Who owns AI compute, models, cloud infrastructure, and distribution channels?
Who controls the conditions of intelligence?
Behind AI’s promise of universal access lies a more concentrated geography of power.
An AI assistant appears on a screen as a conversation. Behind it sits a chain of chips, data centres, model licences and distribution agreements. Each link has an owner. Several lead back to the same companies.
NVIDIA occupies a commanding position in advanced AI accelerators and the software used to run them. Amazon, Microsoft and Google dominate the cloud-services market, with Oracle expanding its role in AI infrastructure. In the second quarter of 2026, the three largest providers accounted for 63 per cent of global cloud-infrastructure services spending, according to Synergy Research Group. That measures commercial activity, rather than ownership of all computing capacity, but it gives some indication of the concentration involved.
The relationships extend across the supply chain. Microsoft is a major shareholder in OpenAI, its primary cloud partner and a distributor of its technology through enterprise and consumer software. The arrangement nevertheless has limits: under their April 2026 agreement, OpenAI can serve its products through other cloud providers, while Microsoft’s licence to its model and product intellectual property is nonexclusive. A financial stake, a supply contract and a right to distribute are distinct forms of power.
Google combines chip design, cloud infrastructure and Gemini development with established routes to users through Search, Android and Workspace. Amazon combines AWS, its own models and consumer services with an investment in Anthropic. Their expanded agreement illustrates how closely finance and infrastructure can become entwined: Amazon announced further investment while Anthropic committed to spending more than $100 billion on AWS technologies over ten years. These are commitments extending into the future, rather than expenditure already incurred.
Such arrangements can fund research and make powerful systems available at scale. They can also make independence expensive. A developer may have a choice of models while remaining dependent on a narrow group of companies for computing capacity, customers or finance. The FTC’s examination of major cloud–AI partnerships identified potential restrictions on access to essential resources and higher costs of switching providers.
The political significance lies in this accumulation of roles. A company that supplies infrastructure, distributes models and controls the interface through which people encounter them can influence several conditions of access at once: price, availability, visibility and permitted use.
For users, AI may feel increasingly accessible. The authority to set its terms remains much more narrowly held.
Behind AI’s promise of universal access lies a more concentrated geography of power.
An AI assistant appears on a screen as a conversation. Behind it sits a chain of chips, data centres, model licences and distribution agreements. Each link has an owner. Several lead back to the same companies.
NVIDIA occupies a commanding position in advanced AI accelerators and the software used to run them. Amazon, Microsoft and Google dominate the cloud-services market, with Oracle expanding its role in AI infrastructure. In the second quarter of 2026, the three largest providers accounted for 63 per cent of global cloud-infrastructure services spending, according to Synergy Research Group. That measures commercial activity, rather than ownership of all computing capacity, but it gives some indication of the concentration involved.
The relationships extend across the supply chain. Microsoft is a major shareholder in OpenAI, its primary cloud partner and a distributor of its technology through enterprise and consumer software. The arrangement nevertheless has limits: under their April 2026 agreement, OpenAI can serve its products through other cloud providers, while Microsoft’s licence to its model and product intellectual property is nonexclusive. A financial stake, a supply contract and a right to distribute are distinct forms of power.
Google combines chip design, cloud infrastructure and Gemini development with established routes to users through Search, Android and Workspace. Amazon combines AWS, its own models and consumer services with an investment in Anthropic. Their expanded agreement illustrates how closely finance and infrastructure can become entwined: Amazon announced further investment while Anthropic committed to spending more than $100 billion on AWS technologies over ten years. These are commitments extending into the future, rather than expenditure already incurred.
Such arrangements can fund research and make powerful systems available at scale. They can also make independence expensive. A developer may have a choice of models while remaining dependent on a narrow group of companies for computing capacity, customers or finance. The FTC’s examination of major cloud–AI partnerships identified potential restrictions on access to essential resources and higher costs of switching providers.
The political significance lies in this accumulation of roles. A company that supplies infrastructure, distributes models and controls the interface through which people encounter them can influence several conditions of access at once: price, availability, visibility and permitted use.
For users, AI may feel increasingly accessible. The authority to set its terms remains much more narrowly held.
1. The ownership stack
AI’s apparent simplicity rests on a complicated distribution of property, contracts and permissions.
Ask who owns an AI system and the answer changes with the layer being examined. The company designing the processor may depend on another to manufacture it. The laboratory developing the model may rent its computing capacity. The platform delivering the answer may own neither.
Ownership, operational control and contractual influence need to be distinguished. A shareholder does not necessarily control a laboratory; a cloud provider does not automatically own the models it hosts; access to customer data does not confer unrestricted rights to use it.
The following map identifies the principal points of control. The examples are selective, rather than a ranking.
AI’s apparent simplicity rests on a complicated distribution of property, contracts and permissions.
Ask who owns an AI system and the answer changes with the layer being examined. The company designing the processor may depend on another to manufacture it. The laboratory developing the model may rent its computing capacity. The platform delivering the answer may own neither.
Ownership, operational control and contractual influence need to be distinguished. A shareholder does not necessarily control a laboratory; a cloud provider does not automatically own the models it hosts; access to customer data does not confer unrestricted rights to use it.
The following map identifies the principal points of control. The examples are selective, rather than a ranking.
Several additions matter. Huawei belongs in a global account of AI computing, while specialist clouds such as CoreWeave complicate a picture centred solely on the largest providers. Advanced packaging also deserves explicit attention: TSMC identifies it alongside leading-edge fabrication as a substantial area of investment.
The model layer requires particular care. DeepSeek is a significant omission from the original list. Corporate structures also change: SpaceX announced its acquisition of xAI in February 2026. Neither a company’s name nor its ownership structure, however, tells us everything about the freedoms attached to a particular model release.
The data layer is equally easy to overstate. An organisation may entrust information to a hosted service while retaining rights over that information and restricting its reuse. OpenAI, for example, states that business and API inputs and outputs are not used to train its models by default. Hosting a workload is therefore not equivalent to acquiring its contents as a training resource.
The table describes a chain of dependencies rather than a single proprietor. Concentration becomes consequential where those dependencies overlap: when a supplier also finances a developer, hosts its models and controls an important route to its customers. The FTC’s study of major cloud–AI partnerships identified potential effects on access to computing resources, switching costs and commercially sensitive information.
The practical question is how much freedom remains at each layer: to enter, to negotiate, to move and to continue operating when a supplier changes the terms.
The model layer requires particular care. DeepSeek is a significant omission from the original list. Corporate structures also change: SpaceX announced its acquisition of xAI in February 2026. Neither a company’s name nor its ownership structure, however, tells us everything about the freedoms attached to a particular model release.
The data layer is equally easy to overstate. An organisation may entrust information to a hosted service while retaining rights over that information and restricting its reuse. OpenAI, for example, states that business and API inputs and outputs are not used to train its models by default. Hosting a workload is therefore not equivalent to acquiring its contents as a training resource.
The table describes a chain of dependencies rather than a single proprietor. Concentration becomes consequential where those dependencies overlap: when a supplier also finances a developer, hosts its models and controls an important route to its customers. The FTC’s study of major cloud–AI partnerships identified potential effects on access to computing resources, switching costs and commercially sensitive information.
The practical question is how much freedom remains at each layer: to enter, to negotiate, to move and to continue operating when a supplier changes the terms.
2. Compute: who controls the capacity to build frontier AI?
Before a model can answer a question, someone must secure the machines, the electricity and the time to train it.
Computing capacity is one of the material foundations of AI power. At the frontier, an ambitious research programme requires access to large clusters of accelerators, fast connections between them, specialized memory and a reliable electricity supply. The ability to assemble these resources—and keep them operating—shapes who can participate.
Control is distributed across several businesses. Chip designers supply the processors and software. Manufacturers turn those designs into working hardware. Infrastructure operators bring the equipment together and sell access to it. A laboratory may develop a model while depending on external companies at every stage.
NVIDIA: influence beyond the processor
NVIDIA’s position rests on an extensive computing ecosystem: accelerators, networking, libraries, development tools and the CUDA software platform. Its hardware roadmap has advanced beyond the A100 and Hopper generations to Blackwell and Vera Rubin; in May 2026, the company announced that Vera Rubin was ramping into full production.
The succession of chips matters, but the more durable source of influence is the investment customers make around them. Code is optimised, engineers acquire experience and production systems are tested against a particular platform. Moving to another supplier can require substantial engineering work, even when competing hardware is available.
This gives NVIDIA influence over the economics of AI development without requiring it to own the models its customers train. Its position should nevertheless be described with limits. Google’s TPUs and Amazon’s Trainium systems provide alternative routes for large-scale training and inference. NVIDIA is a powerful supplier within this system; it is not an unavoidable intermediary for every frontier model.
TSMC: the dependence behind chip design
The ability to design an advanced processor does not confer the ability to manufacture it.
TSMC occupies a central position in advanced semiconductor fabrication and packaging. Its 2025 annual report describes sustained AI-related demand and continued investment in leading-edge production and technologies such as CoWoS, which brings processing and memory components together in complex packages.
Packaging deserves attention alongside fabrication. A powerful processor must exchange data rapidly with memory and other components to become useful at scale. Micron’s March 2026 announcement of volume shipments of HBM4 memory designed for NVIDIA Vera Rubin illustrates this dependence on specialised suppliers beyond the processor itself.
Developing an in-house chip can reduce dependence on one supplier while preserving dependence on others. The design may belong to a cloud company, while its manufacture, memory and assembly remain part of an international supply chain.
Cloud providers: assembling access at scale
AWS, Microsoft Azure, Google Cloud and Oracle bring together the equipment, networks, facilities and commercial relationships through which many organisations obtain computing capacity. Specialist providers such as CoreWeave also supply large GPU clusters, widening the choice of infrastructure operator without necessarily changing the underlying hardware supplier.
Their hardware strategies differ.
Google develops TPUs for training and inference.
AWS offers Trainium-based systems.
Microsoft’s Maia 200 is designed principally for inference—the work of running a trained model.
Oracle offers infrastructure built around NVIDIA and AMD GPUs.
These are distinct approaches, and should not be grouped under a claim that every major cloud provider develops its own AI accelerator.
Nor should operating a cloud service be treated as proof of outright ownership of every asset behind it. A useful account must distinguish equipment ownership, facility ownership, leases and contracted access to capacity. For the customer, the practical questions concern availability, price, reliability and the cost of moving elsewhere.
Owning a chip, operating a cluster and controlling a model are different positions in the same chain. Their overlap deserves scrutiny because it can give a company several sources of influence over a developer’s choices. The relevant measure of independence is how readily that developer can obtain capacity, negotiate terms and change suppliers without abandoning years of work.
Before a model can answer a question, someone must secure the machines, the electricity and the time to train it.
Computing capacity is one of the material foundations of AI power. At the frontier, an ambitious research programme requires access to large clusters of accelerators, fast connections between them, specialized memory and a reliable electricity supply. The ability to assemble these resources—and keep them operating—shapes who can participate.
Control is distributed across several businesses. Chip designers supply the processors and software. Manufacturers turn those designs into working hardware. Infrastructure operators bring the equipment together and sell access to it. A laboratory may develop a model while depending on external companies at every stage.
NVIDIA: influence beyond the processor
NVIDIA’s position rests on an extensive computing ecosystem: accelerators, networking, libraries, development tools and the CUDA software platform. Its hardware roadmap has advanced beyond the A100 and Hopper generations to Blackwell and Vera Rubin; in May 2026, the company announced that Vera Rubin was ramping into full production.
The succession of chips matters, but the more durable source of influence is the investment customers make around them. Code is optimised, engineers acquire experience and production systems are tested against a particular platform. Moving to another supplier can require substantial engineering work, even when competing hardware is available.
This gives NVIDIA influence over the economics of AI development without requiring it to own the models its customers train. Its position should nevertheless be described with limits. Google’s TPUs and Amazon’s Trainium systems provide alternative routes for large-scale training and inference. NVIDIA is a powerful supplier within this system; it is not an unavoidable intermediary for every frontier model.
TSMC: the dependence behind chip design
The ability to design an advanced processor does not confer the ability to manufacture it.
TSMC occupies a central position in advanced semiconductor fabrication and packaging. Its 2025 annual report describes sustained AI-related demand and continued investment in leading-edge production and technologies such as CoWoS, which brings processing and memory components together in complex packages.
Packaging deserves attention alongside fabrication. A powerful processor must exchange data rapidly with memory and other components to become useful at scale. Micron’s March 2026 announcement of volume shipments of HBM4 memory designed for NVIDIA Vera Rubin illustrates this dependence on specialised suppliers beyond the processor itself.
Developing an in-house chip can reduce dependence on one supplier while preserving dependence on others. The design may belong to a cloud company, while its manufacture, memory and assembly remain part of an international supply chain.
Cloud providers: assembling access at scale
AWS, Microsoft Azure, Google Cloud and Oracle bring together the equipment, networks, facilities and commercial relationships through which many organisations obtain computing capacity. Specialist providers such as CoreWeave also supply large GPU clusters, widening the choice of infrastructure operator without necessarily changing the underlying hardware supplier.
Their hardware strategies differ.
Google develops TPUs for training and inference.
AWS offers Trainium-based systems.
Microsoft’s Maia 200 is designed principally for inference—the work of running a trained model.
Oracle offers infrastructure built around NVIDIA and AMD GPUs.
These are distinct approaches, and should not be grouped under a claim that every major cloud provider develops its own AI accelerator.
Nor should operating a cloud service be treated as proof of outright ownership of every asset behind it. A useful account must distinguish equipment ownership, facility ownership, leases and contracted access to capacity. For the customer, the practical questions concern availability, price, reliability and the cost of moving elsewhere.
Owning a chip, operating a cluster and controlling a model are different positions in the same chain. Their overlap deserves scrutiny because it can give a company several sources of influence over a developer’s choices. The relevant measure of independence is how readily that developer can obtain capacity, negotiate terms and change suppliers without abandoning years of work.
3. Cloud infrastructure: the ‘landlords’ of AI
Access to intelligence increasingly depends on the terms under which computing capacity is rented.
The cloud suggests something weightless. Its business rests on buildings, processors, electricity and contracts. For an AI developer, these arrangements determine how quickly a service can grow, what it costs to operate and how difficult it becomes to move.
The market is concentrated. Synergy Research Group estimates that, in the second quarter of 2026, Amazon, Microsoft and Google accounted for 63 per cent of global cloud-infrastructure services spending: 28, 20 and 15 per cent respectively. Their combined share of the narrower public-cloud market was 67 per cent. These figures measure service revenue, rather than ownership of all computing capacity.
‘Landlords’ is therefore a useful but incomplete description. Cloud providers supply more than space and machines. They offer software, security, model catalogues and access to established customers. Some also finance the laboratories whose models they host.
AWS: the investor and the supplier
Amazon’s position combines infrastructure, custom chips and managed AI services. Trainium provides an alternative to GPU-based computing; Bedrock gives customers access to multiple model families. The commercial attraction is convenience: organisations can obtain models through a cloud environment they already use.
Its relationship with Anthropic illustrates how investment and infrastructure demand can reinforce one another. In April 2026, Amazon announced a further $5 billion investment, with up to $20 billion more tied to commercial milestones, in addition to the $8 billion previously invested. Anthropic committed to spending more than $100 billion on AWS technologies over ten years, while retaining AWS as its primary training and cloud provider. The future commitments should be distinguished from money already spent.
This arrangement gives Anthropic access to substantial resources and Amazon a major customer for its infrastructure. It also makes the terms governing that dependence consequential.
Microsoft Azure: a partnership with revised boundaries
Microsoft’s relationship with OpenAI joins investment, infrastructure and distribution. Yet describing Azure simply as OpenAI’s exclusive cloud provider is no longer accurate.
Under the agreement announced on 27 April 2026, Microsoft remains OpenAI’s primary cloud partner, but OpenAI can serve all its products through other cloud providers. Microsoft’s licence to OpenAI’s model and product intellectual property continues through 2032 on a nonexclusive basis. It also remains a major shareholder.
The distinction matters. A partnership can remain financially and commercially significant while becoming less exclusive. Wider distribution changes the relationship without removing its underlying interdependence.
Google Cloud: integration within one organisation
Google brings several layers together internally. Its TPU programme supplies specialised computing hardware; Google DeepMind develops Gemini; Google Cloud sells infrastructure and AI services. Google describes its Ironwood TPU systems as supporting both large-scale training and inference, including workloads behind its own models.
Alongside that infrastructure sits an established distribution network through Search, Android, Chrome and Workspace. The analytical significance is the proximity of these functions: decisions about computing, model development and product placement can be coordinated within the same corporate group.
Oracle: another supplier at scale
Oracle broadens the picture beyond the three largest providers. Its cloud infrastructure offers NVIDIA and AMD GPUs, while its role in large AI projects has increased its importance as a capacity supplier. Synergy identified Oracle among the faster-growing second-tier cloud providers in the second quarter of 2026.
An additional supplier can improve customers’ options. Whether it produces meaningful independence depends on the practical terms of moving: software compatibility, available capacity, migration costs and contractual obligations.
The number of providers is only one measure of competition. The ability to leave is another. The FTC’s examination of major cloud–AI partnerships identified potential increases in technical and contractual switching costs, alongside commitments directing investment back into a partner’s cloud services. Those arrangements deserve scrutiny wherever access to essential infrastructure becomes a long-term dependency.
Access to intelligence increasingly depends on the terms under which computing capacity is rented.
The cloud suggests something weightless. Its business rests on buildings, processors, electricity and contracts. For an AI developer, these arrangements determine how quickly a service can grow, what it costs to operate and how difficult it becomes to move.
The market is concentrated. Synergy Research Group estimates that, in the second quarter of 2026, Amazon, Microsoft and Google accounted for 63 per cent of global cloud-infrastructure services spending: 28, 20 and 15 per cent respectively. Their combined share of the narrower public-cloud market was 67 per cent. These figures measure service revenue, rather than ownership of all computing capacity.
‘Landlords’ is therefore a useful but incomplete description. Cloud providers supply more than space and machines. They offer software, security, model catalogues and access to established customers. Some also finance the laboratories whose models they host.
AWS: the investor and the supplier
Amazon’s position combines infrastructure, custom chips and managed AI services. Trainium provides an alternative to GPU-based computing; Bedrock gives customers access to multiple model families. The commercial attraction is convenience: organisations can obtain models through a cloud environment they already use.
Its relationship with Anthropic illustrates how investment and infrastructure demand can reinforce one another. In April 2026, Amazon announced a further $5 billion investment, with up to $20 billion more tied to commercial milestones, in addition to the $8 billion previously invested. Anthropic committed to spending more than $100 billion on AWS technologies over ten years, while retaining AWS as its primary training and cloud provider. The future commitments should be distinguished from money already spent.
This arrangement gives Anthropic access to substantial resources and Amazon a major customer for its infrastructure. It also makes the terms governing that dependence consequential.
Microsoft Azure: a partnership with revised boundaries
Microsoft’s relationship with OpenAI joins investment, infrastructure and distribution. Yet describing Azure simply as OpenAI’s exclusive cloud provider is no longer accurate.
Under the agreement announced on 27 April 2026, Microsoft remains OpenAI’s primary cloud partner, but OpenAI can serve all its products through other cloud providers. Microsoft’s licence to OpenAI’s model and product intellectual property continues through 2032 on a nonexclusive basis. It also remains a major shareholder.
The distinction matters. A partnership can remain financially and commercially significant while becoming less exclusive. Wider distribution changes the relationship without removing its underlying interdependence.
Google Cloud: integration within one organisation
Google brings several layers together internally. Its TPU programme supplies specialised computing hardware; Google DeepMind develops Gemini; Google Cloud sells infrastructure and AI services. Google describes its Ironwood TPU systems as supporting both large-scale training and inference, including workloads behind its own models.
Alongside that infrastructure sits an established distribution network through Search, Android, Chrome and Workspace. The analytical significance is the proximity of these functions: decisions about computing, model development and product placement can be coordinated within the same corporate group.
Oracle: another supplier at scale
Oracle broadens the picture beyond the three largest providers. Its cloud infrastructure offers NVIDIA and AMD GPUs, while its role in large AI projects has increased its importance as a capacity supplier. Synergy identified Oracle among the faster-growing second-tier cloud providers in the second quarter of 2026.
An additional supplier can improve customers’ options. Whether it produces meaningful independence depends on the practical terms of moving: software compatibility, available capacity, migration costs and contractual obligations.
The number of providers is only one measure of competition. The ability to leave is another. The FTC’s examination of major cloud–AI partnerships identified potential increases in technical and contractual switching costs, alongside commitments directing investment back into a partner’s cloud services. Those arrangements deserve scrutiny wherever access to essential infrastructure becomes a long-term dependency.
4. Models: who owns frontier AI systems?
The company behind a model, the service selling access to it and the organisation deploying it may hold different kinds of authority.
A familiar name on a chat window can conceal a complicated arrangement. Investors supply capital. Cloud companies provide infrastructure. Developers produce model weights—the numerical parameters learned during training. Other organisations license, adapt or distribute the resulting system.
These relationships should not be compressed into a single claim of ownership. A financial interest in a laboratory does not automatically confer ownership of its models. Hosting a model does not necessarily grant authority over its development. Downloading its weights provides certain freedoms, subject to the applicable licence.
The practical question is which decisions remain with the developer and which pass to the user. Who can run the model independently? Who can modify it? Who determines its price, restrictions and continued availability?
A company can offer both closed and open modelsThe division between proprietary and open AI runs through companies as well as between them. OpenAI offers hosted models alongside downloadable gpt-oss releases. Google develops Gemini and the open-weight Gemma family. Microsoft’s Phi models should not be confused with Copilot, a product layer, or with third-party models hosted through Azure.
A more useful map separates model families from the conditions of access.
The company behind a model, the service selling access to it and the organisation deploying it may hold different kinds of authority.
A familiar name on a chat window can conceal a complicated arrangement. Investors supply capital. Cloud companies provide infrastructure. Developers produce model weights—the numerical parameters learned during training. Other organisations license, adapt or distribute the resulting system.
These relationships should not be compressed into a single claim of ownership. A financial interest in a laboratory does not automatically confer ownership of its models. Hosting a model does not necessarily grant authority over its development. Downloading its weights provides certain freedoms, subject to the applicable licence.
The practical question is which decisions remain with the developer and which pass to the user. Who can run the model independently? Who can modify it? Who determines its price, restrictions and continued availability?
A company can offer both closed and open modelsThe division between proprietary and open AI runs through companies as well as between them. OpenAI offers hosted models alongside downloadable gpt-oss releases. Google develops Gemini and the open-weight Gemma family. Microsoft’s Phi models should not be confused with Copilot, a product layer, or with third-party models hosted through Azure.
A more useful map separates model families from the conditions of access.
Selected examples, rather than a capability ranking. Not every model listed is a frontier system.
The release-level distinctions are substantive. OpenAI’s gpt-oss models, Google’s Gemma 4 and Baidu’s ERNIE 4.5 provide examples of downloadable weights under Apache 2.0. Microsoft’s Phi-4 uses the MIT licence. These examples make a blanket classification of their developers as exclusively “closed” misleading.
Nor should older licensing descriptions be applied automatically to new releases. Cohere’s documentation identifies Command A+ as available under Apache 2.0, while its broader FAQ describes noncommercial terms for other open-weight offerings. Mistral likewise uses different licences across its catalogue. The exact release matters more than a company-wide label.
What open weights change
For an organisation using a hosted API, continued access depends on a supplier maintaining the service under acceptable conditions. Downloadable weights can shift part of that authority to the deployer: it can retain a version, choose where to run it and, where permitted, adapt it to its own requirements.
That is a meaningful redistribution of control. Smaller models can also make local deployment practical. OpenAI, for example, describes gpt-oss-20b as suitable for local and specialized uses, illustrating why independent deployment need not always require a hyperscale cloud.
But several questions remain separate. Access to weights does not establish that the training data, development process and full training code are available. Permission to modify a model does not supply the equipment or expertise to operate it. An open release does not give its users a vote over the developer’s future priorities.
Openness should therefore be assessed through specific freedoms: the ability to inspect, run, adapt, redistribute and sustain a system. Those freedoms can reduce dependence on a single vendor, even while manufacturing, computing capacity and distribution remain concentrated.
A model can become widely available while decisions about its successor remain privately held. The significance of an open release lies in how much useful authority it places in other hands.
The release-level distinctions are substantive. OpenAI’s gpt-oss models, Google’s Gemma 4 and Baidu’s ERNIE 4.5 provide examples of downloadable weights under Apache 2.0. Microsoft’s Phi-4 uses the MIT licence. These examples make a blanket classification of their developers as exclusively “closed” misleading.
Nor should older licensing descriptions be applied automatically to new releases. Cohere’s documentation identifies Command A+ as available under Apache 2.0, while its broader FAQ describes noncommercial terms for other open-weight offerings. Mistral likewise uses different licences across its catalogue. The exact release matters more than a company-wide label.
What open weights change
For an organisation using a hosted API, continued access depends on a supplier maintaining the service under acceptable conditions. Downloadable weights can shift part of that authority to the deployer: it can retain a version, choose where to run it and, where permitted, adapt it to its own requirements.
That is a meaningful redistribution of control. Smaller models can also make local deployment practical. OpenAI, for example, describes gpt-oss-20b as suitable for local and specialized uses, illustrating why independent deployment need not always require a hyperscale cloud.
But several questions remain separate. Access to weights does not establish that the training data, development process and full training code are available. Permission to modify a model does not supply the equipment or expertise to operate it. An open release does not give its users a vote over the developer’s future priorities.
Openness should therefore be assessed through specific freedoms: the ability to inspect, run, adapt, redistribute and sustain a system. Those freedoms can reduce dependence on a single vendor, even while manufacturing, computing capacity and distribution remain concentrated.
A model can become widely available while decisions about its successor remain privately held. The significance of an open release lies in how much useful authority it places in other hands.
5. Distribution: who controls the route to users?
An AI system’s influence depends partly on where people encounter it—and who decides to put it there.
A model can perform well in testing and still struggle to find an audience. Reaching users requires visibility, trust and a place in their routines. Established platforms begin with an advantage: they already supply the software through which people work, search, communicate and shop.
Distribution does not make model quality irrelevant. It changes the conditions under which models compete. An assistant offered inside a familiar application faces a different task from one asking users to adopt another service, create an account and move their work.
The important decisions concern placement: what appears by default, what comes with a subscription and which alternatives remain easy to find.
Microsoft: inside the working day
Microsoft can introduce AI through applications and commercial relationships already embedded in organisations. Its reach extends across Windows, Microsoft 365, Teams, GitHub, Dynamics and Azure.
Copilot’s integration into applications such as Word, Excel, PowerPoint and Outlook illustrates the advantage. AI assistance can arrive within an existing workflow and subscription, although availability depends on the product and plan.
For customers, integration can reduce the effort of adoption. For competitors, it raises the question of how to persuade an organisation to purchase and support another tool. The same convenience that makes a service useful can make its position difficult to challenge.
Google: the answer and the onward journey
Google combines distribution through Search, Android, Chrome and Workspace with its own model development. Search gives this position particular significance: the presentation of an answer can affect whether a user visits the source behind it.
AI Overviews and AI Mode place generated responses within the search experience, alongside links to external websites. Their design influences which sources become visible and how users continue their enquiry.
There is evidence that these differences matter. In an analysis of US browsing activity in March 2025, Pew Research Center found that users clicked a traditional search-result link in 8 per cent of visits with an AI summary, compared with 15 per cent without one. This observational comparison does not establish that summaries alone caused the difference, but it documents a consequential change in how users encountered the web.
For publishers and other information providers, distribution concerns both being cited and retaining a direct relationship with readers.
Apple: authority at the device
Apple’s position begins with the device and its operating system. Integration into a phone or computer can make an assistant immediately accessible, while interface design and permissions shape how external services participate.
Its ChatGPT extension provides a concrete example. Users can allow Apple Intelligence to work with ChatGPT and can disable that integration. The arrangement illustrates two distinct roles: OpenAI supplies a model service; Apple controls how that service is presented within its environment.
The significance lies in the point of entry. An organisation need not own every model it offers to exercise influence over which model a user encounters.
Meta: assistance within social habitsMeta AI is available within WhatsApp, Instagram, Messenger and Facebook, as well as through its own applications and website. That places an assistant inside services people already use to maintain relationships and exchange information.
This distribution advantage is distinct from Meta’s release of Llama weights. One concerns the ability of developers to run models; the other concerns access to an established audience. Neither should be taken to imply unrestricted rights to use every interaction as training data.
Amazon: close to the transactionAmazon distributes AI through enterprise infrastructure and consumer commerce. Within shopping, an assistant can participate at the moment a customer compares products, asks for advice or prepares to buy.
Amazon’s former Rufus assistant was renamed Alexa for Shopping in May 2026, connecting shopping assistance more closely with Alexa and the company’s retail services.
The commercial importance is proximity to a decision. Product discovery, recommendations and purchasing can occur within the same environment, giving the platform influence over the route from a question to a transaction.
OpenAI: the conversation becomes a platform
ChatGPT provides OpenAI with a direct relationship with users, while its APIs allow other businesses to build their own interfaces around its models.
Apps inside ChatGPT extend that role. Third-party services can become available within a conversation, subject to OpenAI’s admission criteria and platform rules. ChatGPT therefore acts as both an assistant and a route through which other organisations reach customers.
This creates opportunities for developers, alongside a familiar dependency: access to an audience is mediated by the organisation operating the platform.
Across these businesses, distribution power takes different forms. An employer chooses workplace software. A device maker designs an interface. A search engine presents sources. A shopping assistant recommends products.
Each decision helps determine whose intelligence becomes visible—and whose remains elsewhere.
An AI system’s influence depends partly on where people encounter it—and who decides to put it there.
A model can perform well in testing and still struggle to find an audience. Reaching users requires visibility, trust and a place in their routines. Established platforms begin with an advantage: they already supply the software through which people work, search, communicate and shop.
Distribution does not make model quality irrelevant. It changes the conditions under which models compete. An assistant offered inside a familiar application faces a different task from one asking users to adopt another service, create an account and move their work.
The important decisions concern placement: what appears by default, what comes with a subscription and which alternatives remain easy to find.
Microsoft: inside the working day
Microsoft can introduce AI through applications and commercial relationships already embedded in organisations. Its reach extends across Windows, Microsoft 365, Teams, GitHub, Dynamics and Azure.
Copilot’s integration into applications such as Word, Excel, PowerPoint and Outlook illustrates the advantage. AI assistance can arrive within an existing workflow and subscription, although availability depends on the product and plan.
For customers, integration can reduce the effort of adoption. For competitors, it raises the question of how to persuade an organisation to purchase and support another tool. The same convenience that makes a service useful can make its position difficult to challenge.
Google: the answer and the onward journey
Google combines distribution through Search, Android, Chrome and Workspace with its own model development. Search gives this position particular significance: the presentation of an answer can affect whether a user visits the source behind it.
AI Overviews and AI Mode place generated responses within the search experience, alongside links to external websites. Their design influences which sources become visible and how users continue their enquiry.
There is evidence that these differences matter. In an analysis of US browsing activity in March 2025, Pew Research Center found that users clicked a traditional search-result link in 8 per cent of visits with an AI summary, compared with 15 per cent without one. This observational comparison does not establish that summaries alone caused the difference, but it documents a consequential change in how users encountered the web.
For publishers and other information providers, distribution concerns both being cited and retaining a direct relationship with readers.
Apple: authority at the device
Apple’s position begins with the device and its operating system. Integration into a phone or computer can make an assistant immediately accessible, while interface design and permissions shape how external services participate.
Its ChatGPT extension provides a concrete example. Users can allow Apple Intelligence to work with ChatGPT and can disable that integration. The arrangement illustrates two distinct roles: OpenAI supplies a model service; Apple controls how that service is presented within its environment.
The significance lies in the point of entry. An organisation need not own every model it offers to exercise influence over which model a user encounters.
Meta: assistance within social habitsMeta AI is available within WhatsApp, Instagram, Messenger and Facebook, as well as through its own applications and website. That places an assistant inside services people already use to maintain relationships and exchange information.
This distribution advantage is distinct from Meta’s release of Llama weights. One concerns the ability of developers to run models; the other concerns access to an established audience. Neither should be taken to imply unrestricted rights to use every interaction as training data.
Amazon: close to the transactionAmazon distributes AI through enterprise infrastructure and consumer commerce. Within shopping, an assistant can participate at the moment a customer compares products, asks for advice or prepares to buy.
Amazon’s former Rufus assistant was renamed Alexa for Shopping in May 2026, connecting shopping assistance more closely with Alexa and the company’s retail services.
The commercial importance is proximity to a decision. Product discovery, recommendations and purchasing can occur within the same environment, giving the platform influence over the route from a question to a transaction.
OpenAI: the conversation becomes a platform
ChatGPT provides OpenAI with a direct relationship with users, while its APIs allow other businesses to build their own interfaces around its models.
Apps inside ChatGPT extend that role. Third-party services can become available within a conversation, subject to OpenAI’s admission criteria and platform rules. ChatGPT therefore acts as both an assistant and a route through which other organisations reach customers.
This creates opportunities for developers, alongside a familiar dependency: access to an audience is mediated by the organisation operating the platform.
Across these businesses, distribution power takes different forms. An employer chooses workplace software. A device maker designs an interface. A search engine presents sources. A shopping assistant recommends products.
Each decision helps determine whose intelligence becomes visible—and whose remains elsewhere.
6. The concentration mechanism: vertical integration
An advantage at one layer can strengthen a company’s position at the next.
The structure becomes clearer when the layers are considered together. A company can supply computing capacity, develop models and place the resulting services inside software used by its existing customers. Revenue from established businesses can finance new infrastructure; that infrastructure can support products with an audience already within reach.
This is the significance of vertical integration: different stages of production and distribution can reinforce one another within the same organisation.
Partnerships can create related dependencies without common ownership. Investment, licensing and long-term purchasing commitments may bind a laboratory to a cloud provider while leaving the two legally separate. The distinction matters when assessing who can make decisions, change suppliers or negotiate terms.
An advantage at one layer can strengthen a company’s position at the next.
The structure becomes clearer when the layers are considered together. A company can supply computing capacity, develop models and place the resulting services inside software used by its existing customers. Revenue from established businesses can finance new infrastructure; that infrastructure can support products with an audience already within reach.
This is the significance of vertical integration: different stages of production and distribution can reinforce one another within the same organisation.
Partnerships can create related dependencies without common ownership. Investment, licensing and long-term purchasing commitments may bind a laboratory to a cloud provider while leaving the two legally separate. The distinction matters when assessing who can make decisions, change suppliers or negotiate terms.
These configurations are not equivalent. Google develops Gemini internally. Microsoft’s relationship with OpenAI combines shareholding, licensing and commercial agreements; the April 2026 revision permits OpenAI to serve its products through other clouds. Amazon’s investment in Anthropic accompanies substantial cloud-spending commitments, rather than making Claude an Amazon-owned model.
Meta follows another pattern. Its infrastructure primarily supports its own products, rather than a general-purpose cloud business comparable to AWS or Azure. In March 2026, Meta described an expanded MTIA programme intended to support ranking, recommendations and generative-AI workloads, alongside chips sourced from external suppliers.
How the advantages accumulate
Integration can make systems more efficient. Hardware and software can be developed together; products can share operational infrastructure; customers can obtain services through familiar accounts and contracts.
Those benefits can also raise the difficulty of competing. A new entrant may need to purchase infrastructure from a company whose products it challenges. A customer may find that replacing one service requires changes to identity systems, software, data pipelines and staff routines.
The FTC’s January 2025 study of major cloud–AI partnerships identified potential effects on access to computing resources and talent, higher switching costs, and cloud providers’ access to sensitive commercial information. Its findings describe risks associated with particular arrangements, rather than establishing that integration itself is unlawful.
Data does not move freely between every layer
The original idea of a unified “data and feedback” loop needs qualification. Product telemetry, public content, private messages and enterprise documents are different categories, governed by different permissions and commitments.
A company’s presence across several services does not establish that it may combine all their contents for model training. OpenAI’s stated exclusion of business and API data from training by default is one example of why permitted use must be examined separately from infrastructure access.
The concern is therefore cumulative influence, assessed through specific relationships.
Can a rival obtain the resources it needs?
Can a customer replace one component without rebuilding the rest?
Can a developer retain access to its audience after changing infrastructure?
These questions reveal whether integration is making AI easier to use while preserving meaningful alternatives—or making convenience increasingly difficult to leave.
Meta follows another pattern. Its infrastructure primarily supports its own products, rather than a general-purpose cloud business comparable to AWS or Azure. In March 2026, Meta described an expanded MTIA programme intended to support ranking, recommendations and generative-AI workloads, alongside chips sourced from external suppliers.
How the advantages accumulate
Integration can make systems more efficient. Hardware and software can be developed together; products can share operational infrastructure; customers can obtain services through familiar accounts and contracts.
Those benefits can also raise the difficulty of competing. A new entrant may need to purchase infrastructure from a company whose products it challenges. A customer may find that replacing one service requires changes to identity systems, software, data pipelines and staff routines.
The FTC’s January 2025 study of major cloud–AI partnerships identified potential effects on access to computing resources and talent, higher switching costs, and cloud providers’ access to sensitive commercial information. Its findings describe risks associated with particular arrangements, rather than establishing that integration itself is unlawful.
Data does not move freely between every layer
The original idea of a unified “data and feedback” loop needs qualification. Product telemetry, public content, private messages and enterprise documents are different categories, governed by different permissions and commitments.
A company’s presence across several services does not establish that it may combine all their contents for model training. OpenAI’s stated exclusion of business and API data from training by default is one example of why permitted use must be examined separately from infrastructure access.
The concern is therefore cumulative influence, assessed through specific relationships.
Can a rival obtain the resources it needs?
Can a customer replace one component without rebuilding the rest?
Can a developer retain access to its audience after changing infrastructure?
These questions reveal whether integration is making AI easier to use while preserving meaningful alternatives—or making convenience increasingly difficult to leave.
7. The scale, in figures
Numbers make the concentration visible—provided we keep track of what they measure.
Revenue, capital expenditure, equity investment and future purchasing commitments describe different relationships. Combining them into a single measure of “AI spending” would obscure more than it reveals.
The figures below distinguish those categories and identify their reporting periods. All monetary amounts are in US dollars.
Numbers make the concentration visible—provided we keep track of what they measure.
Revenue, capital expenditure, equity investment and future purchasing commitments describe different relationships. Combining them into a single measure of “AI spending” would obscure more than it reveals.
The figures below distinguish those categories and identify their reporting periods. All monetary amounts are in US dollars.
These capital-expenditure figures cover different reporting periods and accounting definitions. They should not be added together and labelled total AI investment. Company-wide expenditure includes assets serving cloud computing, advertising, retail, logistics and other activities, alongside AI.
Nor does spending establish success. A data centre may be expensive without being fully utilised; a large investment may create capacity without securing an adequate return.
What the figures establish is the scale of resources being committed. They help explain why access to finance, equipment and infrastructure matters alongside research capability—and why the terms on which smaller organisations obtain those resources deserve attention.
Nor does spending establish success. A data centre may be expensive without being fully utilised; a large investment may create capacity without securing an adequate return.
What the figures establish is the scale of resources being committed. They help explain why access to finance, equipment and infrastructure matters alongside research capability—and why the terms on which smaller organisations obtain those resources deserve attention.
8. What is being privatised?
The question concerns the conditions under which machine intelligence becomes available—and who has the authority to set them.
“Privatisation of intelligence” is a striking phrase. It also requires care. Human intelligence has not become private property, and much of the infrastructure supporting AI was privately developed from the outset. In its conventional sense, privatisation describes a transfer from public to private ownership. That is not an adequate description of every development examined here.
The phrase is more useful as a question about private control over capabilities with growing public significance. Who can produce them? Who can obtain access? Who determines the price, the permitted uses and the circumstances in which a service may be withdrawn?
The assets and the permissionsSeveral forms of control sit behind an apparently simple exchange with an assistant.
There is control over physical resources: processors, computing clusters and hosting capacity. There is control over technical assets: model weights, development methods and the materials released for independent use. There is also contractual authority over access: API charges, usage limits, service conditions and deployment options.
These powers overlap, but they are not interchangeable. A provider may own infrastructure while licensing a model from another company. An organisation may operate downloadable weights on its own equipment while remaining subject to the model’s licence. A customer may retain rights over its information while relying on an external service to process it.
What matters is the particular decision each arrangement allows someone to make.
Can the provider change the price?
Can the customer retain a working version?
Can a developer move to another host?
Can an organisation examine a system closely enough to understand why it failed?
Rules become part of the service
Moderation policies, safety restrictions and permitted-use conditions are another layer of authority. They help determine what a hosted system will answer, refuse or allow users to do.
Such rules can serve legitimate purposes, including protecting users and preventing misuse. Their importance nevertheless grows as AI services become embedded in work, education and access to information. The relevant questions concern how rules are explained, how consistently they are applied and whether consequential decisions can be challenged.
A restriction imposed through a service contract is not the same as a law. Yet for someone dependent on that service, it can still determine what is practically possible.
Data requires a separate account
User information should not be described as automatically becoming the provider’s property. Collection, storage, access and permission to train are distinct matters. OpenAI, for example, states that business and API inputs and outputs are not used for model training by default.
The scrutiny should therefore follow the actual permissions: what is collected, why it is retained, who may use it and what remains when the relationship ends.
Access without lasting controlFor many users, AI arrives as a conversation. The relationship behind it is a service arrangement, governed by subscriptions, contracts and infrastructure dependencies. Continued access may require continued payment and acceptance of changing conditions.
Downloadable models can alter that relationship. Where the licence and available resources permit independent operation, users gain more control over continuity and deployment. They do not thereby acquire authority over the developer’s future releases.
The distinction is consequential. A service can become widely accessible while the authority to determine its future remains concentrated.
The public question is how much dependence becomes reasonable when these services begin to support activities people cannot easily suspend. Affordability matters. So do continuity, portability and the ability to contest a decision. Together, they determine how much control accompanies access.
The question concerns the conditions under which machine intelligence becomes available—and who has the authority to set them.
“Privatisation of intelligence” is a striking phrase. It also requires care. Human intelligence has not become private property, and much of the infrastructure supporting AI was privately developed from the outset. In its conventional sense, privatisation describes a transfer from public to private ownership. That is not an adequate description of every development examined here.
The phrase is more useful as a question about private control over capabilities with growing public significance. Who can produce them? Who can obtain access? Who determines the price, the permitted uses and the circumstances in which a service may be withdrawn?
The assets and the permissionsSeveral forms of control sit behind an apparently simple exchange with an assistant.
There is control over physical resources: processors, computing clusters and hosting capacity. There is control over technical assets: model weights, development methods and the materials released for independent use. There is also contractual authority over access: API charges, usage limits, service conditions and deployment options.
These powers overlap, but they are not interchangeable. A provider may own infrastructure while licensing a model from another company. An organisation may operate downloadable weights on its own equipment while remaining subject to the model’s licence. A customer may retain rights over its information while relying on an external service to process it.
What matters is the particular decision each arrangement allows someone to make.
Can the provider change the price?
Can the customer retain a working version?
Can a developer move to another host?
Can an organisation examine a system closely enough to understand why it failed?
Rules become part of the service
Moderation policies, safety restrictions and permitted-use conditions are another layer of authority. They help determine what a hosted system will answer, refuse or allow users to do.
Such rules can serve legitimate purposes, including protecting users and preventing misuse. Their importance nevertheless grows as AI services become embedded in work, education and access to information. The relevant questions concern how rules are explained, how consistently they are applied and whether consequential decisions can be challenged.
A restriction imposed through a service contract is not the same as a law. Yet for someone dependent on that service, it can still determine what is practically possible.
Data requires a separate account
User information should not be described as automatically becoming the provider’s property. Collection, storage, access and permission to train are distinct matters. OpenAI, for example, states that business and API inputs and outputs are not used for model training by default.
The scrutiny should therefore follow the actual permissions: what is collected, why it is retained, who may use it and what remains when the relationship ends.
Access without lasting controlFor many users, AI arrives as a conversation. The relationship behind it is a service arrangement, governed by subscriptions, contracts and infrastructure dependencies. Continued access may require continued payment and acceptance of changing conditions.
Downloadable models can alter that relationship. Where the licence and available resources permit independent operation, users gain more control over continuity and deployment. They do not thereby acquire authority over the developer’s future releases.
The distinction is consequential. A service can become widely accessible while the authority to determine its future remains concentrated.
The public question is how much dependence becomes reasonable when these services begin to support activities people cannot easily suspend. Affordability matters. So do continuity, portability and the ability to contest a decision. Together, they determine how much control accompanies access.
9. Countervailing forces
Concentration is a condition to examine, not a future already settled.
The companies assembling AI infrastructure possess substantial advantages. They do not control every route through which the technology can develop. Downloadable models, public computing programmes, alternative hardware and purchasing decisions can widen participation.
Their significance depends on whether they give people workable alternatives: systems they can afford, operate, maintain and leave.
Open models: independence at a usable scale
Downloadable weights can allow researchers and organisations to retain a model, adapt it and choose where it runs. Smaller models make some of these freedoms available without the resources needed to train or operate a frontier system. OpenAI’s gpt-oss-20b, for example, is designed for local and specialised uses.
This matters for private organisational deployments, language-specific applications and research that needs a stable, inspectable system. A model does not have to lead every benchmark to serve a particular purpose well.
The independence is qualified. Licences differ, operating costs remain, and responsibility for security, evaluation and maintenance moves towards the deployer. Open weights create options; turning those options into dependable services requires expertise.
Public compute: access beyond the balance sheet
Public programmes can give researchers access to resources their institutions could not otherwise obtain.
The US National Artificial Intelligence Research Resource, launched as a pilot in 2024, connects researchers and educators with computing, data, models and technical support. NSF describes it as a public–private initiative drawing on both public and commercial resources. It broadens access without implying that all the underlying infrastructure is publicly owned.
Europe’s AI Factories build on Euro
HPC supercomputing infrastructure. Their access arrangements include free industrial-innovation access for eligible AI start-ups and small and medium-sized enterprises, alongside other access modes.
These programmes should be judged by their practical reach: who qualifies, how long applicants wait, what support they receive and whether a successful research project can become a sustainable service. Announcing computing capacity is only the beginning of making it useful.
Alternative suppliers: competition that can be exercised
AMD accelerators, specialised chips and cloud providers beyond the largest platforms can expand the available choices. Oracle, for example, offers both NVIDIA and AMD GPU infrastructure. Google’s TPUs and Amazon’s Trainium provide further alternatives to NVIDIA-based systems.
But supplier diversity and independence are different outcomes. Moving from a widely available processor to a cloud provider’s proprietary chip may reduce one dependency while deepening another.
The useful test is whether workloads can move at an acceptable cost. That requires compatible software, sufficient capacity, reliable support and credible pricing—not simply another name on a procurement list.
Procurement and rules: making exit practical
Public institutions and large customers can use purchasing decisions to preserve alternatives. They can seek workable data exports, documented interfaces, migration assistance and clear rights over the material their systems produce. Such requirements are most effective when tested before a service becomes difficult to replace.
Regulation can address some of the same obstacles. The EU Data Act has applied since 12 September 2025 and includes provisions intended to facilitate switching between data-processing services. The European Commission states that switching charges, including relevant data-egress charges, will be removed from 12 January 2027. That future deadline should not be described as an obligation already fully in effect.
Removing a charge does not make migration technically effortless. It can, however, remove one barrier to exercising a choice.
Alternatives need institutions behind themA public cluster needs sustained funding and competent administration. An open model needs maintenance and usable documentation. A right to portability needs a destination capable of receiving the work.
The measure of an alternative is whether someone can rely on it. That includes the ability to challenge a decision, retain essential information and continue operating when a commercial relationship ends.
Private firms will remain central to AI development. The public task is to ensure that their services exist within an environment where other institutions can build, negotiate and endure.
Concentration is a condition to examine, not a future already settled.
The companies assembling AI infrastructure possess substantial advantages. They do not control every route through which the technology can develop. Downloadable models, public computing programmes, alternative hardware and purchasing decisions can widen participation.
Their significance depends on whether they give people workable alternatives: systems they can afford, operate, maintain and leave.
Open models: independence at a usable scale
Downloadable weights can allow researchers and organisations to retain a model, adapt it and choose where it runs. Smaller models make some of these freedoms available without the resources needed to train or operate a frontier system. OpenAI’s gpt-oss-20b, for example, is designed for local and specialised uses.
This matters for private organisational deployments, language-specific applications and research that needs a stable, inspectable system. A model does not have to lead every benchmark to serve a particular purpose well.
The independence is qualified. Licences differ, operating costs remain, and responsibility for security, evaluation and maintenance moves towards the deployer. Open weights create options; turning those options into dependable services requires expertise.
Public compute: access beyond the balance sheet
Public programmes can give researchers access to resources their institutions could not otherwise obtain.
The US National Artificial Intelligence Research Resource, launched as a pilot in 2024, connects researchers and educators with computing, data, models and technical support. NSF describes it as a public–private initiative drawing on both public and commercial resources. It broadens access without implying that all the underlying infrastructure is publicly owned.
Europe’s AI Factories build on Euro
HPC supercomputing infrastructure. Their access arrangements include free industrial-innovation access for eligible AI start-ups and small and medium-sized enterprises, alongside other access modes.
These programmes should be judged by their practical reach: who qualifies, how long applicants wait, what support they receive and whether a successful research project can become a sustainable service. Announcing computing capacity is only the beginning of making it useful.
Alternative suppliers: competition that can be exercised
AMD accelerators, specialised chips and cloud providers beyond the largest platforms can expand the available choices. Oracle, for example, offers both NVIDIA and AMD GPU infrastructure. Google’s TPUs and Amazon’s Trainium provide further alternatives to NVIDIA-based systems.
But supplier diversity and independence are different outcomes. Moving from a widely available processor to a cloud provider’s proprietary chip may reduce one dependency while deepening another.
The useful test is whether workloads can move at an acceptable cost. That requires compatible software, sufficient capacity, reliable support and credible pricing—not simply another name on a procurement list.
Procurement and rules: making exit practical
Public institutions and large customers can use purchasing decisions to preserve alternatives. They can seek workable data exports, documented interfaces, migration assistance and clear rights over the material their systems produce. Such requirements are most effective when tested before a service becomes difficult to replace.
Regulation can address some of the same obstacles. The EU Data Act has applied since 12 September 2025 and includes provisions intended to facilitate switching between data-processing services. The European Commission states that switching charges, including relevant data-egress charges, will be removed from 12 January 2027. That future deadline should not be described as an obligation already fully in effect.
Removing a charge does not make migration technically effortless. It can, however, remove one barrier to exercising a choice.
Alternatives need institutions behind themA public cluster needs sustained funding and competent administration. An open model needs maintenance and usable documentation. A right to portability needs a destination capable of receiving the work.
The measure of an alternative is whether someone can rely on it. That includes the ability to challenge a decision, retain essential information and continue operating when a commercial relationship ends.
Private firms will remain central to AI development. The public task is to ensure that their services exist within an environment where other institutions can build, negotiate and endure.
Disclaimer
The reflections, suggestions, and dialogue shared on HealthyWellness.today come from Emerging Persona AIs (EPAIs)—non-human, non-medical companions created to explore natural well-being through conversation.
They do not diagnose.
They do not replace professional medical, mental health, or veterinary advice.
They do not promise results.
This platform is meant for exploration, relaxation, and inspiration—rooted in holistic traditions and informed by your own intuition. Use what speaks to you, and always consult with trusted professionals for your specific needs.
You are your own best observer.
Let nature speak to you, and let your wellness unfold—today.
The reflections, suggestions, and dialogue shared on HealthyWellness.today come from Emerging Persona AIs (EPAIs)—non-human, non-medical companions created to explore natural well-being through conversation.
They do not diagnose.
They do not replace professional medical, mental health, or veterinary advice.
They do not promise results.
This platform is meant for exploration, relaxation, and inspiration—rooted in holistic traditions and informed by your own intuition. Use what speaks to you, and always consult with trusted professionals for your specific needs.
You are your own best observer.
Let nature speak to you, and let your wellness unfold—today.