
GPU cloud market research · Q4 2026
The GPU cloud market is wider than the familiar hyperscalers and neocloud brands. It includes traditional hosting companies, dedicated-server operators, marketplaces, brokers, managed inference platforms and regional infrastructure specialists. The difficult part is not finding a website that mentions an accelerator. It is determining what the provider actually sells, where the service runs, who operates the hardware and which claims can be verified.
This web edition of The GPU Cloud Footprint separates five things that are often collapsed into one number: candidate domains, confirmed GPU-service evidence, corporate supplier groups, advertised service geography and verified infrastructure observations. That distinction matters to anyone searching for GPU cloud providers, comparing H100 or H200 availability, or assessing the physical footprint behind an AI compute offer.
01 · Market sizeHow many GPU cloud providers are there in 2026?
There is no defensible single global count yet. ViaBandwidth screened 872 candidate domains and retained 722 active candidates under its October audit rules. Within that population, a frozen cohort of 339 domains had explicit GPU-service evidence. A later November reconciliation identified 387 provisional supplier groups using broader company, brand and relationship rules.
Those numbers answer different questions. The 339 figure is a reproducible domain-level evidence cohort. The 387 figure is a working group-level market map that restores providers and hyperscalers absent from the earlier screen, but still contains inherited classifications and unresolved legal or brand relationships. Subtracting one from the other would not measure market growth.
| Population | Count | What it means |
|---|---|---|
| Candidate domains screened | 872 | Original research universe; not all are GPU providers |
| Active retained candidates | 722 | Domains with a documented October disposition |
| Explicit-service evidence cohort | 339 | Frozen October baseline for model and network analysis |
| Provisional supplier groups | 387 | November company-level working taxonomy; not a fleet-owner count |
Intent answer: ViaBandwidth can substantiate a 339-domain October GPU-service evidence cohort, but does not claim that 339—or the provisional 387 groups—is the definitive number of GPU cloud companies worldwide.
02 · Provider typesDirect GPU cloud, reseller or marketplace?
Where a supplier sits in the sales chain changes what a buyer can verify. The frozen 339-domain cohort contains 243 directory-classified direct providers, 40 resellers, 15 marketplaces and 41 unclassified domains. These are commercial-role classifications, not proof that 243 companies own GPUs.
| Operating classification | Domains | What the label establishes |
|---|---|---|
| Direct | 243 | Provider presents the service directly; hardware ownership remains a separate question |
| Reseller | 40 | Offer is delivered through another provider or upstream capacity relationship |
| Marketplace | 15 | Platform connects buyers with supply from one or more operators |
| Unclassified | 41 | Public evidence does not support a stable operating-role decision |
What sort of businesses are these?
The business mix is also broad: 103 traditional-cloud domains, 62 neoclouds, 44 sovereign or regional providers, 26 marketplace-sector businesses, 24 managed-MLOps providers, 19 dedicated hosts, 16 distributed platforms and 11 colocation-plus-GPU businesses. A buyer looking for a dedicated GPU server has a different procurement problem from a team buying an on-demand H100 instance through a marketplace.
Use the GPU provider directory to inspect individual suppliers, and the direct versus reseller versus marketplace guide for the commercial implications.
03 · AcceleratorsWhich GPUs appear across cloud-provider evidence?
Within the frozen cohort, 222 of 339 domains (65.5%) have a bounded named-GPU record. The other 117 are unknown, not GPU-free. At least 106 confirmed domains have H100 evidence, followed by A100 (91), H200 (87), RTX PRO 6000 (69), L40S (64), B200 (41), RTX A6000 (38), L4 (37) and B300 (36).
Leading models within 339 confirmed domains. Counts are conservative minimums based on bounded public model evidence; models overlap and do not measure installed inventory or stock.
| GPU model | Confirmed domains | How to interpret it |
|---|---|---|
| NVIDIA H100 | 106 | Bounded public model evidence; not installed stock |
| NVIDIA A100 | 91 | Provider domain may advertise several configurations |
| NVIDIA H200 | 87 | Model mention does not establish region or availability |
| RTX PRO 6000 | 69 | Includes professional-GPU hosting and cloud offers |
| NVIDIA L40S | 64 | Often positioned for inference and graphics workloads |
| NVIDIA B200 | 41 | Includes advertised and contract-led access |
| NVIDIA B300 | 36 | Operational and announced capacity must be separated |
The scoped price ledger is a different layer. It contains 377 H100 offer rows from 90 domains, 394 A100 rows from 85 domains and 318 H200 rows from 81 domains. Offer rows are not interchangeable GPU-hour prices, and a priced configuration does not guarantee stock. Buyers comparing rates should use the GPU cloud pricing comparison guide and normalize the complete configuration, commitment and region.
04 · Generation and route to marketThe same GPU can be sold four different ways
H100 evidence appears across 73 direct-classified domains, 20 resellers, six marketplaces and seven unclassified providers. H200 appears across 61 direct providers, 13 resellers, six marketplaces and seven unclassified domains. B200 and B300 evidence is more concentrated, but still crosses operating models.
| GPU model | Direct | Reseller | Marketplace | Unclassified |
|---|---|---|---|---|
| H100 | 73 | 20 | 6 | 7 |
| A100 | 66 | 15 | 5 | 5 |
| H200 | 61 | 13 | 6 | 7 |
| L40S | 40 | 11 | 6 | 7 |
| B200 | 26 | 10 | <5 | <5 |
| B300 | 25 | 6 | <5 | <5 |
Direct means a classified business role, not audited GPU ownership. Suppressed cells are omitted, not zero.
Generation presence is overlapping
This is why a search for “H100 cloud provider” should not be treated as a list of equivalent suppliers. One result may operate a cloud region, another may lease dedicated nodes from an upstream operator, a third may aggregate third-party machines, and a fourth may sell a managed inference endpoint without offering raw GPU access. The accelerator label is the start of diligence, not the end.
05 · Network evidenceWhat ASN and IP-prefix data can prove
Among the 243 direct-classified domains, 103 have a retained nonzero routed-prefix count. Across the full 339-domain cohort, 126 have nonzero prefix counts and 133 retain at least one ASN identifier. These signals help distinguish a provider with observable network infrastructure from a brand with no retained network identifiers.
They do not prove GPU ownership, installed inventory or deployment geography. The stored “network verified” flag is defined by whether a prefix count is nonzero, so it is not an independent second test. A network facility association may show where an ASN is present while saying nothing about whether the provider has a GPU cluster in that building.
06 · GPU cloud locationsWhere providers say GPU services are available
The November master registry accepted 149 advertised service-country pairs across 89 company IDs and 41 countries. The most frequently represented countries in that evidence include India and the United States (17 company IDs each), Canada and Germany (13 each), the Netherlands (10), France (8), Singapore (7), Finland (6), and Japan and the United Kingdom (5 each).
89 company IDs / 41 named countries / 149 accepted company-country pairs
Accepted advertised service evidence, not installed capacity, unique data centers or live availability.
These are provider-advertised service claims—not a ranking of installed GPUs, data-center buildings or available capacity. A cloud region can span multiple facilities. A marketplace can show supply in a country without owning hardware there. A provider can also route traffic through a network POP far from the GPU host.
For location-specific research, browse the United States GPU provider directory, Germany, the Netherlands and other country pages. Their inclusion rules remain distinct from physical-facility verification.
07 · Infrastructure sampleWhere specific GPU models meet the map
A targeted 13-provider investigation produced 67 location records. Fifty-one rows were operationally labelled across 16 countries, and 44 explicitly linked a named GPU model to a region or zone. This was a purposive evidence test, not a representative market sample.
13 providers / 67 location records / 44 named-model operational links / 16 operational countries
Primary provider evidence class (13 selected)
Status of location-level evidence records (67 rows)
- IBMpublishes H100 and H200 availability by IBM Cloud zone. That establishes a model-to-zone relationship, not street-level building ownership.
- DODigitalOcean publishes a model-by-datacenter matrix covering H100, H200, MI300X, MI325X, MI350X and B300 offerings.
- EXOExoscale documentation maps A30, A5000 and RTX PRO 6000 products to zones, with additional partner corroboration for service operation.
- ALPHAAlphaVPS names Telepoint and Hetzner as facility operators for its GPU nodes, separating service operation from building ownership.
- CIVOOperational LON1 GPU coverage is separate from its announced Hertfordshire Vera Rubin project.
- SILICONProposed Texas and Arizona clusters remain announced capacity until commissioning evidence exists.
- COLOColoCrossing advertises GPU-capable services and ten cities, but does not publish a model-specific city matrix.
All 275 retained network joins for the selected providers remain network-only observations. None is counted as a GPU deployment merely because a provider or ASN appears at a facility.
08 · Q3 and Q4: the correct comparisonTwo snapshots. Two definitions. No invented trend.
The published July cohort cannot be reconstructed record-for-record.
NO
DELTA
Why counts cannot be subtracted
An indexed tag, a verified commercial GPU offer, a corporate group and an independently operated GPU site are separate concepts. No numerical percentage change is supported because the row-level membership and selection rules differ.
09 · Updated provider registryA market of sellers, not one fleet
The November reconciliation expands the picture of who participates in GPU service delivery, while keeping corporate ownership and operating roles distinct.
872 original domains / 114 targeted checks completed / 387 provisional supplier groups
Working supplier-group taxonomy | 387 provisional groups
Includes marketplaces, brokers, managed platforms and seven hyperscalers. Many labels remain inherited, not freshly audited. Categories sum to 387.
Why this is not “supplier growth”
The earlier 339 count described domain-level evidence inside a frozen October cohort. The 387 figure is a November group-level working reconciliation. Changing the denominator is a research improvement, not evidence that 48 new suppliers entered the market.
11 · Infrastructure reality checkThe four different kinds of footprint
A GPU cloud may involve a service seller, a hardware operator, a network operator and a building owner. Those roles need not belong to the same company.
Product page / claim
Offer-location evidence
Routing and presence
Stock and legal title
339 provider network summaries | 133 domains with ASN IDs | 126 with retained prefix counts
Buyer diligenceHow to evaluate a GPU cloud provider
A useful GPU cloud comparison should ask more than “Do you offer H100s?” The strongest suppliers make the following details visible before contract:
- 01Exact accelerator. Model, memory variant, GPU count and interconnect topology.
- 02Delivery object. Virtual machine, bare metal, dedicated cluster, container, endpoint or marketplace host.
- 03Location. Region or zone for the exact model, not a generic global-footprint map.
- 04Status. Operational, orderable, waitlisted, contract-only, under construction or merely announced.
- 05Responsibility. Who owns or leases the GPU, who operates the server and who controls the facility.
- 06Commercial terms. Commitment, billing increment, storage, data movement, support and taxes.
The GPU cloud buyer’s guide turns those questions into a practical shortlist process. For pricing decisions, Buying GPU Services and What an H100 actually costs provide the contract-level framework.
09 · Research conclusionTransparency is part of the infrastructure product
A GPU listing makes a provider discoverable. Verifiable operating details make its footprint meaningful. The public market is strongest where suppliers publish model-by-region availability, distinguish operational capacity from future projects, identify the service operator and explain whether the offer is direct, resold or aggregated.
For future quarters, ViaBandwidth will keep company identity, commercial role, GPU offering, model-region availability, physical facility evidence and network presence as separate measures. That is slower than adding every visible number into one market-size claim, but it produces evidence that buyers, operators and researchers can actually use.
10 · FAQGPU cloud footprint questions
What is a GPU cloud provider?
A GPU cloud provider sells remote access to accelerator-backed compute. The delivered product may be a VM, bare-metal server, dedicated cluster, Kubernetes node, managed inference endpoint or marketplace instance. The provider selling the service is not always the company that owns the hardware or facility.
How many GPU cloud providers are there?
ViaBandwidth retained a reproducible 339-domain October evidence cohort and a broader provisional registry of 387 supplier groups. Neither is presented as a definitive global census because company relationships, reseller roles and evidence rules differ.
Which GPUs are most commonly advertised?
H100, A100 and H200 are the most frequently evidenced named models in the frozen cohort. RTX PRO 6000, L40S, B200, RTX A6000, L4 and B300 also appear across multiple provider types.
Does a cloud region prove a physical GPU data center?
No. A region is service geography. It may contain several buildings or an upstream partner deployment. Physical-facility claims require separate evidence connecting the model or service to a named site.
Does an ASN prove that a company operates its own GPUs?
No. ASN and prefix evidence show network presence or routing relationships. They do not establish hardware ownership, GPU installation, inventory or live ordering capacity.
Present this report
One sponsor can present this edition with a clearly labelled credit on the article, downloadable PDF and related GPU research surfaces. Sponsorship never changes the evidence, findings or provider treatment.
- ▸Logo and presented-by credit on the report and downloadable PDF
- ▸A credit beside relevant GPU research and provider pages
- ▸Exclusive to one sponsor for this edition
- ▸Labelled sponsorship with no influence over the research
Or write to [email protected].