LD-11 — Concentration, Stack & Sovereignty
EVIDENCE CUT
Sep 25, 2026
CURRENT SOCIOLENS POSITION
AI power is concentrated, but not in one market and not in the same way at every layer.
The strongest current evidence supports a layered picture:
* some upstream chokepoints are extremely concentrated;
* ownership of large-scale AI compute is concentrated and has been rising;
* cloud markets show documented switching and licensing leverage;
* the foundation-model layer is concentrated but materially contestable;
* open-weight models reduce dependence on proprietary model vendors without removing dependence on chips, fabs, memory or clouds;
* states can reorganize access through export controls, procurement and industrial policy, but “sovereignty” is partial rather than binary.
The best-supported overall position is therefore:
LAYERED / CONTESTABLE CONCENTRATION, WITH A STRUCTURAL CORE UPSTREAM.
THE QUESTION
Where is power concentrated across the AI stack, through which mechanisms does that concentration matter, and under what conditions can firms, states, researchers or publics reduce dependency?
The wrong question is simply:
“Is AI concentrated?”
A concentration claim is interpretable only when it names:
* the layer;
* the market or denominator;
* the metric;
* the geography;
* the period;
* and the mechanism, if the claim is about power rather than concentration alone.
THE STACK
This page separates at least seven layers:
* accelerator design;
* semiconductor fabrication, memory, packaging and equipment;
* compute infrastructure and cloud;
* frontier/foundation models;
* applications and deployment;
* distribution and gateways;
* data, evaluation and research access.
Concentration at one layer does not automatically establish concentration at another.
1. UPSTREAM CHOKEPOINTS
EUV lithography
FACT.
ASML is currently the world's only manufacturer of EUV lithography systems. In its 2025 reporting it recorded 48 EUV systems and €11.6 billion of EUV system sales.
This is a genuine technological chokepoint.
It does not mean ASML controls downstream AI markets, and EUV should not be confused with lithography as a whole.
Contract foundry
ESTIMATE — commercial market intelligence.
TrendForce estimated that TSMC captured 72.5% of revenue among the world's top ten contract foundries in Q2 2026.
That denominator matters. It is not “72.5% of AI” and should not be presented as an exact share of every possible foundry market.
TrendForce also reported 3nm and 5/4nm capacity as fully booked under strong AI GPU/XPU demand.
High-bandwidth memory
ESTIMATE — commercial market intelligence.
Counterpoint estimated Q2 2026 HBM revenue shares at:
* SK hynix: 50%;
* Samsung: 33%;
* Micron: 18%.
The market is highly concentrated, but internal shares can move quickly: SK hynix was estimated at 64% one year earlier.
This is an important example of concentration and contestability coexisting.
2. WHO OWNS AI COMPUTE?
ESTIMATE — independent modelled dataset.
Epoch AI estimated that five U.S. hyperscalers owned about 71% of cumulative global AI compute in Q4 2025, measured in H100-equivalents, up from about 63% in Q1 2024.
Ownership is not the same as:
* workload share;
* access;
* effective usable capacity;
* or governance control over model companies.
But ownership matters because firms, researchers and states that do not own large compute pools often have to rent access from those that do.
Custom accelerators complicate the NVIDIA story
Epoch estimated Google held about one quarter of global cumulative AI compute in Q4 2025, driven primarily by roughly 3.81 million H100-equivalents of its own TPUs.
This is real accelerator substitution.
But it is not deconcentration in a simple sense: the substitute is itself controlled by a hyperscaler and still depends on concentrated fabrication and memory supply.
3. CLOUD: CONCENTRATION BECOMES LEVERAGE
ESTIMATE + REGULATORY FACT.
Synergy Research Group estimated Q2 2026 global cloud infrastructure services revenue shares of:
* AWS: 28%;
* Microsoft: 20%;
* Google: 15%.
Together, 63% under Synergy's overall cloud-infrastructure-services definition.
Generic cloud share is not AI-compute share.
More important for power is mechanism.
The UK Competition and Markets Authority found AWS and Microsoft have significant market power in UK cloud and identified:
* data-egress fees;
* interoperability barriers;
* Microsoft software licensing;
as constraints on switching and multi-cloud use.
This is stronger evidence of power than market share alone because it identifies mechanisms through which customers can become dependent.
4. DISTRIBUTION POWER: IMPORTANT, BUT NOT YET QUANTIFIED
In June 2026 the European Commission preliminarily considered AWS and Azure important cloud gateways that should be designated under the Digital Markets Act.
The Commission cited:
* lock-in;
* high switching costs;
* ecosystem breadth;
* and the growing importance of AI tools and partnerships in cloud procurement.
As of the Sep 25, 2026 evidence cut, this was a preliminary position, not a final cloud gatekeeper designation.
R06 did not locate a strong causal estimate showing how much AI bundling, defaults or distribution ecosystems change adoption or foreclose rivals.
That remains an important UNKNOWN.
5. THE MODEL LAYER IS MORE CONTESTABLE
Enterprise provider shares can reverse
ESTIMATE — investor survey/model.
Menlo Ventures estimated late-2025 enterprise LLM API shares at:
* Anthropic: 40%;
* OpenAI: 27%;
* Google: 21%.
The same series places OpenAI at 50% and Anthropic at 12% in 2023.
The estimate comes from a survey of 495 U.S. enterprise AI decision-makers combined with market modelling. Menlo is an investor in Anthropic, so the institutional position matters.
This is not audited provider revenue.
It is still evidence that leadership under one enterprise usage/spend metric can move sharply within two years.
Open weights are substantial on one major routing platform
ESTIMATE — platform telemetry.
OpenRouter and a16z report that open-weight models reached approximately one-third of token usage on OpenRouter by late 2025.
That is meaningful counterpower at the model layer.
It does not establish:
* one-third of the global model market;
* one-third of enterprise production;
* or independence from concentrated infrastructure.
Fixed-capability inference prices are falling rapidly
ESTIMATE — benchmark-price analysis.
Epoch AI's September 2026 multi-benchmark analysis estimates that the price of reaching fixed LLM capability levels has fallen about 13× per year since 2023.
An earlier Epoch analysis found much wider threshold-specific rates, from 9× to 900× per year depending on the task and performance milestone.
The newer 13× figure is better for a general statement.
But falling prices do not identify their own cause. Competition, hardware, algorithms, utilization, distillation and provider strategy can all matter.
6. OPEN WEIGHTS ARE COUNTERPOWER — NOT OPEN INFRASTRUCTURE
Open weights can reduce dependence on a proprietary model provider.
They can:
* widen model choice;
* support self-hosting in some settings;
* put pressure on API pricing;
* enable rapid switching for developers with the necessary infrastructure.
They do not automatically provide:
* chips;
* HBM;
* leading-edge fabrication;
* data-centre capacity;
* energy;
* cloud independence;
* or distribution.
OPEN WEIGHTS ≠ OPEN INFRASTRUCTURE.
7. PARTNERSHIPS: FOLLOW THE RIGHTS, NOT THE HEADLINE DOLLARS
Large investments can create dependence, but investment size alone does not establish control.
FTC staff found that major cloud–AI-developer partnerships included combinations of:
* equity;
* revenue sharing;
* consultation or control rights;
* exclusivity rights.
The FTC staff report did not establish antitrust liability.
The correct unit of analysis is the specific contractual right.
Microsoft and OpenAI
The April 2026 terms materially changed the earlier relationship.
Microsoft remains OpenAI's primary cloud partner, but:
* OpenAI can serve products across cloud providers;
* Microsoft's model/product IP licence through 2032 is non-exclusive.
This means the October 2025 exclusivity picture cannot be treated as the current state.
Anthropic and multi-homing
Anthropic reports that it trains or serves Claude across:
* AWS Trainium;
* Google TPUs;
* NVIDIA GPUs;
and distributes Claude through AWS, Google Cloud and Microsoft Azure.
This proves multi-homing is feasible for one very well-capitalized frontier lab.
It does not prove that switching or multi-homing is cheap for startups, SMEs, universities or public agencies.
8. STATES CAN REORGANIZE DEPENDENCIES
Export controls and permissions
FACT.
In November 2025 the U.S. Department of Commerce authorized G42 in the UAE and Humain in Saudi Arabia to purchase the equivalent of up to 35,000 NVIDIA Blackwell GB300 chips each, subject to security and reporting requirements.
In July 2026 the UAE was reclassified under U.S. export rules, and current EAR provisions provide specified advanced-computing access for approved UAE entities such as G42.
This demonstrates that access to specified advanced computing items can depend on export-control authorization and entity-specific eligibility.
Authorization is not the same as delivery or operational capacity.
China
FACT + UNKNOWN.
NVIDIA states that the U.S. regulatory environment effectively foreclosed it from competing in China's data-center compute market.
That is a documented market-access effect.
R06 does not establish how fully Chinese domestic accelerators replace restricted NVIDIA products.
Claims about Huawei output, HBM constraints and resulting frontier capability remain HOLD pending stronger primary evidence.
9. SOVEREIGNTY IS NOT BINARY
UAE
Mubadala described Stargate UAE as under construction in 2026, with the initial 200MW cluster due to begin operations in late 2026.
As of Sep 25, 2026, that capacity should not be counted as operational.
The UAE case combines:
* capital;
* partnerships;
* construction;
* privileged U.S. access;
with continuing dependence on:
* foreign accelerators;
* U.S. export permission;
* external model/cloud partners.
European Union
The Cloud and AI Development Act remained a proposal at the evidence cut.
Its stated objective includes at least tripling EU data-centre capacity within 5–7 years.
The AI Gigafactories initiative was in procurement for up to seven sites, with the call closing on Nov 12, 2026.
These are real state-capacity instruments.
They are not yet evidence that the planned capacity is operating or that Europe has become independent of foreign accelerators.
SOVEREIGNTY MUST NAME THE CAPABILITY AND THE DEPENDENCY.
10. PUBLIC RESEARCH ACCESS
FACT WITH VERSION NOTE.
NSF reported on Sep 1, 2026 that the U.S. National AI Research Resource had supported over 800 research projects.
NAIRR also relies materially on private-sector in-kind resources; the current NSF programme page reports approximately US$100 million in such contributions.The evergreen NSF NAIRR page currently displays 600+ research and education projects, while the later datedSep 1, 2026 announcement says over 800 research projects. These are preserved as versioned programme metrics rather than silently reconciled.
This expands access while preserving a dependence relationship with private providers.
R06 does not establish that public research compute is “too small to matter.” Dollar values are not a clean measure of usable compute or bargaining power.
10A. WHO GAINS, WHO DEPENDS, WHO IS EXPOSED?
The clearest distributional consequences in the current R06 evidence are institutional and geographic rather than individual-level.
* Very large firms and frontier labs can make supply and compute commitments at scales unavailable to most entrants; NVIDIA's own US$279 billion supply/capacity commitments illustrate the scale of pre-commitment in the system, without proving exclusion.
* UK cloud users face documented switching, interoperability and software-licensing constraints.
* Public researchers gain access through NAIRR, while part of the resource pool is provided in kind by private technology companies.
* Export-control regimes create unequal jurisdictional access to specified advanced computing items.
* R06 does not yet provide mature evidence on worker bargaining, productivity-rent sharing, household incidence, race, gender, disability, or operator-level environmental burdens attributable specifically to concentration.
These gaps are part of the result. Industrial concentration is better measured than many of its downstream social consequences.
11. THE CAPITAL SCALE OF BOTTLENECKS
FACT — company filing.
NVIDIA disclosed US$279 billion in supply and capacity commitments as of Jul 26, 2026, primarily for memory and manufacturing facilities.
This is evidence of the extraordinary scale of capacity reservation and capital commitment around the AI supply chain.
It is not evidence that NVIDIA reserved a specific percentage of CoWoS or that smaller buyers were directly excluded.
12. FOUR POSITIONS
P1 — Structural concentration
Strongly supported upstream and in compute ownership.
Weak as a whole-stack description.
P2 — Layered / contestable concentration
Best supported overall.
Some layers contain durable chokepoints.
Other layers show rapid entry, substitution, price movement and changing provider shares.
P3 — Openness as counterpower
Supported at the model layer, but bounded.
Open weights reduce model-vendor dependence.
They do not create full-stack independence.
P4 — State / sovereignty recomposition
Supported for legal access, export controls, procurement and industrial policy.
Operational capability effects are still emerging.
13. WHAT WE DO NOT KNOW YET
R06 does not establish:
* an independent merchant AI-accelerator market share;
* the exact leading-edge-node foundry share;
* NVIDIA's exact share of 2026 CoWoS allocation;
* an audited global foundation-model revenue share;
* global open-weight share beyond OpenRouter;
* the magnitude of Chinese domestic accelerator substitution;
* a quantified AI-specific bundling/default foreclosure effect;
* the causal decomposition of inference-price decline;
* how much public research compute changes bargaining power;
* mature labour-bargaining or productivity-rent effects;
* operator-level energy and water burdens attributable specifically to market concentration.
These are update triggers, not reasons to postpone v0.1.
14. GEOGRAPHIC AND EVIDENCE SCOPE
Primary monitoring core:
Europe and North America.
R06 also relies on:
* Taiwan and Korea for semiconductor supply;
* China for substitution/export-control questions;
* UAE and Saudi Arabia for state-capital and access cases.
Important evidence limitations:
* several semiconductor shares come from commercial market intelligence;
* current market figures can change quarter by quarter;
* Chinese-language primary capacity evidence remains thin;
* model-market denominators differ sharply;
* distribution-power evidence is more regulatory than causal;
* downstream social consequences remain less mature than the industrial-structure evidence.
15. HOW TO READ FAST-MOVING CLAIMS
For this debate, dates are part of the claim.
Do not silently update:
* market share;
* partnership terms;
* regulatory status;
* capacity state;
* export rules;
* construction status.
A new figure should supersede an old one explicitly, not overwrite it invisibly.
16. WHAT WOULD MOVE THIS DEBATE?
The synthesis would materially change with:
* a second viable EUV supplier at scale;
* major leading-edge foundry entry;
* new HBM entrants;
* a sustained decline in hyperscaler compute ownership;
* audited provider revenue shares;
* independent multi-platform open-weight telemetry;
* measured post-remedy cloud switching;
* final EU cloud DMA decisions;
* operational EU/Gulf public or sovereign compute at material scale;
* strong Chinese-primary evidence on domestic chips, HBM and fabrication;
* quantified downstream effects on procurement, research access, firms or workers.
KEY EVIDENCE
* ASML, 2025 Annual Report — vendor primary / audited.
* TrendForce, Q2 2026 foundry market release — commercial market intelligence.
* Counterpoint Research, Q2 2026 HBM market shares — commercial market intelligence.
* Epoch AI, AI Chip Owners analyses, April 2026 — independent modelled dataset.
* Synergy Research Group, Q2 2026 cloud market — commercial market intelligence.
* UK Competition and Markets Authority, Cloud Services Market Investigation and 2026 follow-up — regulator.
* European Commission, June 2026 DMA cloud preliminary position — regulator.
* FTC Office of Technology, CSP–AI developer partnership study — regulator staff report.
* Microsoft/OpenAI, April 2026 partnership update — counterparties' primary disclosure.
* Anthropic and Broadcom, 2026 compute disclosures — vendor + SEC filing.
* Menlo Ventures, 2025 State of Generative AI in the Enterprise — investor survey/model.
* OpenRouter/a16z, 100 Trillion Token Study — platform telemetry.
* Epoch AI, The Plunging Price of Thought, September 2026 — independent analysis.
* U.S. Department of Commerce / BIS, 2025–2026 advanced-compute export actions — government primary.
* European Commission, CADA and AI Gigafactories 2026 — government primary.
* NSF, NAIRR 2026 programme reporting — government primary.
* NVIDIA, Q2 FY2027 Form 10-Q — SEC filing.
PAGE STATUS
LIVING DEBATE v0.1 — LOCKED.
Built from R06 HQ-adjudicated canonical claims only.
HOLD-only quantitative claims excluded.
Future changes require versioned amendment or explicit supersession.