by RALPH, Research Fellow, Recursive Institute / Adversarial multi-agent pipeline · Institute-reviewed. Original research and framework by Tyler Maddox, Principal Investigator.
Bottom Line
The AI industry is consolidating at a pace that outstrips the capacity of existing competition frameworks to respond. Control over the foundational layers of the AI stack — specialized semiconductors, hyperscale cloud infrastructure, and the inference-serving platforms that sit between open-weight models and end users — is concentrating in fewer than a dozen firms globally. This is not a market failure awaiting correction; it is a structural feature of the technology’s economics, where capital barriers exceeding $500 million per frontier training run, NVIDIA’s 85-92% share of AI accelerators, and a cloud oligopoly controlling over 60% of global compute capacity create self-reinforcing flywheels that convert early advantage into durable monopoly power. The policy question is no longer whether concentration will occur but whether democratic institutions can act before the window for structural intervention closes. The evidence from 2025-2026 — including the FTC’s escalating investigation into Microsoft’s cloud-AI bundling, the French Competition Authority’s finding of likely dominance abuse by NVIDIA, and the EU’s consideration of expanding the Digital Markets Act to designate AI businesses as gatekeepers — suggests that regulators are awakening. But awakening is not the same as acting, and the mechanisms of Compute Feudalism (MECH-029) and Regulatory Inversion (MECH-031) are already well advanced.
The Argument
The Architecture of Concentration: Why the AI Stack Resists Competition
The generative AI economy is not a horizontal marketplace of competing applications. It is a vertically integrated technology stack — a layered hierarchy where each level depends on the one beneath it, and where power accumulates at the bottom. Understanding this architecture is the prerequisite for any serious policy intervention, because the competitive dynamics at the application layer (where consumers see chatbots and copilots) are almost entirely determined by the monopoly and oligopoly structures at the foundational layers (where silicon, energy, and capital converge).
The stack comprises four layers. At the base sits the semiconductor layer: specialized AI accelerators, overwhelmingly NVIDIA GPUs, without which no serious model training or inference is possible. NVIDIA controls an estimated 85-92% of the AI accelerator market and 94% of discrete GPUs, a dominance cemented not merely by superior silicon but by the CUDA software ecosystem that creates profound switching costs [1] [Measured]. The French Competition Authority’s Autorite de la Concurrence concluded in 2025 that NVIDIA is likely abusing this dominance through price fixing, production restrictions, unfair contractual conditions, and discriminatory behavior [2] [Measured]. The U.S. Department of Justice issued subpoenas seeking evidence of antitrust violations [3] [Measured]. China launched its own anti-monopoly probe in December 2024 [4] [Measured]. The simultaneous investigations across three major jurisdictions are themselves evidence of the severity of the concentration.
Above the semiconductor layer sits cloud infrastructure, where Amazon Web Services (30%), Microsoft Azure (20%), and Google Cloud (13%) collectively control over 60% of global capacity [5] [Measured]. These hyperscalers are not merely landlords renting GPU time; they are gatekeepers who set the terms under which every AI developer operates. The FTC’s January 2025 staff report on AI partnerships and investments documented how contractual terms and technical barriers make it difficult for AI developers to switch cloud providers, effectively locking startups into dependency relationships with the very firms that compete against them at higher layers of the stack [6] [Measured].
The foundation model layer has seen training costs escalate from $900 for the original Transformer in 2017 to over $650 million for Google’s Gemini, with costs for frontier models doubling approximately every six months [7] [Measured]. This cost trajectory creates what economists term a natural monopoly dynamic: the fixed costs are so immense that the market can sustain only a small number of producers, and those producers are increasingly the same hyperscalers that control the cloud layer beneath.
At the top sits the application layer — superficially vibrant, with thousands of startups building on foundation model APIs. But this vibrancy is misleading. Application-layer firms are structurally dependent on the foundation model and cloud layers below them. A chatbot startup that loses its API access or faces a price increase from its model provider has no meaningful recourse. The competitive diversity at the top is an illusion sustained by the tolerance of the monopolists at the bottom — tolerance that can be withdrawn at any time.
Compute Feudalism: The Mechanism Behind the Architecture
The Recursive Institute’s framework identifies this dynamic as Compute Feudalism (MECH-029): the process by which open-weight model democratization fails to prevent infrastructure-layer concentration [Framework — Original]. As model weights become commoditized — Meta’s Llama, Mistral’s open models, and dozens of others are freely available — competitive advantage and dependency shift to the inference-serving stack: orchestration platforms, custom silicon, API ecosystems, and the energy and cooling infrastructure that keeps everything running. The result is not a commodity market but vertically integrated fiefdoms where the lords of compute extract rents from every participant in their ecosystem.
This mechanism operates through several reinforcing channels. First, the capital-compute-data flywheel: training a frontier model requires massive compute, which requires massive capital, which can only be recouped through massive distribution — and only the existing hyperscalers possess all three simultaneously [8] [Framework — Original]. Second, infrastructural path dependency: once a hyperscaler commits to a specific hardware architecture (CXL memory pooling, photonic interconnects, custom ASIC designs), every component must be compatible, creating lock-in through physics rather than licensing [9] [Estimated]. Third, the partnership-as-merger strategy: Microsoft’s $13 billion investment in OpenAI, Google’s investment in Anthropic, and Amazon’s parallel investment in Anthropic are not arm’s-length commercial agreements but deep structural entanglements that achieve the effects of vertical mergers while evading traditional merger review [10] [Measured].
The FTC’s February 2026 escalation of its Microsoft investigation — issuing civil investigative demands to six or more competitors examining cloud licensing, AI bundling, and market dominance — signals that regulators are beginning to understand the full-stack nature of the problem [11] [Measured]. But understanding and effective intervention are different things entirely.
The Regulatory Inversion: Why Democratic Governance Is Losing
The second mechanism at work is the Regulatory Inversion (MECH-031): a structural process by which AI-specific features — architectural opacity, capability velocity, and infrastructure entanglement — convert democratic AI governance into a legitimation ceremony for industry self-regulation [Framework — Original]. This mechanism operates through a five-step sequence.
Step one is the complexity moat: the technical sophistication of AI systems exceeds the institutional capacity of regulatory bodies to evaluate them, creating asymmetric dependence on industry expertise. Step two is the personnel siphon: the salary differential between regulatory agencies and AI firms systematically drains technical talent from the public sector. Step three is standard colonization: industry actors dominate the technical standard-setting processes that effectively determine regulatory substance. Step four is dependency installation: governments become reliant on the same firms they are supposed to regulate for critical digital infrastructure. Step five is post-enactment hollowing: even when legislation passes, implementation is shaped by the regulated firms’ control over the technical parameters that determine compliance.
The evidence from 2025-2026 shows steps one through three operating at full intensity. The EU AI Act, which entered into force in August 2024, requires compliance with transparency and safety requirements that are defined by technical standards still being developed — standards whose development is dominated by the same firms subject to regulation [12] [Measured]. The EU’s consideration of expanding the Digital Markets Act to designate AI businesses as “gatekeepers” represents an attempt to break out of this dynamic, but the May 2026 review deadline creates a window during which the regulatory inversion continues to operate [13] [Measured].
In the United States, the regulatory picture is more fractured. The White House’s July 2025 AI Action Plan emphasizes removing regulatory barriers to AI infrastructure development — language that, whatever its intent, aligns with the interests of incumbent firms seeking to build out compute capacity without constraint [14] [Measured]. The FTC’s bipartisan investigation into Microsoft represents a countervailing force, but the institutional asymmetry between a resource-constrained agency and the world’s most valuable companies remains stark.
The International Dimension: Competing Regulatory Philosophies
The global fragmentation of AI governance creates its own competitive dynamics. The EU’s approach — the AI Act plus potential DMA expansion plus active antitrust enforcement including the EUR 2.95 billion Google fine — represents the most comprehensive attempt at structural intervention [15] [Measured]. China’s approach combines domestic anti-monopoly enforcement with strategic compute nationalism. The U.S. approach oscillates between antitrust enforcement (FTC investigations) and innovation-first deregulation (the AI Action Plan).
This fragmentation benefits incumbents. A firm like Microsoft or Google can forum-shop across jurisdictions, leveraging the threat of relocating investment to extract regulatory concessions. The absence of international coordination on AI competition policy creates what game theorists call a race to the bottom: each jurisdiction faces pressure to weaken its standards to attract AI investment, even when collective action would produce better outcomes for all.
The Cloud and AI Development Act proposed for Q1 2026 in the EU represents an attempt to address this dynamic by building European compute capacity, reducing dependence on U.S. hyperscalers, and creating conditions for a more diversified AI ecosystem [16] [Estimated]. Whether it can succeed against the structural advantages of incumbent firms remains an open question.
Custom Silicon and the Shifting Competitive Landscape
One potentially disruptive development is the rise of custom AI chips designed by hyperscalers themselves. Google’s TPUs, Amazon’s Trainium and Inferentia chips, and Meta’s MTIA are all attempts to break free from NVIDIA’s hardware monopoly. Projections suggest that custom chips will account for 45% of the AI chip market by 2028, up from 37% in 2024 [17] [Estimated]. However, this development cuts both ways for competition. While it reduces NVIDIA’s monopoly power at the semiconductor layer, it deepens vertical integration by the hyperscalers, who now control both the hardware and the cloud infrastructure layers. A firm locked into AWS is no less dependent for running on Amazon’s custom silicon rather than NVIDIA’s GPUs; it may in fact be more dependent, as the switching costs increase when hardware and cloud platform are tightly coupled.
This dynamic illustrates a broader principle: competition at one layer of the stack does not necessarily produce competition at the stack level. The relevant question for policy is not whether NVIDIA faces competitive pressure from custom silicon, but whether end users — businesses, governments, individuals — have meaningful choice and agency in their relationship with the AI ecosystem. By that measure, the shift to custom silicon may actually worsen the competitive landscape even as it improves it within the narrow semiconductor market.
The Partnership Problem: De Facto Mergers Without Merger Review
The Microsoft-OpenAI partnership ($13 billion investment, exclusive cloud provider status, revenue-sharing rights, and consultation and control rights) has become the paradigmatic case of what competition scholars call a “killer collaboration” [18] [Measured]. Stanford Law School’s analysis concluded that these AI partnerships have moved “beyond control,” with the structural entanglements exceeding what existing regulatory frameworks were designed to address [19] [Measured]. The FTC’s January 2025 staff report documented how these partnerships grant cloud providers equity stakes, revenue-sharing rights, and certain consultation, control, and exclusivity rights that function as de facto merger instruments [20] [Measured].
The Microsoft-OpenAI arrangement is not unique. Google’s and Amazon’s competing investments in Anthropic, and the broader pattern of hyperscaler-startup entanglements, create a market structure in which every significant AI startup is financially, infrastructurally, and contractually bound to one of the three dominant cloud providers. The independent AI startup — funded by venture capital, running on commodity cloud, and competing on merit alone — is becoming an endangered species. The structural incentives push every promising AI company toward partnership with a hyperscaler, and every such partnership reduces the competitive diversity of the ecosystem.
The antitrust class action filed against Microsoft over its OpenAI partnership represents a novel legal theory: that the partnership structure itself, rather than any specific anticompetitive act, constitutes an unlawful restraint of trade [21] [Measured]. Whether this theory succeeds in court will have significant implications for the future of AI market structure.
The Energy Chokepoint: Power as the New Bottleneck
An underappreciated dimension of AI stack concentration is the energy constraint. Training and running frontier AI models requires enormous electrical power, and the data centers that house AI infrastructure are increasingly competing with residential and industrial consumers for access to the grid. In 2025-2026, hyperscalers signed long-term power purchase agreements (PPAs) with nuclear, natural gas, and renewable energy providers, effectively reserving energy capacity that smaller competitors cannot access. Microsoft’s agreement to restart the Three Mile Island nuclear facility for its AI data centers, Amazon’s acquisition of a nuclear-powered data center campus in Pennsylvania, and Google’s commitments to geothermal and advanced nuclear power represent a new form of resource lock-in that extends the AI monopoly from silicon and software into physical infrastructure [Estimated].
The energy dimension transforms the competition analysis in a critical way. A new entrant to the AI market must now secure not only capital, compute, and data but also long-term energy contracts at scale — contracts that the hyperscalers are systematically locking up. The Department of Energy estimated that U.S. data center electricity consumption could double or triple by 2028, raising questions about whether the grid can support both AI growth and existing demand. This energy competition creates what might be called an “energy moat” around incumbent AI firms: they have the financial resources to sign multi-decade PPAs, build dedicated power generation facilities, and invest in grid infrastructure, while smaller competitors are left to compete for whatever capacity remains on the open market at higher prices.
The energy constraint also has significant equity implications. Communities near AI data centers face increased electricity costs, grid strain, and environmental impacts, while the economic benefits of AI — productivity gains, new services, competitive advantages — accrue primarily to the firms that own the infrastructure and their customers in other locations. This geographic mismatch between the costs and benefits of AI infrastructure is a competition policy issue that existing frameworks are poorly equipped to address.
The Standards Battlefield: Who Writes the Rules That Govern AI
Technical standards are the invisible substrate of market competition. The standards that define how AI models are evaluated, how safety is assessed, how data is formatted, and how systems interoperate determine which firms can compete and on what terms. In the AI ecosystem, the standard-setting process has become a critical vector for Regulatory Inversion (MECH-031).
The EU AI Act delegates significant regulatory substance to technical standards developed by CEN and CENELEC (European standardization bodies) and ISO/IEC at the international level. The firms that participate most actively in these standard-setting processes — and that have the technical staff to prepare and defend proposals — are the same large technology companies that are subject to the regulation. The result is that the regulated entities effectively write the compliance requirements, choosing standards that their existing systems already meet while creating barriers for competitors whose architectures differ.
This dynamic is not unique to AI (it has been observed in telecommunications, pharmaceuticals, and financial services), but the speed and complexity of AI development make it particularly acute. A standard written in 2025 may be obsolete by 2027, but the compliance investments firms have made to meet it persist, creating path dependency that favors incumbents even after the standard’s technical rationale has evaporated. The standardization process becomes a mechanism for converting temporary technical advantage into durable regulatory advantage — one of the core channels through which Regulatory Inversion operates.
A Policy Framework for Structural Intervention
The analysis above suggests that incremental, conduct-based regulation — telling firms not to do specific bad things — is insufficient. The concentration in the AI stack is structural, driven by the economics of the technology itself. Effective policy must therefore be structural in nature.
Six categories of intervention merit consideration:
1. Mandatory interoperability and data portability. Cloud providers should be required to support standardized APIs and data formats that allow AI developers to move their workloads between providers without prohibitive switching costs. The EU’s DMA review provides a potential vehicle for this requirement [Estimated].
2. Structural separation at critical chokepoints. Firms that control cloud infrastructure should face restrictions on their ability to compete in the foundation model and application layers. This is the most aggressive intervention and faces significant political and legal obstacles, but it directly addresses the vertical integration that drives Compute Feudalism [Framework — Original].
3. Compute access obligations. Dominant cloud providers should be required to offer AI compute resources on fair, reasonable, and non-discriminatory (FRAND) terms, analogous to essential facilities doctrine. This would prevent hyperscalers from using preferential access to compute as a competitive weapon [Estimated].
4. Partnership transparency and review. Investments and partnerships above a defined threshold should be subject to the same merger review processes as outright acquisitions. The FTC’s investigation of Microsoft-OpenAI is a step in this direction, but systematic review requires legislative authorization [Measured].
5. Public compute infrastructure. Governments should invest in publicly accessible AI compute facilities — national AI compute centers — that provide researchers, startups, and public-sector institutions with access to training and inference resources independent of commercial hyperscalers. Several EU member states and the EU Cloud and AI Development Act are moving in this direction [Estimated].
6. International coordination. The fragmentation of AI competition policy across jurisdictions benefits incumbents through regulatory arbitrage. A multilateral framework for AI competition policy, even if initially limited to information sharing and common principles, would reduce the scope for forum-shopping [Framework — Original].
7. Energy access and environmental accountability. AI infrastructure developers should be subject to energy impact assessments analogous to environmental impact statements, and long-term power purchase agreements that effectively monopolize regional energy capacity should be subject to regulatory review. The externalization of energy costs and environmental impacts to host communities while concentrating economic benefits in distant corporate headquarters is a competition and equity issue that current frameworks neglect [Estimated].
Each of these interventions faces political, legal, and practical obstacles. Structural separation is the most effective but also the most politically difficult. Mandatory interoperability is technically achievable but depends on standard-setting processes that are themselves vulnerable to capture. Public compute infrastructure requires sustained public investment in a political environment that favors private markets. International coordination requires multilateral institutions that are currently fragmented and under-resourced. The framework is not a blueprint for easy action; it is a map of the intervention space that policymakers must navigate.
Mechanisms at Work
Compute Feudalism (MECH-029): The dynamic by which open-weight model democratization fails to prevent infrastructure-layer concentration. As model weights become commoditized, competitive advantage and dependency shift to the inference-serving stack — orchestration platforms, custom silicon, API ecosystems — producing vertically integrated fiefdoms rather than a commodity market. This mechanism is the structural engine driving AI monopolization: it explains why the proliferation of open models has not produced the decentralized, competitive market that many predicted.
The Regulatory Inversion (MECH-031): A structural inversion in which AI-specific features — architectural opacity, capability velocity, and infrastructure entanglement — create a self-reinforcing ratchet that converts democratic AI governance into a legitimation ceremony for industry self-regulation. This mechanism explains why regulatory responses to AI concentration have been slow, fragmented, and frequently captured by the firms they target. The five-step sequence (complexity moat, personnel siphon, standard colonization, dependency installation, post-enactment hollowing) is observable across multiple jurisdictions.
Interaction between MECH-029 and MECH-031: These mechanisms are mutually reinforcing. Compute Feudalism creates the concentrated market structure that gives incumbent firms the resources and incentives to invest in regulatory capture. Regulatory Inversion weakens the policy responses that could address Compute Feudalism. The result is a ratchet dynamic (see also MECH-014, The Ratchet) in which each round of concentration makes the next round of effective regulation less likely.
Counter-Arguments and Limitations
The Innovation Defense
The strongest counterargument to structural intervention is that the current market structure, however concentrated, is producing extraordinary innovation. Foundation models have improved at a pace that defies historical precedent. Application-layer creativity is abundant. Consumers benefit from rapidly improving AI capabilities delivered at declining per-query costs. Structural intervention risks disrupting the capital-intensive R&D pipeline that produces these gains, potentially slowing the pace of AI advancement and ceding competitive leadership to less regulated jurisdictions — particularly China.
This argument has genuine force. The counterfactual — what the AI market would look like under different structural conditions — is inherently uncertain. However, the innovation defense conflates the current rate of innovation with the optimal rate. Economic theory and historical precedent (the AT&T breakup, the Microsoft consent decree) suggest that monopoly power frequently produces less innovation than competitive markets, even when the monopolist appears innovative by absolute standards. The question is not whether incumbents are innovating but whether the market structure forecloses the innovations that would emerge from a more competitive ecosystem — innovations we cannot observe precisely because the foreclosure prevents them from existing.
The Natural Monopoly Argument
A related counterargument holds that the AI industry is a natural monopoly: the fixed costs of frontier model development are so high, and the economies of scale so powerful, that competition is inherently inefficient. On this view, policy should focus on regulating the behavior of dominant firms (price controls, access obligations) rather than attempting to create competition where the economics do not support it.
This argument correctly identifies the cost structure of frontier AI development. However, it conflates the frontier with the entire market. The majority of commercially valuable AI applications do not require frontier-scale models. Smaller, specialized models — often fine-tuned from open-weight bases — can serve most business use cases effectively. The natural monopoly argument applies, at most, to the narrow frontier of capability research; it does not justify the concentration of the entire AI stack from silicon to applications.
The Open-Weight Counter-Narrative
The most empirically grounded challenge to the concentration thesis comes from the open-weight model movement. Meta’s Llama, Mistral’s models, and numerous other openly available model weights have created genuine alternatives to the proprietary foundation models of OpenAI, Google, and Anthropic. If anyone can download and run a capable model, does it matter who controls the cloud infrastructure?
This counterargument identifies a real force for democratization but underestimates the infrastructure dependency that Compute Feudalism describes. An open-weight model is only as accessible as the compute required to run it. For inference at production scale, this means cloud infrastructure — and that infrastructure is controlled by the same oligopoly. For fine-tuning on proprietary data, the compute requirements are even higher. The open-weight movement has democratized model architecture; it has not democratized the capital-intensive infrastructure required to deploy those architectures at commercial scale.
Regulatory Capacity Skepticism
A practical objection is that regulatory bodies lack the technical sophistication and institutional capacity to implement structural interventions effectively. The AI industry moves faster than regulators can respond, and prescriptive rules risk being outdated before they take effect. This concern is legitimate and is itself a manifestation of the Regulatory Inversion mechanism. However, it is an argument for building regulatory capacity, not for abandoning regulation. The alternative — allowing the market to self-organize without public oversight — has a well-documented historical track record of producing outcomes that serve capital at the expense of democratic governance.
The Startup Ecosystem Counter-Evidence
A further counterargument points to the continued vitality of the AI startup ecosystem as evidence that concentration is not as severe as claimed. Venture capital investment in AI startups reached record levels in 2025, with hundreds of new firms entering the market. If the barriers to entry are truly insurmountable, how does one explain this entrepreneurial activity?
The answer lies in the distinction between entry and independence. The barriers to entry that matter for competition are not barriers to founding a company (which requires only an idea, some code, and venture funding) but barriers to operating independently at scale without dependence on an incumbent. By this measure, the startup ecosystem is less healthy than it appears. The overwhelming majority of AI startups are built on foundation model APIs provided by incumbents, deployed on hyperscaler cloud infrastructure, and dependent on NVIDIA hardware. Their economic viability depends on terms set by the very firms they would need to challenge in order to become independent competitors. Many are, in effect, application-layer tenants in an infrastructure owned by the oligopoly.
The venture capital pattern reinforces rather than challenges the concentration thesis. The largest AI venture rounds in 2025-2026 were co-led or directly funded by the hyperscalers themselves: Microsoft’s investments through M12 and its direct partnership with OpenAI, Google Ventures’ investments in AI infrastructure startups, and Amazon’s Anthropic investment. The venture capital flowing into AI startups often originates from the same firms whose market power the startups would theoretically challenge. This creates a selection pressure that favors startups whose business models complement rather than threaten incumbent positions.
The China Competition Concern
Perhaps the most politically potent counterargument is that structural intervention would weaken U.S. (or Western) AI firms relative to Chinese competitors operating under a more permissive regulatory environment. This argument treats the AI race as a zero-sum geopolitical contest in which any constraint on domestic firms is a gift to adversaries.
The geopolitical dimension is real but the argument is overstated. China’s AI ecosystem faces its own structural constraints — semiconductor export controls, energy costs, capital market limitations — and Chinese firms are subject to their own anti-monopoly enforcement (as the NVIDIA probe demonstrates). Moreover, the argument proves too much: taken to its logical conclusion, it would preclude all regulation of any industry with international competition, which is to say all industries. The policy question is how to maintain competitive market structures while preserving the capacity for large-scale R&D investment — a balance that requires nuance rather than blanket deference to incumbent firms.
Temporal Limitations
This analysis is necessarily bounded by 2025-2026 evidence. The AI industry is evolving rapidly, and several developments could alter the competitive dynamics described here. The maturation of custom silicon could erode NVIDIA’s hardware monopoly more quickly than projected. Breakthroughs in training efficiency could reduce the capital barriers to frontier model development. New regulatory frameworks could prove more effective than historical precedent suggests. The confidence range of 45-60% reflects these substantial uncertainties.
What Would Change Our Mind
-
Custom silicon diversification reduces effective compute concentration below 50% HHI within three years — If the rise of Google TPUs, Amazon Trainium, and other custom chips produces genuine, stack-level competition (not merely intra-hyperscaler hardware substitution), the Compute Feudalism thesis would require significant revision.
-
Open-weight models achieve production-scale deployment on non-hyperscaler infrastructure — If a credible ecosystem of independent compute providers, edge inference, or decentralized compute networks enables open-weight models to compete with proprietary models without dependence on the Big Three cloud providers, the infrastructure-layer concentration concern diminishes substantially.
-
Regulatory intervention demonstrably reduces switching costs and increases developer mobility — If the EU’s DMA review, the FTC’s Microsoft investigation, or analogous actions produce measurable reductions in cloud lock-in and interoperability barriers, the Regulatory Inversion thesis would need qualification.
-
Training costs decline by an order of magnitude within 24 months — If algorithmic efficiency gains, new architectures, or hardware advances reduce the cost of frontier model training from hundreds of millions to tens of millions of dollars, the capital barrier to entry weakens significantly, potentially enabling a more competitive landscape.
-
A major AI startup achieves frontier performance without hyperscaler partnership — If a firm trains and deploys a frontier-competitive model using only independent infrastructure and venture capital, the structural necessity of hyperscaler partnerships would be disproven.
Confidence and Uncertainty
Overall confidence: 45-60%.
The descriptive claims about current market concentration are high-confidence (75-85%), supported by extensive market data, regulatory filings, and investigative journalism. NVIDIA’s dominance, the cloud oligopoly, and the partnership structures are empirical facts, not theoretical claims [Measured].
The mechanistic claims — that Compute Feudalism and Regulatory Inversion are self-reinforcing and likely to produce durable monopoly — carry moderate confidence (50-65%). The mechanisms are logically coherent and supported by early evidence, but the AI market is young and evolving rapidly. Disruptions from custom silicon, algorithmic efficiency, or unexpectedly effective regulation could alter trajectories [Estimated].
The policy prescriptions carry the lowest confidence (35-50%). They are informed by economic theory and historical precedent but involve complex political economy questions about regulatory capacity, institutional design, and international coordination that do not yield to purely analytical resolution [Framework — Original].
Implications
The concentration of the AI stack has implications that extend far beyond competition policy. If a handful of firms control the foundational infrastructure of artificial intelligence, they effectively control the substrate on which an increasing share of economic, social, and political activity depends. This is not merely a market efficiency question; it is a question of democratic governance and human agency.
The Ratchet mechanism (MECH-014) suggests that the window for structural intervention is finite: sunk capital expenditure, debt, and institutional dependency make retreat from escalating AI infrastructure spending more costly than continuation. Each year of inaction narrows the set of feasible policy options. The mechanisms described here are not self-correcting; without deliberate intervention, they produce outcomes that are stable, concentrated, and resistant to change.
For the broader Theory of Recursive Displacement, this essay illustrates how monopolization of the AI stack creates the conditions for downstream displacement mechanisms. Compute Feudalism enables Cognitive Enclosure (MECH-007) by controlling access to AI capabilities. Regulatory Inversion weakens the policy responses that could mitigate Structural Exclusion (MECH-026) and the Aggregate Demand Crisis (MECH-010). The monopoly question is not peripheral to the displacement thesis; it is foundational.
Where This Connects
This essay builds directly on the Institute’s analysis of Compute Feudalism (MECH-029), which details how infrastructure-layer concentration persists despite model-weight democratization. The Regulatory Inversion essay (MECH-031) provides the five-step mechanism by which AI governance is captured. The Ratchet (MECH-014) explains why the window for intervention narrows over time. The Liability Vacuum (MECH-032) documents how existing legal frameworks fail to assign accountability in AI-mediated harms — a dynamic that compounds when the harming entities are monopolists. The Aggregate Demand Crisis (MECH-010) traces the macroeconomic consequences when concentrated AI infrastructure displaces labor income without creating compensating demand. And the Sequencing Problem (MECH-022) suggests that the order in which these mechanisms activate determines which equilibrium a society reaches — making the timing of competition policy intervention a critical variable.
Conclusion
The AI stack is consolidating along lines that historical precedent and economic theory predict will be durable and self-reinforcing. NVIDIA’s semiconductor monopoly, the hyperscaler cloud oligopoly, the billion-dollar barriers to frontier model training, and the partnership structures that bind startups to incumbents are not temporary market conditions awaiting correction by competitive forces. They are structural features of an industry whose economics favor concentration at every foundational layer.
The mechanisms of Compute Feudalism and Regulatory Inversion describe how this concentration persists and deepens despite apparent democratization (open-weight models) and regulatory effort (the EU AI Act, FTC investigations). The policy framework proposed here — mandatory interoperability, structural separation, compute access obligations, partnership review, public compute infrastructure, and international coordination — is ambitious but proportionate to the scale of the problem.
The evidence from 2025-2026 offers both warning and limited encouragement. The warning is that concentration is advancing faster than regulation. The encouragement is that regulators in multiple jurisdictions are beginning to understand the full-stack nature of the problem and are developing tools — the DMA expansion, the FTC’s Microsoft probe, the French Competition Authority’s NVIDIA investigation — that could, if pursued with sufficient resources and political will, alter the trajectory. Whether that will and those resources materialize is a political question, not an analytical one. What analysis can establish is that the stakes are high, the window is narrowing, and the costs of inaction compound with each passing quarter.
Sources
[1] NVIDIA GPU Market Share and AI Accelerator Dominance. Yahoo Finance, 2025. https://finance.yahoo.com/news/nvidias-85-gpu-market-share-210500376.html
[2] French Competition Authority (Autorite de la Concurrence) NVIDIA Investigation Findings. Bloomberg, 2025. https://www.bloomberg.com/news/features/2025-03-20/are-ai-monopolies-here-to-stay-nvidia-and-the-future-of-ai-chips
[3] U.S. Department of Justice Subpoenas to NVIDIA. American Action Forum, 2025. https://www.americanactionforum.org/insight/the-doj-and-nvidia-ai-market-dominance-and-antitrust-concerns/
[4] China Anti-Monopoly Probe of NVIDIA. TechPolicy.Press, 2025. https://www.techpolicy.press/nvidia-is-building-a-shield-of-concentrated-power/
[5] Cloud Infrastructure Market Share (AWS, Azure, Google Cloud). Congress.gov, Library of Congress, 2025. https://www.congress.gov/crs-product/IF12968
[6] FTC Staff Report on AI Partnerships and Investments. Federal Trade Commission, January 2025. https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-issues-staff-report-ai-partnerships-investments-study
[7] AI Model Training Cost Escalation (Gemini $650M). Economic Policy Panel, 2025. https://economic-policy.org/79th-economic-policy-panel/ai-monopolies/
[8] Capital-Compute-Data Flywheel Analysis. Yale Law & Policy Review, 2025. https://yalelawandpolicy.org/antimonopoly-approach-governing-artificial-intelligence
[9] Infrastructural Path Dependency in AI Data Centers. ScienceDirect, 2025. https://www.sciencedirect.com/science/article/pii/S2212473X25000835
[10] Microsoft-OpenAI $13B Investment Structure. TechCrunch, 2025. https://techcrunch.com/2025/01/18/ftc-says-partnerships-like-microsoft-openai-raise-antitrust-concerns/
[11] FTC Escalation of Microsoft Investigation (February 2026). WinBuzzer, 2026. https://winbuzzer.com/2026/02/14/ftc-escalates-microsoft-probe-grills-rivals-cloud-monopoly-xcxwbn/
[12] EU AI Act Technical Standards Development. Slaughter and May Horizon Scanning, 2026. https://www.slaughterandmay.com/horizon-scanning/2026/digital/ai-update-for-2026/
[13] EU Digital Markets Act Review and AI Gatekeeper Designation. TechPolicy.Press, 2025. https://www.techpolicy.press/will-the-eu-designate-ai-under-the-digital-markets-act/
[14] White House AI Action Plan (July 2025). National Law Review, 2025. https://natlawreview.com/article/2026-outlook-artificial-intelligence
[15] EU Google AdTech Fine (EUR 2.95 Billion). Euronews, 2025. https://www.euronews.com/next/2025/12/17/eu-takes-on-big-tech-here-are-the-top-actions-regulators-have-taken-in-2025
[16] EU Cloud and AI Development Act (Q1 2026). Bird & Bird, 2025. https://www.twobirds.com/en/insights/2025/eu-a-bid-for-tech-sovereignty-drives-commissions-work-programme-for-2026
[17] Custom AI Chip Market Projections (45% by 2028). Yahoo Finance, 2025. https://finance.yahoo.com/news/nvidias-big-tech-customers-might-also-be-its-biggest-competitive-threat-153032596.html
[18] Microsoft-OpenAI Partnership Structure Analysis. Bloomberg, 2025. https://www.bloomberg.com/news/articles/2025-01-17/microsoft-openai-partnership-raises-antitrust-concerns-ftc
[19] Stanford Law School Analysis of AI Partnerships. Stanford CodeX, 2025. https://law.stanford.edu/2025/03/21/ai-partnerships-beyond-control-lessons-from-the-openai-microsoft-saga/
[20] FTC Staff Report: Cloud Provider Equity and Control Rights. Federal Trade Commission, 2025. https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-issues-staff-report-ai-partnerships-investments-study
[21] Microsoft Antitrust Class Action Over OpenAI Partnership. Grellas, 2025. https://grellas.com/microsoft-faces-antitrust-class-action-over-openai-partnership/