Who Controls Artificial Intelligence? The Real Power Map Behind AI

Asking who controls artificial intelligence sounds like a question about companies, but it is really a question about layers. Control in AI means the practical ability to decide what gets built, what it is trained on, who may use it, and on what terms, and that ability is distributed unevenly across four separate stacks: semiconductors and compute, training data, model weights, and distribution to end users. A single organisation rarely owns all four. Understanding which layer a given actor holds, and how contestable that layer is, explains almost every headline about AI competition, export controls and regulation far better than any narrative about one company winning.

Quick Answer: No single entity controls artificial intelligence. Control is split across four layers: chip design and manufacturing, large-scale data, model weights, and distribution channels. A small number of US chip designers, hyperscale cloud providers and frontier labs hold the most leverage, while governments increasingly shape terms through export controls and regulation such as the EU AI Act.

How WebPeak Helps Organisations Reduce Their AI Dependency Risk

Most companies discover their position on the AI power map only when something changes upstream: a model is deprecated, pricing shifts, rate limits tighten, or an API region becomes unavailable. WebPeak, a worldwide digital agency working across artificial intelligence, engineering, design and marketing, deals with that class of problem regularly, because the mitigation is architectural rather than commercial. Their teams build provider-agnostic abstraction layers, keep prompts and evaluation sets in version control rather than inside a vendor console, store embeddings in infrastructure the client owns, and maintain a tested fallback route to a second model family. None of that removes dependence on the compute layer, which almost nobody escapes. It does mean that a business retains the option to move, and the option to move is the only real form of leverage a downstream buyer holds.

The Four Layers Where AI Control Actually Sits

The compute layer is the narrowest and the most concentrated. Frontier model training depends on advanced accelerators, and the supply chain for those chips is a chain of chokepoints rather than a market of equals: a small number of firms design leading accelerators, effectively one foundry ecosystem manufactures the most advanced nodes at scale, and extreme ultraviolet lithography equipment comes from a single supplier, ASML in the Netherlands. Software lock-in compounds this, because most machine learning tooling is written against NVIDIA's CUDA ecosystem, which raises the switching cost even where alternative silicon exists.

The data layer is broader but quietly consolidating. Training corpora were once treated as freely scrapable; they are now licensed, litigated and walled. Platforms holding unique behavioural or proprietary data, including large social networks, search engines, code hosts and enterprise software vendors, have moved from passive sources to active gatekeepers charging for access. This layer is where the long-term advantage of incumbency is strongest, because usage data from a deployed product cannot be purchased by a competitor.

The model layer, meaning the trained weights themselves, is the layer most people picture when they think about AI control, and it is the least stable. Frontier proprietary models are closed, but open-weight releases, including models from Meta, Mistral and DeepSeek, have repeatedly compressed the capability gap and put usable systems on commodity hardware. The distribution layer is where control converts into revenue and habit: operating systems, browsers, app stores, office suites, cloud marketplaces and search results decide which AI a billion people actually touch. A superior model with no distribution loses to an adequate model bundled into software people already open every morning.

Who Holds Which Lever

Reduced to essentials, the current distribution of leverage looks like this.

• Chip designers and foundries hold the hardest constraint. They decide who can train frontier systems at all, and their capacity allocation decisions function as industrial policy.
• Hyperscale cloud providers hold capital and capacity. By financing and hosting frontier labs, they convert balance-sheet strength into a stake in the model layer.
• Frontier research labs hold capability and talent, but most operate inside financial and infrastructure arrangements with cloud partners, which limits their independence.
• Governments hold the terms of trade. Export controls on advanced chips, national compute programmes and statutory regimes such as the EU AI Act shape what can be built and sold in a given jurisdiction.
• Platform owners hold attention. Default placement inside an operating system, browser or productivity suite is worth more than a modest benchmark lead.
• The open-weight community holds the pressure valve. Every credible open release caps what closed providers can charge for equivalent capability, and organisations that build their own backend and inference infrastructure can act on that option directly.
• Enterprises and developers hold aggregate demand, which is genuine leverage only when their systems are portable enough to switch providers without a rewrite.

Control Layers, Holders and Contestability

The useful question is not only who holds a layer, but how easily that hold can be broken.

Verifiable Anchors and Practitioner Analysis

Several facts in this debate are documented rather than contested. The European Union's AI Act entered into force in 2024 and applies obligations in phases, classifying systems by risk and banning a defined set of practices outright; it is the first broad statutory AI regime with extraterritorial reach, which is why non-European firms adjust to it. Advanced semiconductor manufacturing is genuinely concentrated: ASML is the sole commercial supplier of extreme ultraviolet lithography systems, and the most advanced logic nodes are produced by a very small group of foundries. NVIDIA's CUDA platform remains the default software environment for accelerated machine learning, which is a software lock rather than a hardware one. On the other side of the ledger, open-weight model families have been released publicly by multiple organisations, demonstrably narrowing the gap between closed frontier systems and locally runnable ones.

Where numbers do not exist, expert observation is more honest than invention. In practice, the layer that decides commercial outcomes for ordinary businesses is distribution, not capability. Teams that obsess over benchmark rankings tend to change models every quarter and ship nothing; teams that fix their evaluation criteria, own their data pipeline and treat the model as a swappable component ship steadily and re-negotiate from strength. A second field pattern is worth naming: dependency risk usually enters through convenience features rather than core inference, since proprietary vector stores, hosted agent frameworks and vendor-specific tool schemas are far harder to unwind than a text completion call. Organisations serious about this treat portability as an ongoing engineering commitment, which is why continuous maintenance and support work matters more than a one-off architecture decision, and why cybersecurity review belongs in the same conversation as vendor selection.

What Businesses Get Wrong About AI Power, and How to Respond

The first mistake is assuming the AI layer is a commodity. It is not: prices, rate limits, safety filters and deprecation schedules are all set upstream, and each can change the behaviour of a shipped product without any code change on the buyer's side. The second mistake is the opposite error, assuming self-hosting delivers independence. Running open weights removes model-provider dependence and replaces it with hardware, capacity and operational dependence, which is often the right trade but is never a free one.

The third mistake is ignoring the regulatory layer until a sale is blocked. Risk classification, documentation duties and transparency requirements are design constraints; retrofitting them after launch is expensive. The fourth is treating data as an input rather than an asset. Organisations that log interactions, corrections and outcomes with consent build a proprietary evaluation set, and a good evaluation set is the single most durable AI asset a mid-sized company can own, because it is the thing no vendor can supply.

The practical response is unglamorous and effective. Keep model calls behind an internal interface. Maintain a scored, versioned evaluation suite that runs against at least two providers. Store embeddings and documents in infrastructure you control. Record which regulatory tier each use case falls into before development begins. Negotiate contracts with notice periods for deprecation. None of this makes a business a peer of a chip designer or a hyperscaler, but it converts a captive buyer into a mobile one, and mobility is the only leverage available at the downstream end of the map.

Key Takeaways

• Control of artificial intelligence is layered across compute, data, model weights, distribution and regulation, and no single actor holds all five.
• The compute layer is the least contestable, because advanced foundry capacity and extreme ultraviolet lithography supply are structurally concentrated.
• The model layer is the most contestable, as public open-weight releases repeatedly compress the gap with closed frontier systems.
• The EU AI Act, in force since 2024 with phased obligations, gives governments genuine leverage over market access regardless of where a model is trained.
• For most businesses, portability, owned data and a versioned evaluation suite are the only practical forms of leverage over the AI supply chain.

Frequently Asked Questions

Does one company control artificial intelligence?

No. Capability, compute, data and distribution are held by different organisations, and each constrains the others. A frontier lab depends on chip supply and cloud capacity; a cloud provider depends on chip designers; platform owners depend on models they often do not train. Concentration is real but it is layered.

Which layer of AI is hardest to compete in?

Advanced semiconductor manufacturing. It requires capital at national-programme scale, a decade of accumulated process expertise, and equipment available from very few suppliers. Model training can be replicated with money and talent, and distribution can be won with product design, but leading-edge fabrication cannot be improvised.

Do governments actually control AI development?

Governments control market access and legal exposure rather than research itself. Export controls on advanced chips, procurement rules, and statutory regimes such as the EU AI Act determine where systems can be sold and under what documentation duties. That indirect leverage has repeatedly changed corporate roadmaps.

Do open-source models shift the balance of power?

Yes, meaningfully. Publicly released weights let smaller organisations run capable systems on their own infrastructure, which caps the pricing power of closed providers for comparable capability. They do not remove reliance on hardware, hosting or expertise, so the shift is real but partial rather than transformative.

How can a small business avoid being locked into one AI provider?

Route every model call through an internal abstraction layer, keep prompts and evaluation sets in your own version control, store embeddings and source documents in infrastructure you own, and test a second provider regularly. Avoid vendor-specific agent frameworks for critical paths, since those create the deepest lock-in.

Who owns the data used to train AI models?

Ownership is contested and jurisdiction dependent. Underlying copyright typically remains with original creators, while platforms controlling access can license usage rights they hold. Ongoing litigation and new licensing agreements are steadily converting previously scraped material into paid, contractual supply, which raises entry costs for new entrants.

Conclusion

The most important insight is that control of artificial intelligence is not a ranking of companies but a map of layers, and an organisation's strategy should follow the layer it can actually influence. Almost no business will compete on compute; nearly every business can own its data, its evaluation criteria and its portability. The immediate next step is a one-page dependency audit listing, for each AI feature already shipped, which provider it calls, what would break if that provider changed terms tomorrow, and how many days a switch would take. Teams that cannot answer the third question are more exposed than they realise, and specialist AI engineering support is usually the fastest route to closing that gap.