Who Owns the Computing Power?
Gary WhittakerHow chipmakers, cloud platforms, AI laboratories, data-centre landlords, financiers and governments divide control over advanced artificial intelligence.
By Jack Righteous · Mont-Real · Fact-checked July 19, 2026
The Name on the Building May Not Be the Owner
A person opens an AI application owned by one company. The model may be developed by that company, licensed partly to another, hosted through a third cloud platform and operated on processors designed by a fourth.
Those processors may sit inside a facility leased from a data-centre operator. The building may belong to a real-estate partnership. Debt may finance the hardware. A regulated utility may control delivery of the electricity. National law may determine whether the chips, model or service can be exported.
Each participant owns or controls something important. None necessarily controls the complete system alone.
Asking who owns the data centre can produce a legally correct answer while still failing to explain who controls the computing power inside it.
Ownership Is a Chain, Not a Name
One AI service can depend on at least seven separately owned layers.
1. Land
The site may belong to a technology company, developer, government, utility, landlord or joint venture.
2. Building
The facility can be owned directly, leased, built to suit or held through an infrastructure partnership.
3. Power systems
Utilities, facility operators, power producers and special project companies may own different parts of the electrical chain.
4. Hardware
Processors and servers may be purchased, leased, financed or owned by the cloud platform rather than the AI company.
5. Cloud service
The provider controls scheduling, networking, security, billing, availability and allocation of the underlying capacity.
6. AI model
Rights can be divided across model weights, source code, product interfaces, licences, fine-tuned versions and research methods.
7. Financing and contracts
Lenders, investors, landlords and major customers can influence construction, expansion and priority access without operating the servers.
Legal Ownership Is Only One Kind of Power
Property control
The ability to sell, lease, expand, redevelop or restrict access to land and facilities.
Technical control
The ability to operate processors, software, networks, cooling, security and model deployment.
Contractual control
Rights created through cloud agreements, leases, power contracts, intellectual-property licences, take-or-pay commitments and equipment financing.
Capital control
The ability to finance expansion, require collateral, restructure debt, acquire assets or force a sale after default.
Regulatory control
The authority to permit facilities, restrict exports, impose sanctions, require data residency, approve mergers or limit uses for national security.
The party with the strongest control changes according to the decision being made.
The Processor Layer: A Chokepoint Before the Building Opens
An empty building with a powerful electrical connection is not an AI system. Advanced computing requires accelerators, high-speed networking, memory, storage and software designed to make the hardware work together.
NVIDIA reported approximately US$193.7 billion in data-centre revenue for its fiscal year ending January 25, 2026. Its platform includes processors, networking, interconnects, software libraries, models, application frameworks and development tools.
That scale does not mean NVIDIA owns every facility or model using its technology. It does mean that organizations building advanced AI can become dependent on its processor designs, networking architecture, software ecosystem, product allocation and release timing.
NVIDIA’s filing also describes concentrated demand from a limited number of direct and indirect customers. Concentration exists on both sides: a limited supply chain serves a limited group of buyers capable of purchasing at extraordinary scale.
The Hyperscalers: Renting Computing Power at Global Scale
Most cloud customers do not buy a complete physical system. They buy access to processing, storage, networking, databases, security tools, AI services and deployment platforms.
The cloud provider decides how the underlying physical infrastructure is organized, maintained and expanded. It can distribute a customer’s workload across facilities while presenting one service interface.
Amazon Web Services, Microsoft Azure and Google Cloud combine global data-centre networks, enterprise software, security systems, billing relationships, development platforms and government contracts. Their advantage is not one building. It is the connected system.
That integration can reduce cost and complexity for customers. It can also create dependency when an application is built around one provider’s databases, identity tools, storage formats, security systems, discounts and software licences.
A business may be able to download its data while still finding that moving the complete application is expensive and technically difficult.
When Switching Becomes Difficult, Infrastructure Becomes Leverage
The United Kingdom’s Competition and Markets Authority completed a cloud-services investigation in July 2025. It identified significant market power held by Amazon Web Services and Microsoft and raised concerns involving data-transfer fees, interoperability, switching and software licensing.
In March 2026, the regulator said Amazon and Microsoft had taken material steps to reduce some transfer charges and improve interoperability, while announcing continued review and a separate investigation of Microsoft’s business-software ecosystem.
The strongest argument for integrated cloud platforms is real. They can provide security, global reliability, compliance tools and technical services that smaller providers may struggle to match.
The competition concern begins when the cost of leaving becomes greater than the customer reasonably understood when entering.
Review the UK cloud-services investigation and the March 2026 follow-up.
OpenAI: One Model Company, Several Infrastructure Relationships
OpenAI demonstrates why the power structure behind one AI service cannot be reduced to one partnership.
Under an amended agreement announced April 27, 2026, Microsoft remains OpenAI’s primary cloud partner. OpenAI products are to launch first on Azure unless Microsoft cannot and chooses not to provide the required capability. OpenAI can also serve its products through other cloud providers.
Microsoft’s licence to OpenAI model and product intellectual property continues through 2032 and is now non-exclusive. Microsoft also remains a major shareholder.
Two months earlier, OpenAI announced US$110 billion in new investment: US$50 billion from Amazon, US$30 billion from NVIDIA and US$30 billion from SoftBank. It also announced a strategic AWS partnership and next-generation inference computing with NVIDIA.
These relationships overlap without being identical.
Shareholder, primary cloud partner, IP licensee and distribution partner
Investor, AWS partner, enterprise distributor and custom-processor provider
Investor, processor supplier and systems partner
Investor and infrastructure partner through Stargate relationships
OpenAI can direct and own its products while depending on several companies for capital, compute, licensing and distribution. That does not prove that one partner secretly owns the others.
The correct reporting separates equity, cloud rights, intellectual-property licences, customer distribution, compute commitments and governance.
Review the amended Microsoft agreement · OpenAI’s February 2026 investment announcement · the Amazon partnership
CoreWeave: Control Through Leases, Debt and Customer Commitments
Specialized AI cloud companies offer another ownership model.
At the end of 2025, CoreWeave reported operating 43 data centres with more than 850 megawatts of active power and approximately 3.1 gigawatts of contracted power expected to be deployed over future periods.
The company said it primarily financed infrastructure through asset-level debt supported by take-or-pay customer contracts. It reported US$60.7 billion in remaining contracted performance obligations, with committed contracts averaging approximately five years.
CoreWeave relies heavily on third-party facilities and long leases, while beginning to develop some of its own sites. Its filing states that all GPUs then used in its infrastructure were NVIDIA GPUs because of obligations in its customer contracts.
Customers generally purchase cloud services rather than control one identified server. CoreWeave retains discretion over how servers are selected and deployed while meeting the promised service.
The chain can look like this
Landlord owns the facility → CoreWeave leases capacity → lenders finance processors → customers sign take-or-pay commitments → CoreWeave allocates and operates the computing service.
Leasing can help a newer cloud company compete without owning every building. It also creates a chain in which each participant depends on the next contract remaining intact.
The Data-Centre Landlords Behind the Cloud
Colocation companies provide secure space, electricity, cooling, fibre connections, physical security and interconnection with networks and cloud platforms. The customer can control its equipment without owning the complete building.
Equinix reported a footprint of 280 data centres at the end of 2025. Its standard facilities provide customers with space, power and interconnection. Its xScale facilities are designed for hyperscale cloud and AI workloads and are developed through joint-venture partnerships.
Those joint ventures matter because the company operating and marketing the facility may share ownership, financing and major decisions with investment partners.
Digital Realty also holds substantial investments through partnerships. Its filings identify unconsolidated entities involving firms such as Blackstone, GI Partners, Mapletree, Mitsubishi, Realty Income and others.
A community may therefore approve a project for one named developer while the ultimate investors, tenants, lenders and operators change over the life of the facility.
Public documents should identify the landowner, development company, facility operator, joint-venture partners, expected tenant, power customer and party responsible for environmental obligations.
Review Equinix’s 2025 filing and Digital Realty’s 2025 filing.
Computing Ownership Is Not the Same as Data Ownership
Rights over a cloud service, model and customer information can be divided just as physical ownership is divided.
Text, images, music, business records or other information sent into the system
Usage, security, performance and diagnostic information created by the platform
The numerical parameters that make the trained model function
Models adapted for a particular customer, task or dataset
Text, music, images, code or analysis produced through the service
Embeddings, indexes and representations created from customer information
One platform may allow customers to retain ownership of their input while reserving limited rights to operate the service. Another may permit training unless the customer opts out. Enterprise contracts may differ from consumer terms.
The correct answer must come from the current service terms, data-processing agreement, enterprise contract, privacy policy and model licence.
A company may own its data and still lack the technical ability to use it without the platform that stores, processes or interprets it.
When a Country Wants Computing Power It Cannot Lose
Sovereign computing is often described as keeping servers inside a country. Location is only one layer.
Canada’s 2026 AI Sovereign Compute Infrastructure Program defines sovereignty through domestic location, data residency, operational control, governance and decision-making authority.
The program requires core computing and storage infrastructure to be owned or contractually controlled by Canadian entities, with safeguards limiting the ability of a foreign party to restrict access unilaterally. Decisions over who may use the infrastructure and for what purposes must rest with Canadian institutions or organizations.
Canada has allocated approximately $890 million toward the design, construction and ongoing operation of a large-scale public AI supercomputing system over seven fiscal years.
Sovereign compute can reduce foreign dependency and protect sensitive data. It can also concentrate authority domestically. The governing institution may decide which researchers receive access, which projects qualify, which security restrictions apply and how usage is monitored.
Sovereignty can protect a country from foreign control. It does not remove the need to examine domestic governance.
Review Canada’s sovereign-compute program and its governance requirements.
The Ultimate Test of Control Is the Ability to Say No
Practical control becomes visible when one party can prevent another from using the system.
Can restrict expansion or end a lease under the contract
Can delay or limit power delivery under regulatory rules
Can suspend service, reallocate capacity or change availability
Can change model access, capabilities and permitted uses
Can limit allocation, delay products or withdraw support
Can impose export, sanctions, security or procurement restrictions
Can enforce collateral or restrict capital after default
Can withdraw future revenue where contractual terms permit
The existence of a legal right does not prove it will be abused. Citizens and customers should still examine notice, appeal rights, portability, due process, emergency powers and contract termination.
Scale Can Create Benefits That Fragmentation Cannot
Concentrated infrastructure is not harmful merely because it is large.
Major integrated providers can offer global availability, advanced security, faster hardware deployment, disaster recovery, compliance systems, specialized expertise and clusters that smaller organizations could not finance.
Many customers do not want to own servers, cooling systems, networking equipment or data-centre property. They want a reliable service that allows them to build.
Long-term contracts can also give infrastructure providers enough certainty to finance new facilities.
The concern begins when scale becomes a barrier to meaningful competition, switching, independent research, national resilience or public oversight.
What Happens When Too Few Organizations Hold the Keys?
- Higher prices: Customers may have limited alternatives after deep integration.
- Switching barriers: Moving data is not the same as rebuilding applications and security systems.
- Research dependence: Universities may rely on companies they are expected to evaluate independently.
- National dependence: Governments may rely on foreign providers for essential systems.
- Policy by contract: Private service terms can shape permitted uses without a public legislative process.
- Systemic failure: One outage, cyberattack or financing problem can affect many dependent services.
- Unequal access: Smaller firms and researchers may receive less favourable prices or hardware.
- Private influence: Capital may flow toward the projects offering the strongest return rather than the greatest public value.
Decentralization is not an automatic solution. More providers can also create inconsistent security, weak accountability, duplicated infrastructure and lower utilization. The goal is not fragmentation for its own sake. It is enough competition, portability, transparency and public capacity to prevent dependency from becoming submission.
What We Know—and What Must Be Examined Relationship by Relationship
Creators Rarely Own the Infrastructure Behind Their Work
An AI music creator may own original lyrics, arrangement decisions, released recordings and a public brand, subject to the platform terms and applicable law.
The creator usually does not own the model, servers, generation platform, cloud account or underlying training system.
That does not make the work meaningless. It does mean creators should understand which parts of the process depend on someone else’s permission.
Creators can reduce dependency by exporting project files, preserving lyrics and prompts, keeping local backups, documenting licences, maintaining direct audience relationships and avoiding a brand built completely on one platform.
We do not need to own a data centre to create meaningful work. We do need to protect the parts of the work that can remain ours.
Questions to Ask Before Accepting “The Company Owns It”
Physical ownership
- Who owns the land and building?
- Is the facility held through a joint venture?
- Who owns the substation and internal power systems?
- Who is responsible for environmental obligations?
- Can the land, building and equipment be sold separately?
Computing and cloud control
- Who owns or finances the processors?
- Who controls server allocation and operating software?
- Which cloud provider operates the service?
- Is the cloud agreement exclusive?
- Can the same model operate through another provider?
Model and data rights
- Who owns the model weights and product interface?
- Who holds intellectual-property licences, and when do they expire?
- Who owns customer input data and service logs?
- Can data be used for training?
- Can customers export data and fine-tuned systems?
- Which country’s laws govern the data?
Financing and government authority
- Who provided equity and debt?
- What assets can lenders seize after default?
- How long are the leases and customer commitments?
- Which export controls and security laws apply?
- Can a foreign party restrict access?
- Do obligations continue after a sale?
- Who can citizens hold accountable when ownership changes?
Primary Sources Used
- NVIDIA Fiscal 2026 Annual Report
- UK Competition and Markets Authority Cloud Investigation
- OpenAI and Microsoft Amended Partnership
- OpenAI February 2026 Investment Announcement
- CoreWeave 2025 Annual Report
- Equinix 2025 Annual Report
- Digital Realty 2025 Annual Report
- Canada AI Sovereign Compute Infrastructure Program Guide
Ownership Is Not the Same as Control
The computing power inside a data centre may be divided among a landlord, utility, chipmaker, cloud company, AI laboratory, investment fund, lender and government.
Each may control a different part of the system.
That structure can make advanced AI possible at a scale no single organization could build alone. It can also make responsibility difficult to locate when access is restricted, costs rise, promises fail or citizens demand answers.
The question is therefore larger than who owns the building.
Who can keep the processors running?
Who decides which model receives capacity?
Who controls the data?
Who can change the terms?
Who can shut the system down?
Those answers reveal where the real power sits.
And once again, that power becomes physical inside the data centre.