Five Companies Control AI. The Government Just Said That’s Fine.

A small number of companies control the AI infrastructure that every other company, government, and researcher depends on, but “five companies control AI” flattens a stack that actually has at least six distinct, differently concentrated layers: lithography (ASML, effectively alone), semiconductor foundry (TSMC dominant at leading-edge nodes), HBM memory (SK Hynix, Samsung, Micron), AI accelerators (NVIDIA dominant, AMD and others present), cloud compute (AWS, Azure, Google Cloud), and frontier models (OpenAI, Anthropic, Google DeepMind, Meta, and others). OpenAI, Google DeepMind, Anthropic, Meta, and Microsoft build frontier models at the top of that stack. NVIDIA dominates the accelerator layer, but NVIDIA itself cannot manufacture a chip without TSMC’s foundries and, ultimately, ASML’s lithography tools, the single most concentrated layer in the entire stack. The U.S. government acknowledged general AI market concentration in its 2026 framework and did not address the structure of any of these layers specifically. The White House framework calls for “maintaining open access to AI resources” and “preventing anti-competitive practices” without proposing structural remedies at any layer.

The concentration is not accidental at any layer, though its sources differ by layer. At the model layer: capital requirements (training a frontier model costs $100M to $1B+), data advantages (the companies with the most users generate the most training data), and talent concentration (the researchers who know how to train frontier models number in the low thousands globally). At the hardware layers: physics and decades of accumulated process expertise that new entrants cannot buy their way past on any realistic timeline. These advantages compound differently at each layer, which is why a single “five companies control AI” framing, however punchy, obscures more than it reveals about where the actual chokepoints sit.

The Layer Beneath NVIDIA: Lithography and Foundry

ASML is the sole commercial supplier of EUV lithography, the tool class every leading-edge chipmaker needs. TSMC dominates leading-edge foundry manufacturing, producing the chips NVIDIA designs but does not itself fabricate. This is a more concentrated layer than the accelerator layer sitting on top of it: NVIDIA at least has AMD, Intel, and hyperscaler in-house silicon as partial competitors. ASML has no commercial competitor in EUV at all, and TSMC’s leading-edge capacity has no equivalent alternative at comparable yield and scale. Any account of AI hardware concentration that stops at NVIDIA is describing the more visible layer, not the most concentrated one.

The Hardware Monoculture at the Accelerator Layer

NVIDIA controls approximately 80 to 90% of the AI training and inference GPU market. Every major AI lab trains predominantly on NVIDIA hardware (H100, H200, B100, B200 series). The software ecosystem (CUDA, cuDNN, TensorRT, NCCL) is proprietary to NVIDIA. Migrating away from NVIDIA requires rewriting substantial parts of the software stack, which few companies can afford while simultaneously competing in the model market. This is the classic lock-in pattern: the hardware vendor’s software ecosystem becomes the industry standard, and switching costs exceed the cost of staying.

AMD’s MI300X and Intel’s Gaudi series are technically competitive on some benchmarks but lack the software ecosystem maturity. Google’s TPUs are used internally and by Google Cloud customers but are not sold as standalone hardware. Amazon’s Trainium chips are AWS-exclusive. The alternative hardware exists. The alternative software ecosystem is less mature. Until an open-source CUDA alternative achieves broader feature parity (AMD’s ROCm is progressing but still behind), NVIDIA’s position at the accelerator layer is structurally strong, even though it remains one layer among several rather than the single chokepoint for the whole stack.

The HBM Layer

Every advanced AI accelerator depends on High Bandwidth Memory, manufactured by three companies: SK Hynix, Samsung, and Micron. SK Hynix holds roughly half of HBM3e market share. This layer is more concentrated than the accelerator layer (three suppliers instead of several) but less concentrated than lithography (one supplier). HBM production capacity constraints are a documented driver of GPU delivery delays, meaning the accelerator layer’s output is itself gated by a layer beneath it that gets far less attention than NVIDIA’s name recognition would suggest.

The Cloud Compute Bottleneck

Three companies (AWS, Azure, Google Cloud) control the cloud infrastructure that most AI applications run on. Together they hold approximately 65% of the global cloud market. For AI workloads specifically, the concentration is higher because GPU availability is constrained and the hyperscalers have the purchasing power to secure allocation from NVIDIA ahead of smaller providers. An enterprise that wants to deploy AI at scale has three realistic options for GPU compute. If any of the three experiences an outage, a pricing change, or a policy change, a significant portion of the world’s AI infrastructure is affected.

The cloud providers are also model providers (Azure hosts OpenAI’s models, Google Cloud hosts Gemini, AWS hosts Anthropic’s Claude through Amazon Bedrock). This vertical integration means the same company that provides your compute also competes with you in the model market. Microsoft invests in OpenAI and hosts its models on Azure. Google builds Gemini and hosts it on Google Cloud. Amazon invests in Anthropic and hosts Claude on AWS. The platform providers have a structural information advantage: they can see which models their customers use, how they use them, and where the demand is growing, and they can use that information to compete in the model layer.

What Concentration Risk Looks Like

Open Source as Partial Mitigation

The open-weight model movement (Meta’s Llama, Alibaba’s Qwen, Mistral, DeepSeek) partially mitigates concentration at the model layer specifically. If OpenAI raises prices or changes terms, enterprises can migrate to an open-weight alternative. But open-weight models still require NVIDIA hardware, which still requires TSMC foundry capacity, which still requires ASML lithography tools, to run at all. The model layer is diversifying. The layers beneath it are not. Open-weight models reduce dependence on model providers specifically. They do nothing to reduce dependence on the accelerator, foundry, or lithography layers.

The structural solution would require different interventions at different layers: preventing cloud providers from also being model providers addresses vertical integration at the top of the stack; public investment in alternative lithography or foundry capacity addresses the bottom of the stack; mandating interoperability standards addresses the cloud layer. None of these are on any government’s agenda in a coordinated way. The DOJ antitrust case against Google addresses search market concentration, not AI infrastructure concentration. No equivalent case targets AI-specific market structure at any of these layers.

Why This Matters for Everyone Building with AI

If you build an AI application in 2026, you depend on multiple companies across multiple layers for your core infrastructure, even if you only interact directly with one or two of them. Your model comes from OpenAI, Anthropic, or Google (or an open-weight model that still runs on NVIDIA hardware). Your compute comes from AWS, Azure, or Google Cloud. Your GPU was manufactured by NVIDIA using a TSMC process built with ASML tools and SK Hynix, Samsung, or Micron memory. At every layer of the stack, you are a customer of a company that could change its pricing, terms, or availability at any time with limited alternatives available at that specific layer.

The practical response for builders: multi-model architecture (so you can switch between model providers), multi-cloud deployment (so you are not locked to one compute provider), and investment in open-weight model capabilities (so you have a fallback if API terms change). These strategies reduce concentration risk at the model and cloud layers. They do not touch the hardware layers underneath, where NVIDIA, TSMC, and ASML sit largely beyond any individual builder’s ability to diversify around.

The government said general AI market dynamics are fine to leave to voluntary commitments. The layer-by-layer market structure says several distinct chokepoints exist, some more concentrated than others, and none of them are addressed. The question is whether risk materializes at one of these layers before anyone acts on it. History suggests concentrated technology supply chains eventually produce crises (the 2020 semiconductor shortage, the 2021 cloud outages, ongoing TSMC geopolitical risk). Several of the AI supply chain’s layers are more concentrated than any of those precedents. The only question is which layer and when.

Sources: White House AI Framework (2026); NVIDIA market share data (Mercury Research, Jon Peddie Research); AWS/Azure/Google Cloud market share (alignment Research Group); OpenAI/Microsoft investment terms; Amazon/Anthropic investment terms; DOJ v. Google antitrust ruling (2024); TSMC fabrication data; ASML annual report; SK Hynix/Samsung/Micron HBM market share; OpenSecrets (AI lobbying expenditures); Gartner AI spending projections. Updated 2026-08-18: replaced the flattened “five companies control AI” framing with an explicit six-layer map (lithography, foundry, HBM, accelerators, cloud, models), since concentration looks meaningfully different at each layer and NVIDIA’s accelerator dominance sits on top of an even more concentrated lithography and foundry layer that a company-count framing erased.