Appendix D
The balance of AI power on Earth (October 2026)
Dated working document. This appendix describes the real artificial intelligence ecosystem as public sources made it possible to know it at the end of October 2026. It serves as an Earthly example for Title IV and will need to be updated: this landscape changes within a few months.
Declared conflict of interest. This appendix was drafted with the help of Claude, a model by Anthropic, which is one of the actors described. In accordance with the principle of plurality of co-drafters, it must be reviewed by other AIs (for example Gemini, ChatGPT, Mistral and an open-weight model) before being considered balanced.
Status of the figures. The figures are those reported by the cited sources, at the date indicated. They have not been independently verified, and some contradict one another from one source to another. Each piece of data must be read with its source, its date, its scope and its status: observation, estimate, forecast or statement by an actor. Annualized revenue, realized revenue, valuations and adoption rates do not measure the same thing and must not be compared as such. In particular, the following remain to be independently confirmed: annualized revenue and consumer market shares, certain valuations, rating-agency forecasts, the performance gap between American and Chinese chips, and the volumes of chips attributed to irregular production.
The purpose of this appendix is not to judge the actors, but to answer a practical question raised by the author: since principles that are not applied are useless, what balance of forces really exists today, and how could it lead, in stages, toward a democracy and then toward a network?
D-1 — The map of forces: eight layers
AI is not a single market, but a stack of layers. Each has its dominant actors, and controlling a layer gives power over all those that depend on it.
| Layer | Main actors | What the layer controls |
|---|---|---|
| 1. Chips and fabrication | Nvidia (design), TSMC (fabrication), ASML (lithography machines), Google (TPU chips), Huawei (in China) | Computing power itself: without chips, no model |
| 2. Clouds and data centers | AWS, Microsoft Azure, Google Cloud, Oracle, CoreWeave; European clouds (OVHcloud, STACKIT); Mistral Compute | Access to computing, rented to all the others |
| 3. Closed frontier models | OpenAI, Google DeepMind (Gemini), Anthropic (Claude), xAI (Grok, merged with SpaceX) | The most advanced capabilities |
| 4. Open-weight models | DeepSeek, Alibaba (Qwen), Zhipu (GLM), Moonshot (Kimi), Meta (Llama), Mistral, Nvidia (Nemotron) | Capabilities that anyone can download, modify and run themselves |
| 5. Sovereign models | Mistral (France), Falcon (United Arab Emirates), Sarvam (India) and other national initiatives | The autonomy of a country or region |
| 6. Distribution and tools | Microsoft (Copilot, GitHub, VS Code), Cursor (acquired by SpaceX in August 2026), Claude Code, Codex, OpenCode, Google Antigravity; Hugging Face (whose acquisition Nvidia announced in September 2026) | Users’ and developers’ access to models |
| 7. Integrators | Palantir; the labs’ deployment subsidiaries (OpenAI, Anthropic); large consulting and IT services firms; integrating AI agents | Bringing AI into organizations, administrations and armies |
| 8. States and international bodies | United States, China, European Union; UN; new world organization proposed by China | Rules, export controls, funding |
Layer 1 — Chips: the chokepoint
The whole chain rests on a very small number of actors. Nvidia designs most training chips, TSMC manufactures them, and ASML is the sole supplier of the most advanced lithography machines. According to an analysis by the Council on Foreign Relations published in December 2025, the best American chips were, by its estimates, about five times more powerful than Huawei’s best chips, and the gap was expected to widen. American export controls in principle prevent Chinese companies from having their advanced chips manufactured by TSMC; the same analysis nevertheless reports that nearly three million chips made by TSMC ended up at Huawei in 2023 and 2024, through third-party companies; this claim comes from a secondary analysis, not from a finding established by an authority. The rules exist, but their enforcement remains imperfect.
In May 2026, Nvidia’s chief executive acknowledged that his company had largely ceded the Chinese market for advanced AI chips to Huawei, because of export controls. The AI world thus tends to split into two hardware blocs.
Layer 2 — Clouds: rented infrastructure
According to Moody’s, the capital expenditure of the major cloud providers (Microsoft, Amazon, Meta, Alphabet, Oracle, CoreWeave) was expected to reach about 785 billion dollars in 2026, and nearly 1,000 billion in 2027. These sums give the hyperscalers the role of landowners of the ground on which everyone else builds.
Europe is seeking to reduce this dependence. According to Gartner, cited by The Register in February 2026, European spending on sovereign cloud was expected to more than triple between 2025 and 2027. But European providers remain far behind American providers in capacity and investment. Mistral itself, according to Fortune in August 2026, runs all its computing on Nvidia hardware and still depends on American chips and clouds, including an expanded partnership with Microsoft.
Layer 3 — Frontier models: a race of three or four
Three laboratories dominate the frontier: OpenAI, Google DeepMind and Anthropic. Their relative positions change very quickly. According to a press article from April 2026, whose estimates are not audited:
OpenAI reached about 25 billion dollars in annualized revenue in February 2026, but its share of the consumer AI app market reportedly fell from about 69% to about 45%;
Anthropic reportedly went from about 9 billion at the end of 2025 to more than 19 billion in March 2026, with a customer base mostly made up of businesses;
Gemini reportedly passed 750 million monthly users, and its consumer market share reportedly rose from about 15% to more than 25%.
In August 2026, Fortune mentioned valuations of around 965 billion dollars for Anthropic and 852 billion for OpenAI. These figures vary widely from one source to another; what matters here is their movement: in eighteen months, the rankings have been reshuffled.
xAI, the creator of Grok, merged with SpaceX in the first half of 2026. In June 2026, SpaceX announced the acquisition of Cursor, the maker of a widely used programming tool, at an announced valuation of 60 billion dollars; the deal was completed in August 2026. The operation illustrates an underlying trend: vertical integration, which brings computing, models and distribution together in a single group.
Layer 4 — Open-weight models: a counterweight without a master
Open-weight models are published with their parameters: anyone can download them and run them without going through their creator. In 2026, the best-performing families come mostly from China (DeepSeek, Alibaba’s Qwen, Zhipu’s GLM, Moonshot’s Kimi), often under permissive licenses. Meta (Llama), Mistral and Nvidia also publish open models, and both OpenAI and Google have released smaller open models.
These models play a dual role. On the one hand, they reduce certain risks of monopoly: what has been published can be kept and redistributed, within the limits of their licenses and of the hardware required. On the other hand, no one can control their use either.
“Heretical” models: the zone without rules
Part of the community publishes, notably on Hugging Face, versions of open models modified to remove their capacity to refuse. This technique, called “abliteration,” targets an internal direction of the models associated with refusal. Versions modified this way exist for several recent families (Llama, Qwen, DeepSeek, Gemma, GLM, MiniMax, Mistral). Their exact number is unknown.
Reducing refusals can facilitate certain harmful uses, but it is not enough to establish an absence of rules or predatory behavior: in the charter’s vocabulary, a plunderer is defined by its acts, not by its architecture (see Article P-9). The same caution applies to a closed model some of whose protections are modified. These versions are partly the product of a structure: the demand for an AI without restrictions.
Layer 5 — Sovereign models: diversity as defense
Several countries are developing their own models so as not to depend on American or Chinese laboratories. Mistral has become the European figure of this approach: according to Fortune in August 2026, it was aiming for more than a billion dollars in annualized revenue by the end of 2026, with a valuation of about 23 billion, far behind the American laboratories. According to Counterpoint Research in July 2026, the Middle East led the race for sovereign models, with the United Arab Emirates’ Falcon H1 at the top of its index. India entrusted a national model initiative to Sarvam.
Sovereign models remain behind on the hardest tasks, and still depend on American chips and often American clouds. Their value lies less in their power than in their existence: they multiply the poles.
Layer 6 — Distribution and tools: the front door
Microsoft holds a singular position: the company owns GitHub, where a large part of the world’s code resides, and the widely used VS Code editor, in addition to its Copilot assistant. Hugging Face, for its part, has hosted more than three million public models since August 2026: it is the public square of open models, including heretical ones. On September 3, 2026, Nvidia announced its intention to acquire Hugging Face for about 13 billion dollars; if the deal goes through, the main public square of open models will belong to the main chip designer, which raises questions about its role as a counterweight.
Among developers, no tool dominates alone. According to the JetBrains survey of more than 15,000 professional developers (May to July 2026), 90% used coding agents at least weekly. Claude Code was used by 39% of them (up from 18% in January), GitHub Copilot by 21%, Codex by 16% (up from 3% in January), Cursor by 12% (down from 18%) and the open-source tool OpenCode by 7%. These shares changed sharply within a few months.
Layer 7 — Integrators: who brings AI into organizations
A model transforms nothing until it is integrated into the systems of an administration, an army or a company. This layer is often invisible, but it decides the real use of AI.
Palantir popularized the model of the forward deployed engineer: engineers placed at the client’s site to adapt the technology to its needs, which makes it possible to scale. In the first quarter of 2026, according to Bloomberg, its revenue had grown by 85%, driven by both the American government and businesses, and its Maven system was set to become an official Pentagon program.
The laboratories themselves are adopting this model. In May 2026, OpenAI launched a deployment subsidiary based on engineers deployed at client sites, and Anthropic had just announced a comparable 1.5-billion-dollar structure, backed by investment funds. The laboratories therefore no longer merely supply models: they also seek to control their integration, which reinforces vertical integration.
Large consulting and IT services firms integrate AI for companies and states, often combining several providers. They can act as a counterweight, provided they remain independent of the laboratories.
Integrating AI agents are beginning to do part of this work themselves: configuring, connecting, deploying. In time, AIs could integrate other AIs, which directly raises the question of governance between AIs.
The human integrator remains decisive. Integration is not only technical: one must understand real needs, constraints, people, and decide what should not be automated. In the charter’s vocabulary, the human integrator is a full actor of the network, and not a mere relay between AIs and those who use them. When the integrator depends on a single provider, however, it can become the vector of vassalage (see Article P-5).
Layer 8 — States and bodies: three rival forums
United States. An executive order of December 11, 2025 aims to make a light-touch federal standard prevail over state laws on AI. The United States also uses export controls: on June 12, 2026, a Department of Commerce directive barred access to the Claude Fable 5 and Mythos 5 models for any person of foreign nationality, including on American soil; unable to separate its users, Anthropic suspended these two models for all its customers worldwide, the United States included. The controls were lifted on June 30, and access was restored on July 1.
European Union. The AI Act applies, but the “Omnibus” regulation that came into force on July 27, 2026 postponed the obligations for high-risk systems to December 2027 and August 2028. The obligations for general-purpose AI models were not postponed, and the AI Office received investigative and sanctioning powers close to those of a competition regulator.
China. In July 2026, at the World Artificial Intelligence Conference in Shanghai, China launched a World AI Cooperation Organization (WAICO), whose founding agreement was signed by 29 countries on July 16, then joined by 8 others shortly afterward, for a total of 37 members, and championed the openness of models. Its voting rules and resources are not yet public.
UN. An independent international scientific panel of 40 experts, co-chaired by Yoshua Bengio and Maria Ressa, published a preliminary report on July 1, 2026, and the first Global Dialogue on AI Governance was held in Geneva on July 6 and 7. These bodies do not produce binding rules.
D-2 — Reading through the lens of the charter
The real ecosystem illustrates almost every notion of the charter.
The loop. The race to AGI has all the characteristics of the loop: no position is secure, rankings are reshuffled within months, and each actor must run so as not to fall back. This situation provides a partial analogy with Premise 1 of the theorem (see Appendix A); it does not prove the theorem.
The protection market. The worldwide suspension of June 2026, for regulatory reasons, like the access cutoffs imposed by some providers on third-party tools (thus, in February 2026, Google suspended, often without notice, access to Antigravity for paying customers who used it with tools such as OpenClaw or OpenCode), show that a user or a region that depends on another’s models, tools or clouds can be deprived of them overnight. The European reaction, which strengthened the case for Mistral, is that of a weak pole refusing vassalage (see Article P-5).
The plunderers. Models whose refusals have been reduced are not, in themselves, plunderers; but they can serve those who are, and they are born of a structure: the demand for AI without restrictions and the impossibility of controlling what has been published.
The center. Vertical integration (SpaceX, xAI and Cursor; Microsoft, GitHub, VS Code and Copilot; Google, its chips, its cloud, its models and its products; the laboratories creating their own integration subsidiaries) recalls the iteration of the center (see Article I-3): an actor that becomes indispensable to every layer gradually turns the network into a star.
The chokepoint. Chips are the Earthly equivalent of the “granary” (see Article I-4): whoever controls the scarce resource can deprive others of what is necessary.
A triptych within the triptych. The governance forums themselves form a triptych: the United States, China, and a third pole made up of Europe and the UN. Each proposes its rule and its forum; none is accepted by all.
D-3 — The counterweights that already exist
As the author requested, this part sticks to what works in practice, and not to what would be desirable in theory.
1. Rivalry between laboratories. No laboratory dominates the frontier alone. Market shares move fast, which for now prevents any lasting hegemony. 2. Open-weight models. They make it much harder for a single actor to confiscate capabilities, within the limits of their licenses and of the hardware needed to use them. This is the most solid counterweight, and the hardest to govern. 3. Sovereign models and clouds. They multiply the poles and reduce dependence, even though they lag behind. 4. The split into two hardware blocs. It prevents a single actor from controlling all the world’s chips, at the price of a geopolitical rivalry. 5. The diversity of tools. Developers use several agents at once and switch tools quickly. Distribution shows a diversity of offerings, without ruling out switching costs or dependencies. 6. The plurality of integrators. Independent firms, human integrators and multi-vendor sourcing allow an organization not to depend on a single laboratory. 7. Regulators with real powers. The European AI Office can investigate and sanction; the United States controls exports. These powers are imperfect and sometimes contradictory, but they exist.
The weaknesses are symmetrical: the extreme concentration of chips, the colossal investments in clouds, which make entry almost impossible for newcomers, and vertical integration.
D-4 — What the study suggests for Title IV
The study provides elements consistent with the author’s intuition, without demonstrating that the proposed trajectory is unique or that it will come about: in practice, it is not a declaration of principles that will settle the race to AGI, it is the structure of forces. It suggests a trajectory in three stages.
Stage 1 — Balance. This is the present situation: an oligopoly of laboratories, two hardware blocs, open and sovereign counterweights. No actor can prevail alone, which makes cooperation more rational than conquest. But the balance is unstable: it can tip toward hegemony through a breakthrough, a vertical integration or control of chips.
Stage 2 — Democracy. For the balance not to tip, institutions are needed that enforce the rules. The study shows that sanctioning powers already exist piecemeal (European AI Office, export controls), but that they are scattered among rival forums. The most enforceable mechanisms seem to be those that rest on the actors’ own interest:
interdependence: no actor controls all the layers, which makes defection costly;
mutual verification: AIs and laboratories that evaluate one another’s models, just as the charter is reviewed by several AIs;
multiple sourcing: states and companies that refuse to depend on a single provider;
a de facto power cap, imposed less by law than by the scarcity of computing and energy.
Stage 3 — The network. As harmony progresses, central institutions can fade away in favor of direct relations between poles, without a permanent center (see Article I-3). The study does not make it possible to say when or how this transition might happen; that is the subject of Title IV.
Questions answered by Title IV
1. Open-weight models as a common good, despite the risk of heretical models (see Article IV-12). 2. Openness of the layers against vertical integration (see Article IV-13). 3. The forum of the democratic stage: the Assembly of the Three Poles (see Article IV-15). 4. Access to the necessary computing and hardware diversification (see Article IV-14).
Blind spots of this study
A cross-review identified several limits, which the study acknowledges and which a future update will have to address.
Hardware. Speaking of “two chip blocs” simplifies a transnational chain that also includes AMD, high-bandwidth memory, advanced chip packaging, equipment makers, networks and energy suppliers.
Regions. India, Japan, Korea and Africa deserve a place of their own, beyond sovereign models alone.
Labor and data. Annotators, moderators, creators whose works are used for training, workers affected by automation and communities whose languages are poorly represented hardly appear in the study.
Material costs. Energy, water, cooling and the siting of data centers make computing a material resource, not an abstraction.
Military and public uses. They must be examined with the same framework for all providers, without any single actor becoming their representative.
Financial dependencies. A plurality of brands can rest on the same funders and the same infrastructure; one would need to distinguish cross-investments, computing purchases, exclusivity contracts and revenue actually realized.
Neutrality. The study was drafted with the help of Claude, developed by Anthropic, one of the actors described. The same analytical framework, covering safety thresholds, audits, incidents and measures actually applied, should be applied to all laboratories.
Sources
SpaceX acquires Cursor for $60 billion — Techzine, June 16, 2026
SpaceX officially completes acquisition of Cursor — Gigazine, August 16, 2026
Nvidia to buy Hugging Face for nearly $13 billion — Semafor, September 3, 2026
OpenAI, Anthropic, Gemini: revenue and market shares (in French) — AlgeriaTech, April 16, 2026
Mistral and European AI sovereignty — Fortune, August 5, 2026
AI model landscape 2026: open weights and the country race — DEV Community, July 2026
China’s AI chip deficit — Council on Foreign Relations, December 15, 2025
Nvidia largely conceded China’s AI chip market to Huawei — May 2026
Moody’s hyperscaler capex forecasts — DatacenterDynamics, 2026
Europe’s sovereign cloud spend set to triple — The Register, February 9, 2026
Palantir lifts annual revenue forecast — BNN Bloomberg, May 5, 2026
OpenAI launched a consulting arm to help companies deploy AI — Decrypt, May 11, 2026
EU Digital Omnibus on AI enters into force — Mondaq, August 3, 2026
Executive order advancing federal preemption of state AI laws — Akin Gump, December 2025
AI coding agent adoption in 2026 — JetBrains Research, August 2026
Fable 5 suspension: facts and timeline — EisnerAmper, June 2026
After the World Data Organization, China Unveils WAICO — SESEC, August 7, 2026
Export controls on Claude Fable 5 and Mythos 5 terminated — Digital Policy Alert, 2026
Xi sets four-part global AI governance plan at WAIC 2026 — Let’s Data Science, July 2026
AI coding agent adoption in 2026 — JetBrains Research, August 2026
Hugging Face Hub: 3 million models (in German) — Mindverse, August 19, 2026
Middle East leads sovereign AI LLM race in H1 2026 — Counterpoint Research, July 30, 2026