; separate annual 2026 GPU infrastructure capex undisclosed
nicheIndian enterprise cloud and AI-ready campuses; FY2026 capex INR13.3B (est.), predominantly data centers
nicheSells enterprise cloud and managed services on an owned H100/H200 fleet; 2026 AI capex undisclosed
nicheCloud Service Generative AI Platform shares GPU servers in Fujitsu data centers while isolating customers' confidential data
nicheGPU Cloud's managed Slurm service supports Japanese B300 training clusters; domestic service launched in March
nicheOffers AI Platform Service across cloud and on-premises systems; announced phased AIaaS rollout from July 2026
niche
Sources: group.ntt, newsroom.kddi.com, sakura.ad.jp, iij.ad.jp, fujitsu.com, jpn.nec.com
Glossary
- Capex
- Spending on long-lived assets; cash capex, finance leases and customer-funded construction are not interchangeable.
- Hyperscaler
- A cloud operator with enough infrastructure and customers to deploy capacity across many large regions.
- Captive compute
- Infrastructure primarily built to run its owner's models or products rather than rent to unrelated customers.
- ASIC
- A chip designed for a specific workload, such as Google TPU, AWS Trainium or Microsoft Maia.
Heat, spend, exposure and shares are editorial estimates. How to read the map ↗