Focus Intermarium
Structural geopolitical analysis between the Baltic and the Black Sea
2035

Contours of the New World

Thesis

Two technology cores and a potential third, the EU; four state AI architectures; partial mobilization autarky; and hinge states.

≈ 31 min read · 20 sections

01 / POWER
Unequal power centers
Several power centers will operate on top of two technology cores.
02 / AUTARKY
The cost of autarky
Both cores could survive mutual disconnection, but neither could reproduce the former efficiency of the whole economy.
03 / HINGE STATES
Hinge states
Third countries’ resources, markets, and routes will become the main arena of bargaining between the cores.
04 / CONTROL PLANE
Models are replaceable; the integration layer is not
Data, permissions, ontologies, and workflows can create deeper lock-in than any single frontier model.
I

Key Judgment

In brief By 2035 the world becomes an armed technological polycentrism: there are more power centers than technology cores, and the technology stack becomes the basic unit of politics.

By 2035, the world is unlikely to have settled into either a new, stable liberal order or a simple US–China bipolar system. Instead, it will take the form of armed technological polycentrism:

  • several unequal ;
  • two —the United States and China;
  • the EU as a potential third core;
  • militarized regional systems;
  • controlling resources, markets, and routes needed by multiple power centers at once;
  • constant bargaining over infrastructure, , energy, logistics, and access rules.

There will be more power centers than technology cores. States will retain their flags, armed forces, and diplomatic room for maneuver, but their effective freedom of action will depend on external control over chips, clouds, models, payment channels, energy routes, and military infrastructure.

The basic unit of world politics is no longer the isolated state or the formal alliance. It is becoming the : what a state receives from a single core or assembles from several—chips, clouds, models, and updates, and with them security, finance, logistics, and standards.

The American and Chinese cores may approach mobilization autarky—the ability to continue functioning and arming themselves after mutual cutoff. But neither will be able to reproduce the entire modern technological ecosystem internally without major penalties in cost, quality, and the pace of innovation. Direct links between the two cores may therefore contract sharply without eliminating interdependence: it will shift into third countries that control raw materials, energy, food, production sites, and transport hubs.

Within states, AI will be organized not as a market of stand-alone models but as a state-corporate AI . Its layers will include interchangeable models, accredited cloud and compute, data and access rights, , application workflows, and government demand. Four distinct architectures are already visible: the United States preserves competition among frontier models while concentrating the integration of mission workflows; China combines central regulation and compute with state-owned enterprises, telecom operators, and municipal platforms; the EU combines supranational rules and funds with distributed national deployments; and Russia is building a protected market around the state, Sber, and Yandex while facing constrained access to advanced accelerators.

Confidence: high in the direction of structural change; medium in the configuration of power centers; low in the precise boundaries of future alliances and the form that settlements of ongoing wars will take.

II

Political Multipolarity Will Overlay a Technological Oligopoly

In brief There are several power centers but only two technology cores, the US and China. The American core is anchored by an integrator, the Chinese one by state coordination of a federation of platforms.

By 2035, several power centers are likely to exist, but only two technology cores: only the United States and China will be positioned to claim a near-complete set of advanced AI and related technologies.

United States

The United States will retain the most complete portfolio of systemic power:

  • global military reach;
  • dollar and sanctions infrastructure;
  • advanced semiconductor design;
  • the largest cloud and model platforms;
  • a network of allies controlling critical links in production;
  • the ability to license other countries’ access to the technology stack.

The US strategy is increasingly built around exporting its stack, a linked package of “chips + models + software + applications + standards”. Access to this package is becoming an instrument of foreign policy and can be exchanged for security guarantees, reciprocal investment, restrictions on Chinese presence, and compatibility with US rules. US agreements with the UAE demonstrate an early form of this exchange.

US policy, however, will remain fluid and transactional. The withdrawal of the AI Diffusion Rule before it took full effect, and the shift to case-by-case licensing for selected accelerator shipments to China, show that controls will be used not as an immutable boundary but as a bargaining instrument.

The US State-Corporate AI Architecture

US defense AI is already being built around something other than a single “national model.” CDAO awarded OpenAI a prototype OTA with a ceiling of $200 million but an initial obligation of about $2 million, and subsequently assembled a portfolio comprising Anthropic, Google, OpenAI, and xAI, each with a separate ceiling of up to $200 million. This indicates that the government intends to preserve competition among closed frontier models and avoid dependence on a single laboratory. These contract ceilings must not be added together as actual spending: they define authorized capacity, not guaranteed procurement volume.

CDAO → OpenAI: prototype OTA
≈ $2M
initial obligation against a $200M ceiling — about 1%

By 2026, the multivendor principle had been extended to classified networks: agreements for IL6/IL7 covered SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, AWS, and Oracle. The official objective is lawful operational use without AI vendor ; financial terms and IDs have not been disclosed. The names xAI/AIQ Phase LLC in 2025 contracts and SpaceX in the 2026 list refer to different observed records and must not be retrospectively conflated into “SpaceX AI”.

But multivendor sourcing at the model layer does not mean multivendor sourcing across the entire architecture. Claude 3/3.5 became available to defense and intelligence users through an Anthropic–Palantir AIP–AWS/SageMaker configuration in an accredited environment. Here, Palantir serves not as another model laboratory but as the operational : it connects the model to classified data, access rights, ontology, and mission workflows.

The trajectory of Maven demonstrates the scale of this layer’s institutionalization. Palantir’s initial contract for the Maven Smart System had a ceiling of $480 million; less than a year later, total capacity was increased to $1.275 billion. In July 2025, the Army consolidated 75 contracts into a ten-year enterprise agreement with a maximum potential value of $10 billion through 2035. All three figures are ceilings, not expenditures, but this sequence of decisions lowers the cost of subsequent procurement and entrenches Palantir as the platform for deploying and managing operational applications.

Palantir Maven Smart System: contract ceilings
Maximum value, US dollars. All three figures are ceilings, not spending.
Initial contract
$480M
Less than a year later
$1.275B
July 2025: Army enterprise agreement through 2035
$10B

The US architecture therefore reduces lock-in at the individual-model layer while potentially increasing it at the level of data, ontology, permissions, and applications. For allies, this accelerates interoperability with the United States but expands their dependence beyond US chips and clouds to include whoever controls the integration layer and updates to mission workflows.

US: a single integration hub
Relationships as described in the text. Hover over or tap a box to see its links.
Rules & direction
Compute & capital
Demand & procurement
Models
Integration
Users
External
Select a box to see its links.
All links as a list
  • CDAO → OpenAI: prototype OTA: $200M ceiling
  • CDAO → Anthropic: CDAO portfolio: ceiling up to $200M
  • CDAO → Google: CDAO portfolio: ceiling up to $200M
  • CDAO → xAI: CDAO portfolio: ceiling up to $200M
  • Anthropic → Palantir AIP / Maven: Claude via Palantir AIP
  • Palantir AIP / Maven → AWS / SageMaker: runs in the AWS/SageMaker IL6 environment
  • US Army → Palantir AIP / Maven: enterprise agreement: up to $10B through 2035
  • Palantir AIP / Maven → Defense & intelligence: control plane: data, access rights, ontology, workflows
  • IL6/IL7 classified networks → Defense & intelligence: multivendor access to classified networks
  • US export policy → Allies (UAE): stack exports traded for security guarantees, investment, and limits on Chinese presence

China

China will remain the only systemic competitor to the United States capable of developing all of the following in parallel:

  • national compute infrastructure;
  • domestic chips and equipment;
  • mass sectoral deployment;
  • its own models and an open-source ecosystem;
  • exports of technology, expertise, and standards;
  • control over critical materials.

China’s strategy combines the domestic mobilization of compute, a broad “AI+” deployment program, development of domestic hardware and model layers, and an international offer to partners encompassing data, compute, open models, workforce training, and standards.

China’s State-Corporate AI Architecture

The observable Chinese architecture is neither a single vertical centered on DeepSeek nor a system with a single equivalent to Palantir. It is federated but state-coordinated. The CAC sets the market-access regime for public generative services; the filing procedure (bei’an) does not in itself signify government procurement, access to restricted data, or operational deployment within an agency. The progression from the 2023 interim rules to a public registry mechanism and accumulated figures of 1,112 services and 731 applications demonstrates the scale of the regulatory layer, not the number of frontier models or government deployments.

The compute backbone is being built by the NDRC/NDA and operators of the national integrated computing network. By August 2026, the government reported aggregate capacity of 2.45 million FP16, of which 1.45 million PFLOPS was connected to the monitoring layer; these figures do not demonstrate node interchangeability, access to advanced accelerators, or the ability to train frontier models across the full nominal capacity. A national AI fund with announced capital of 60 billion yuan and regional subsidies reduce the cost of ecosystem development, but announced capital and maximum subsidy levels are not actual expenditures.

China: national computing network, August 2026
1.45M PFLOPS
connected to the monitoring layer, out of 2.45M PFLOPS FP16 aggregate capacity — about 59%

SASAC is organizing demand among central state-owned enterprises. A dedicated “AI+” initiative was launched in February 2024 and was later supplemented by DeepSeek integration and a registry of 40 scenarios. This creates a channel for scaling through sectoral data, telecommunications, energy, and industry, but corporate announcements of integration do not disclose the intensity or effectiveness of operational use.

At the execution layer, state clouds, telecom operators, municipal data groups, and platforms are at work. Beijing Yizhi combines Baidu, Zhipu, and other models; Nanjing uses Qwen and DeepSeek in its 12345 public-service hotline system and in an internal application for officials; and the National Bureau of Statistics provides a service based on Qwen and DeepSeek. This confirms the emergence of multi-model control planes, in which the developer of the foundation model is not necessarily the lead systems integrator.

There are also narrowly scoped, confirmed workflows: the Lucheng District archive reported processing 10,788 files while retaining manual review; within the judicial system, Faxin and the Shenzhen court system have been confirmed. At the same time, the 19,080,666-yuan award for Guangzhou’s unified AI platform remains an award, not proof of payment and acceptance; PLA tender 2025-JQ02-F1290 records procurement intent, not military or combat use.

China’s advantage, therefore, lies not only in possessing models. It is the ability to regulate market access, build a compute and data backbone, direct capital, generate demand from state-owned enterprises, and distribute interchangeable models through multiple regional and sectoral control planes—all at once. Its weak points are the advanced hardware layer and the opacity of actual system utilization. Beijing will not necessarily seek to surpass the United States in every component; rather, it will aim to secure sufficient domestic autonomy and make its stack the most accessible alternative for states that either cannot obtain US technology or will not accept US terms.

China: federated, state-coordinated
Relationships as described in the text. Hover over or tap a box to see its links.
Rules & direction
Compute & capital
Demand & procurement
Models
Integration
Users
External
Select a box to see its links.
All links as a list
  • State AI strategy → SASAC → state-owned firms: the “AI+” program: mass sectoral deployment
  • State AI strategy → Partner countries: international offer
  • CAC → DeepSeek: market access for public services (bei’an)
  • CAC → Qwen: market access for public services (bei’an)
  • CAC → Baidu (ERNIE): market access for public services (bei’an)
  • CAC → Zhipu (GLM): market access for public services (bei’an)
  • SASAC → state-owned firms → DeepSeek: DeepSeek integration in state-owned firms
  • Baidu (ERNIE) → Beijing Yizhi: model on the Beijing Yizhi platform
  • Zhipu (GLM) → Beijing Yizhi: model on the Beijing Yizhi platform
  • Qwen → Nanjing: 12345: 12345 hotline
  • DeepSeek → Nanjing: 12345: 12345 hotline
  • Qwen → Statistics bureau: statistics bureau service
  • DeepSeek → Statistics bureau: statistics bureau service

The Main Axis of Competition

The systemic contest will not be fought only over the best model or the most advanced chip. What is at stake is the number of countries whose government, industrial, and military processes operate within either the American or the Chinese stack.

This will create an unusual structure: a multipolar world built on two technology cores.

The Limits of Autarky

The American and Chinese cores have asymmetric bottlenecks relative to each other. The United States and its allies control a substantial share of advanced computing chips, fab equipment, and manufacturing know-how; China controls a substantial share of critical-material processing and intermediate-component production. This is already reflected in mirror-image export-control regimes: US, Dutch, and Japanese restrictions on chips and equipment, and Chinese controls on gallium, germanium, and rare-earth materials.

Almost any such bottleneck, however, can be partially offset through some combination of new capacity, recycling, technological substitution, higher energy consumption, and demand rationing. The cores do not need a particular material or machine as such; they need the function it performs. By 2035, both cores will therefore probably be able to sustain the critical minimum required by the state and the defense-industrial base without direct supplies from each other, but not the previous level of economy-wide efficiency.

The chief irreplaceable resource is the time required to build an ecosystem: qualified suppliers, tacit manufacturing knowledge, failure statistics, servicing capacity, scale, and trust. It cannot be rapidly replaced by a single mine, factory, or subsidy. Full autarky will mean duplicated capacity, higher costs, lagging civilian sectors, and priority allocation of scarce components.

The likely objective of both cores, therefore, is not the complete absence of trade but the ability to survive a cutoff and set their own terms of access. Case-by-case licensing of accelerator shipments to China and export safeguards show that even strategic restrictions may remain negotiable rather than forming an absolute curtain.

III

The EU Can Become a Third Core—but Only Through Production and Deployment

In brief The EU has rules, a market and money, but it becomes a third core only if it turns sites, grants and awards into working compute and mass deployment.

The European Union has the market, industrial base, capital, research institutions, and regulatory power needed to claim the role of a third technology core. Its AI architecture differs fundamentally from both the American federal-corporate demand model and Chinese party-state coordination: supranational institutions set the rules and provide funding and some of the compute capacity, while operational deployment is distributed across the European Commission, member states, regions, and municipalities.

Regulation and Organized Demand

The AI Act, the European AI Office, and the GPAI Code create a supranational compliance layer. It can influence foundation-model developers and centralize oversight of , but it is neither procurement nor government deployment. The 2026 AI Omnibus both expanded the AI Office’s role and postponed the application of the main high-risk rules: for Annex III, until December 2027; for systems embedded in regulated products under Annex I, until August 2028. This makes the trajectory more flexible but lengthens the gap between the rules and their full enforcement.

The AI Continent Action Plan and the Apply AI Strategy seek to convert regulatory power into demand by linking , gigafactories, data labs, open source, AI-first, and buy-European priorities. But an EU-wide strategy does not mean uniform implementation across all member states, and buy-European is not an automatic procurement quota.

Compute and Capital

EuroHPC is building a distributed backbone. After the first seven AI Factories, two further batches of six sites each were selected, bringing the network to 19. Site selection does not constitute delivery or commissioning: the hosting agreement, tender, vendor contract, delivery, acceptance, and operational access remain separate stages.

The transition to the physical layer is already visible. EuroHPC signed a contract with HPE for HammerHAI and its sixth next-generation contract—with Bull for LUMI-AI. But a signed contract is still not an operational supercomputer. As of the cutoff date, AI Gigafactories remain at an even earlier stage: the open call allows for up to seven projects and announces up to €10 billion in public support and at least €20 billion in mobilized private investment, but these are not awards or commissioned capacity.

At the model layer, the EU is funding OpenEuroLLM, with a total project budget of €37.4 million and a €20.6 million contribution from Digital Europe. This is a significant attempt to connect open models, European languages, and public compute, but a grant and a STEP Seal do not prove the existence of a competitive, deployment-ready model or its operational use by government.

Sovereign Cloud and Procurement

The Cloud III DPS creates a shared procurement channel for EU institutions, with a planned aggregate value of €550 million. Under it, the Commission announced a sovereign-cloud competition with a €180 million ceiling and subsequently awarded four framework contracts. This sequence demonstrates a transition from policy → framework → tender → award, but even an award does not prove the actual volume of orders, production use, or the hosting of sensitive and classified workloads.

Europe’s architecture is deliberately multivendor. Awardees and partners include POST Luxembourg, STACKIT, Scaleway, Proximus, and Mistral, while one consortium uses a controlled sovereign layer built on Microsoft technology. EU digital sovereignty should therefore be understood not as mandatory technological autarky, but as control over jurisdiction, keys, operations, portability, and vendor dependence.

Real Government Workflows

Operational use already exists, but it does not form a single European control plane. Following its internal AI@EC strategy, the European Commission launched GPT@EC, a secure multi-model tool for its own staff. In France, Albert was confirmed as a pilot used by approximately 60 staff members across about 30 France Services offices; the announced general rollout and future tax, judicial, or medical workflows do not count as a completed rollout.

Baden-Württemberg deployed F13 in the state’s data center and later released a production application for reuse across public agencies. Aleph Alpha participated in the prototype but not in the full version, demonstrating that the application layer can be separated from the original model vendor; releasing it as open source, however, does not prove adoption by other authorities. Hamburg took a different route: LLMoin runs on GPT through Azure, combining a government-owned interface and governance with an American model and cloud layer.

Luxembourg moved from a strategic partnership with Mistral to a confirmed on-premises deployment, although the promised access to a sovereign chatbot for all civil servants had not yet occurred as of the report date. The French Ministry of the Armed Forces concluded a framework agreement with Mistral under AMIAD’s coordination; it opens a procurement channel but does not prove task orders, acceptance, operational use, or combat employment. Three Digital Europe GenAI pilots for public administrations signed grant agreements, but national procurements and deployments must be documented through subsequent events.

EU: strong at the top, fragmented below
Relationships as described in the text. Hover over or tap a box to see its links.
Rules & direction
Compute & capital
Demand & procurement
Models
Integration
Users
External
Select a box to see its links.
All links as a list
  • AI Act · AI Office → OpenEuroLLM: compliance layer for GPAI
  • AI Act · AI Office → Mistral: compliance layer for GPAI
  • AI Act · AI Office → Aleph Alpha: compliance layer for GPAI
  • AI Continent · Apply AI → EuroHPC: AI Factories
  • AI Continent · Apply AI → AI Gigafactories: gigafactories
  • Digital Europe → OpenEuroLLM: €20.6M of €37.4M
  • Cloud III DPS → Cloud contractors: four framework contracts
  • Mistral → Luxembourg: on-premises deployment
  • Mistral → French armed forces ministry: framework agreement (AMIAD)
  • Aleph Alpha → Baden-Württemberg: F13: prototype only
  • US stack → Hamburg: LLMoin: GPT via Azure
  • US stack → Cloud contractors: Microsoft technology in one consortium

Europe’s Limit

Europe’s strength lies in diversification: multiple developers, public compute, national clouds, open source, and the ability to shape the global market through regulation. Its weakness is a long implementation chain and the absence of a single integration layer. Between strategy and a functioning workflow lie national budgets, procurement procedures, data interoperability, energy shortages, and divergent models of digital sovereignty.

The same logic applies in the military domain. Russia retains short-term advantages in mobilization speed and cost but risks losing a prolonged contest to Europe’s institutionalized industrial potential. The material test is not declarations about percentages of GDP, but budgets, contracts, production lines, personnel, stockpiles, and reinforcement infrastructure.

If the EU brings both its computing and its military base into working order by 2035, it will become a third technology core and an autonomous power center. If not, it will remain a powerful market and regulator, and a collection of national AI islands dependent on the American stack and American security guarantees.

IV

Russia Will Remain a Military Power but Risks Becoming a Technologically Dependent Power Center

In brief Russia keeps its military weight, but its protected AI market runs into a compute ceiling and depends on foreign hardware.

Russia possesses a nuclear arsenal, a defense industry, a mobilization apparatus, raw materials, territory, and experience adapting to a prolonged war. This is sufficient to preserve its role as an autonomous power center.

But its long-term position is complicated by three factors:

  1. Europe’s economic and industrial potential is substantially greater;
  2. access to the advanced American stack is restricted;
  3. redirecting former European energy flows to China requires costly infrastructure and Beijing’s consent.

The most rational Russian strategy is therefore a combination of:

  • inexpensive mass-produced weapons;
  • asymmetric capabilities;
  • pressure on the political cohesion of Western coalitions;
  • attempts to constrain European deployment through a pause and arms control;
  • deeper technological and economic ties with China.

Russian military-economic analysis already acknowledges the risk of a prolonged contest with a larger Europe and shows an interest in constraining military infrastructure before Europe fully mobilizes. Yet it still does not assign Belarus a distinct status, retaining the formula of a common “Russia/Belarus” space.

Russia’s State-Corporate AI Architecture: A Protected Market Under a Compute Ceiling

By 2026, the Russian architecture had taken on a state-corporate form. The president sets the strategy and issues directives; the government and the AI Development Center coordinate deployment; Sber/GigaChat and Yandex/YandexGPT are the two main model developers. The updated strategy through 2030 linked foundation and generative models to domestic infrastructure and compute resources, while a separate AI Development Center was tasked with coordinating the federal project and monitoring deployment.

Government demand is becoming the principal scaling mechanism. GigaChat has already been used to classify submissions to the president’s “Direct Line,” although company-reported user and quality metrics are not independent benchmarks. The 2026 directives called for a national deployment plan, funding for foundation models and the electronic component base, and priority for Russian solutions in public administration and critical information infrastructure. Law No. 243-FZ institutionalized the categories of large foundation, sovereign, and national models, along with support measures and access to government data; its practical effect will depend on implementing regulations and the conferral of these statuses.

The principal constraint is compute. Sber’s Christofari Neo and supercomputing capacity of the same generation were built on NVIDIA A100 GPUs. Following the broad restrictions imposed in 2022, the United States separately extended licensing requirements to the A100, H100, DGX, and subsequent products above specified thresholds. Gray-market supplies, inventories, and Chinese accelerators can sustain operations, but they do not guarantee regular scaling of training to the global frontier.

Russia is seeking to bypass this constraint by linking energy and infrastructure. Authorities have been directed to plan data centers through 2036 near available generating capacity and to include data centers in nuclear export projects. This could create a specialization in “nuclear generation + data centers + national models”, but megawatts cannot substitute for accelerators, interconnects, memory, and the software ecosystem.

Russia is building its external ties through BRICS and the AI Alliance Network. These ties could provide standards, application markets, sector-specific models, and technology diplomacy, but they do not yet include a confirmed shared compute pool, a binding budget, or a joint frontier model.

By 2035, Russia will most likely remain a major regional AI power in the upper second tier globally: strong in the Russian language, government and banking deployment, cybersecurity, industry, energy, and selected dual-use systems, but without sustained parity with the United States and China in frontier-model training. A protected market increases the speed of government deployment and the resilience of Russia’s architecture, while simultaneously deepening concentration around Sber and Yandex and dependence on an external hardware layer.

Russia: a vertical chain on a borrowed hardware base
Relationships as described in the text. Hover over or tap a box to see its links.
Rules & direction
Compute & capital
Demand & procurement
Models
Integration
Users
External
Select a box to see its links.
All links as a list
  • President → Government & AI Development Center: directives and a national deployment plan
  • President → Data centers to 2036: directive to plan data centers to 2036
  • President → BRICS / AI Alliance Network: external ties
  • Law No. 243-FZ → Sber: GigaChat: model statuses and access to government data
  • Law No. 243-FZ → Yandex: YandexGPT: model statuses and access to government data
  • Government & AI Development Center → Public administration & CII: priority for Russian solutions
  • Christofari Neo → Sber: GigaChat: Sber’s compute base
  • US export licensing → Christofari Neo: restricted access to A100/H100
  • Gray imports, stockpiles, Chinese accelerators → Christofari Neo: sustaining operations
  • Sber: GigaChat → “Direct Line”: classifying submissions
Four state AI architectures, layer by layer
Assessment based on sections II–IV. Hover over or tap a cell to see the reasoning.
StrengthPartialWeak linkNot assessed
LayerUSChinaEURussia
Chips & equipment
Compute
Models
State demand
Integration layer
Select a cell in the table to see the reasoning behind the assessment.
V

India and the Gulf States Will Become Distinct Power Centers

In brief India is building a mass AI ecosystem outside the US–China pair, while the UAE and Saudi Arabia buy autonomy with capital and energy.

India

India is the most likely candidate to build an autonomous mass-market AI ecosystem outside the United States–China pair. The IndiaAI Mission combines shared compute, national models, data, talent, and safety tools. The record shows a shift from a procurement configuration of 18,693 GPUs to a government announcement of more than 45,000 GPUs and an expansion of AIKosh.

IndiaAI Mission: shared compute
> 45,000 GPUs
per government announcement; the original procurement configuration was 18,693 GPUs

By 2035, India could become:

  • a provider of models and services for the Global South;
  • a platform for alternative AI diplomacy;
  • a major deployment market;
  • a US partner in trusted supply chains;
  • an autonomous actor not fully aligned with either stack.

Its hardware dependence will persist longer than its dependence at the model and application layers.

The UAE and Saudi Arabia

The UAE and Saudi Arabia are converting capital, energy, and foreign-policy flexibility into compute capacity. The UAE is using government-backed models and trusted-AI agreements; Saudi Arabia has established HUMAIN and is expanding its planned AI data-center network, including a 250 MW agreement.

Their likely role:

  • major sites for hosting compute;
  • investors in American, European, and Asian technology companies;
  • providers of regional cloud and model services;
  • intermediaries between technology cores;
  • holders of the energy resources required for AI growth.

They will not replace the United States or China, but they will be able to purchase substantial autonomy and influence the terms on which technology spreads.

VI

Hinge States Will Convert the Cores’ Interdependence into Agency

In brief A hinge state controls a function several power centers need and can change the terms of access, provided it keeps the decision and the rent at home.

A hinge state is not merely an intermediary, a “bridge,” or neutral territory. It is a state that controls a function needed by several competing power centers, retains access to at least two of them, and can alter the terms of that access. Such an asset may be energy, food, minerals, a market, a strait, a land corridor, a manufacturing platform, capital, talent, or a politically acceptable jurisdiction.

A hinge role requires four attributes at once:

  1. a scarce asset—loss of access to it would materially weaken one or more power centers;
  2. multiple buyers—the state can redirect flows rather than submit to a single patron;
  3. control over the decision—an external power cannot control or allocate the asset without the local state’s consent;
  4. domestic capacity to retain rents—revenue is converted into infrastructure, capabilities, and freedom of maneuver, rather than merely into profits for elites or a foreign operator.

Not every commodity exporter is a hinge state. If a deposit, port, or route is controlled from abroad, and the supplier cannot change partners or terms, it remains a resource enclave or client. Hinge-state status is not an attribute of geography alone, but the capacity to convert geography and scarcity into the political right to choose.

The most likely types of hinge states by 2035 are:

  • India—market scale, talent, digital services, and the ability to participate in trusted Western supply chains without fully relinquishing its autonomy;
  • the Gulf states—energy, capital, sites for compute, and the ability to connect American technology with Asian demand;
  • Brazil and Indonesia—a combination of food, minerals, industrial scale, markets, and foreign-policy flexibility;
  • the Central Asian states—uranium, metals, energy, and overland routes among China, Russia, Europe, and South Asia;
  • transit powers and jurisdictions—control of straits, ports, cables, payment channels, and re-export routes.

Their importance will increase precisely as direct United States–China ties contract. Interdependence will not disappear but become triangular: both technology cores will seek access to the same resource periphery, offering security, infrastructure, and a technology stack while simultaneously demanding that states there limit the rival’s presence. The political map of 2035 will therefore be defined not only by the boundaries of alliances, but also by which hinge states can refuse an exclusive alignment.

The principal risk for a hinge state is becoming an arena of coercion. The more critical its asset, the stronger the pressure exerted through sanctions, changes in licensing terms, control of logistics, domestic elites, and military infrastructure. Resilient hinge-state status requires diversifying not only buyers, but also routes, finance, technology, and security guarantees.

VII

Sovereignty Will Become Modular

In brief Sovereignty splits into separate domains (military, energy, finance, compute), and in each a country can be the client of a different power center.

By 2035, formal independence will be an increasingly poor measure of a state's actual agency. Sovereignty will have to be disaggregated across distinct domains:

  • military command;
  • foreign forces and nuclear systems;
  • energy;
  • finance and payments;
  • logistics and ports;
  • compute infrastructure;
  • government data;
  • models and software updates;
  • human capital;
  • the ability to sign and implement agreements independently.

Belarus already provides an instructive case of such modular sovereignty. The economic and political integration of the Union State remains incomplete, but military-nuclear integration has advanced considerably further: the use of Belarusian territory for the invasion of Ukraine, the abandonment of neutral and non-nuclear status, the deployment of Russian nuclear weapons and the Oreshnik system, and simplified military-technical exchange.

At the same time, Minsk maintains separate channels with the United States, the EU, China, and the Gulf states. But a multiplicity of contacts does not amount to agency. Focus Intermarium rightly measures multi-vector policy by resources actually secured, not by visits and memoranda.

This model will spread. A country may be:

  • the military client of one center;
  • the technology client of a second;
  • export-dependent on a third;
  • legally interoperable with a fourth;
  • formally independent in every respect.
VIII

Alliances Will Be Defined by Technology and Infrastructure Stacks

In brief Alliance membership is defined by chips, clouds, data and updates. Expect a hierarchy of access and hybrid stacks rather than a world split into two sealed zones.

Alliance membership will increasingly be determined by more than a mutual-defense treaty. The US experience shows that even competition among several frontier models does not eliminate dependency if one integrator controls the data, ontology, permissions, and mission applications. More revealing questions will be:

  • whose accelerators and fabrication equipment a country uses;
  • where government models run;
  • who has access to data and telemetry;
  • whose clouds support critical functions;
  • where updates come from;
  • which security standards are recognized;
  • whether an external provider can terminate service.

The United States uses export controls on advanced chips, HBM, and equipment; the Netherlands and Japan license critical manufacturing technologies. China is responding with controls on gallium, germanium, and rare-earth materials.

By 2035, a hierarchy of access is likely:

  1. trusted allies receive the advanced stack;
  2. conditional partners receive quotas and licenses subject to safeguards;
  3. competitors rely on aging generations, import substitution, and gray-market channels;
  4. the highest-risk actors face bans and tighter re-export controls.

The world is unlikely to split completely between the two cores into hermetically sealed zones. Many states will assemble hybrid stacks: US accelerators, Chinese open models, European regulation, and national data centers. This will create room for maneuver, but it will also create multiple external points at which access can be cut off.

Even if direct trade between the American and Chinese cores falls to a minimum, goods and technologies will continue to flow through processing, assembly, and re-export in third countries. Official statistics may indicate decoupling while functional dependency persists at the level of materials, machine tools, components, and service updates. Hinge states will simultaneously serve as circumvention channels, negotiating venues, and targets in a contest for exclusivity.

Alliance membership will therefore be measured not by the absence of every external connection, but by the answer to a harder question: can a state sustain critical operations after the opposing stack is disconnected and one major intermediary fails?

IX

Energy, Compute, and Security Will Converge into a Single System

In brief The chain of data centers, grids, generation, cooling and chips becomes strategic infrastructure and a potential target.

The data center — power grid — generation — cooling — chips nexus will become as strategic as an oil pipeline, port, or military airfield.

This will increase the importance of states capable of providing:

  • cheap, scalable energy;
  • rapid grid connections;
  • physically secure sites;
  • water and cooling;
  • capital;
  • expedited permitting;
  • access to advanced accelerators.

The EU already links technological sovereignty to the energy system, Canada links national compute to clean energy, and the United States and the UAE link trusted AI to energy cooperation.

The implication is that data centers, power grids, undersea cables, and semiconductor fabs will become assets covered by strategic deterrence—and potential targets.

X

War Will Become a Permanent Condition Between Peace and Total Mobilization

In brief War becomes a standing condition between peace and mobilization: strike, infrastructure shock, mediation, incomplete deal, contest over verification.

Focus Intermarium data reveal a recurring pattern of conflict:

  • long-range strikes without a formal declaration of war;
  • maritime and sanctions blockades;
  • cyberattacks and information operations;
  • strikes on production and civilian logistics;
  • brief ceasefires that do not resolve the causes of conflict;
  • negotiations conducted in parallel with coercive pressure.

The Iran Lesson

The US-Israeli operation against Iran demonstrated the limits of superiority. A decapitation strike did not destroy the regime; Hormuz retained its role as a source of systemic leverage; the largest release from strategic reserves did not offset the physical constraint on transit; and an armed convoy operation did not restore reliable shipping.

A ceasefire and memorandum emerged without achieving the key stated objectives: regime change did not occur, the stockpile of highly enriched uranium was not destroyed, and actual IAEA verification was absent.

A likely pattern in future wars is:

strike → continued governability → infrastructure shock → mediation → incomplete deal → contest over verification.

The Russia-Ukraine Lesson

The campaign of strikes against Russian refineries, export terminals, industrial facilities, and later civilian logistics networks shows that the target set is expanding from armed forces to the systems that sustain the home front. The series of attacks on Wildberries and Ozon warehouses demonstrates a network logic of target selection.

By 2035, the boundary between military and civilian infrastructure will become even less stable. Logistics, energy, clouds, data, and compute will be treated as a single system underpinning a nation's capacity to continue a war.

XI

AI Will Transform Warfare Through Integration, Not Through a Single “Supermodel”

In brief Advantage comes not from a single model but from integration: data, communications, command, cheap mass platforms and the integration layer.

AUKUS has already brought AI and autonomy into a shared military experimentation environment. US procurement is taking the next step: the CDAO is retaining several closed frontier models, classified networks are adopting a multivendor supplier base, and Palantir is linking models to data and command workflows. A Chinese tender for local DeepSeek, MaaS, and aviation agents confirms only procurement intent, while the French Mistral framework agreement provides a channel for access to technology; neither record by itself proves combat use. Access to a model therefore does not by itself create a decisive advantage.

Outcomes will be determined by:

  • sensors and secure communications;
  • the quality and currency of data;
  • integration into command structures;
  • mass production of low-cost platforms;
  • resilience to electronic warfare;
  • the speed of updates based on battlefield experience;
  • the ability to replace losses;
  • energy and compute in a distributed environment.

This will allow less technologically advanced states to remain an asymmetric threat if they integrate available models into cheap, mass-deployed systems more rapidly. Leadership at the frontier does not guarantee operational victory; industrial and organizational integration may matter more than the quality of any individual model.

Competition centers not only on model weights but also on the integration layer. A state that can replace Claude with GPT, Gemini, or Grok without restructuring its data, permissions, and applications reduces its dependence on any one lab. But if migration requires rewriting the ontology and mission workflows, strategic lock-in simply shifts to the integration platform. Palantir illustrates the US variant of a centralized control plane; Chinese MaaS platforms illustrate a federation of regional and sectoral layers; and Europe’s GPT@EC, Albert, F13, and LLMoin represent several incompatible national and institutional solutions. Russia’s architecture is seeking similar manageability through state demand and national models, but rests on a substantially narrower hardware base.

XII

Open-Weight Models Will Become Instruments of External Influence

In brief Open weights have become an instrument of influence. Whether they weaken the hardware oligopoly depends on which grows faster: model efficiency or compute requirements.

Open weights enable local deployment, reduce dependence on a provider's cloud, and accelerate the spread of an ecosystem. development is therefore not merely an engineering practice, but also a geopolitical instrument.

This approach is being pursued by:

If model performance grows faster than compute requirements, open weights will weaken the hardware oligopoly. If frontier-level performance continues to require rapidly increasing resources, they will broaden adoption but will not eliminate dependence on the owners of chips and compute.

XIII

International Institutions Will Become Leaner but More Necessary

In brief The universal order is weakening, but institutions remain needed as providers of narrow technical verification.

The universal rules-based order will weaken. It will be replaced by a combination of:

  • regional agreements;
  • minilateral coalitions;
  • licenses and exemptions;
  • technical standards;
  • inspections;
  • verification mechanisms.

The OECD, UNESCO, the G7, the Bletchley Process, the network of AI safety institutes, and the Council of Europe Convention are creating a common vocabulary, but not a unified enforcement regime. The refusal of the United States and the United Kingdom to sign the Paris AI Summit declaration demonstrates the limits even of Western consensus.

The same pattern is evident in the military and nuclear domains. What matters is not a promise of inspections, but actual access; not a declaration of neutrality, but verifiable force constraints; not an announced budget, but an operational production line.

By 2035, international institutions will be less capable of shaping an overarching order, but more indispensable as providers of narrowly focused technical verification.

XIV

Sanctions and Human Rights Will Become Part of a Transactional Order

In brief Human rights stay in diplomacy, but they work at once as a norm, a condition for easing sanctions and a bargaining chip.

The Belarus evidence base reveals a fragmentation of Western policy. The United States can ease selected restrictions and maintain a channel for securing prisoner releases while the EU simultaneously tightens sanctions in another sector. The release of political prisoners may coincide with new detentions, forced expulsions, and the preservation of the repressive apparatus. Administrative normalization in one EU country does not amount to a common European position.

By 2035, human rights will not disappear from diplomacy, but will increasingly function simultaneously as:

  • a normative demand;
  • a condition for sanctions relief;
  • a bargaining chip;
  • an instrument of domestic politics in receiving states;
  • a source of legitimacy for particular coalitions.

The universal vocabulary will endure, while its application becomes more selective.

XV

Arms Control Will Return after Rearmament

In brief After rearmament, arms control without trust becomes likely: force ceilings, inspections and regulated border theaters.

If Europe’s military mobilization materializes, Russia will face an unfavorable protracted arms race. This will create demand for a new system of mutual constraints:

  • force ceilings;
  • limits on missile and nuclear systems;
  • rules governing permanent foreign basing;
  • notifications and inspections;
  • constraints on rapid reinforcement;
  • specially regulated border theaters.

This will not be a return to the peaceful 1990s. The parties will first build new capabilities and then try to constrain the most dangerous among them. One possible formula is arms control without trust.

The theater spanning Belarus, Poland, and the Baltic states could become one of the principal testing grounds for such a regime. A viable agreement would require reciprocity across Russian and NATO force postures, Belarus’s separate consent, and verifiable consequences for violations.

XVI

Baseline Global Configuration by 2035

In brief The resulting line-up: two technology cores, the EU as a potential third, a military Russia with a compute ceiling, and hinge states between the cores.

United States

The principal owner of a full-spectrum military, financial, and AI stack; a provider of technology and guarantees on a conditional basis.

China

The only systemic competitor to the United States; the source of alternative production, financing, models, and infrastructure for much of Eurasia and the Global South. Its state AI architecture is not a single platform, but a coordinated federation of the CAC, NDRC/NDA, SASAC, state-owned telecommunications companies, funds, and municipal MaaS/control planes.

European Union

A potential third technology core and military power center. Its architecture combines supranational rules and funds with EuroHPC, a multivendor , and national workflows. Its status depends on converting site selections, grants, and framework awards into deployed compute capacity, broad-based adoption, and portable government control planes.

Russia

An independent military and nuclear power center with a protected state-corporate AI market centered on Sber and Yandex. It is likely to become a major regional AI power in the upper second tier globally, but with a compute ceiling and growing hardware and economic asymmetry in its relationship with China.

Hinge States

India, the Gulf states, Brazil, Indonesia, selected Central Asian states, and transit jurisdictions are not a passive periphery: they maneuver between the technology cores. They supply markets, energy, food, minerals, capital, production sites, and routes. Their agency is determined by their ability to retain multiple buyers and stacks, authorize access independently, and convert external competition into domestic capabilities.

A hinge role is not guaranteed by size or resource endowments. A state can rise within this hierarchy by diversifying routes and partners, or fall out of it if a critical asset comes under external control or a single patron becomes indispensable.

XVII

Three Scenarios

In brief Three paths: armed interdependence, verifiable coexistence, or theaters converging into a single crisis.

Baseline: Armed Interdependence

American and Chinese stacks emerge alongside a partially autonomous EU and a network of hinge states. Direct trade between the cores contracts selectively, but materials, components, and technologies continue to move through third countries. Regional wars and sanctions conflicts persist, but the largest power centers avoid direct, all-out war. Arms control and technology controls develop in a fragmented manner.

Negotiated: Verifiable Coexistence

The cost of the military and compute race compels the major power centers to establish mutual constraints: inspections, rules for military AI, controls on missile and nuclear systems, access regimes for compute and critical materials, and secure crisis-communication channels. Hinge states gain recognized rules for nonexclusive access and reduce the risk of being forced to choose sides. The world remains armed, but becomes more predictable.

Crisis: Convergence of Theaters

Conflicts in Europe, the Middle East, and the Indo-Pacific become interconnected through energy, chips, maritime routes, and sanctions. The cores forcibly accelerate autarky, ration civilian demand, and require hinge states to make an exclusive choice. A disruption at one narrow chokepoint triggers cascading effects across food, industry, and compute. Institutions cannot verify agreements quickly enough, while temporary ceasefires merely allow the parties to regroup.

XVIII

Indicators That Could Change the Forecast

In brief 23 signals that should prompt a revision of the forecast, from the utilization of European AI Factories to model portability in China.

EU

  • 1Actual delivery, acceptance, and utilization of European AI Factories; contract awards and commissioning of AI Gigafactories following the 2026 call.
  • 8The institutionalization of European rearmament beyond a single political cycle.
  • 22The conversion of Europe’s Cloud III and sovereign-cloud awards into specific orders and sensitive workloads; and portability among European providers without reverting to lock-in.
  • 23The spread of GPT@EC, Albert, F13, LLMoin, and the Luxembourg deployment beyond pilots and individual jurisdictions; and the emergence of a common European application layer or, alternatively, the entrenchment of national fragmentation.

China

  • 2China’s ability to replace restricted accelerators, , and semiconductor-fabrication equipment at scale.
  • 14China’s ability to reproduce advanced semiconductor-fabrication equipment and its servicing without Western suppliers, rather than merely substituting individual chips.
  • 20The share of Chinese municipal and sectoral AI platforms able to switch among DeepSeek, Qwen, ERNIE, and GLM without rebuilding data systems and workflows; and actual utilization of the national compute-monitoring layer.
  • 21A shift in Chinese SASAC use cases and municipal awards from reports of integration to measured production effects; and the emergence—or absence—of a unified cross-regional control plane.

US

  • 3The transformation of US full-stack AI exports into a durable allied system.
  • 16The DoD’s ability to switch frontier models in practice without rewriting ontology, permissions, and mission applications; the share of operational workflows dependent on Maven/AIP; and the ratio of actual obligations to contract ceilings.

Russia

  • 4A material reorientation of Russian West Siberian gas exports toward China, and the terms governing it.
  • 17The emergence in Russia of independently verified compute capacity for training frontier models: a mass-produced accelerator, sustained legal imports, or a shared compute mechanism with China or BRICS.
  • 18Implementation of Russia’s data-center plan in terms of GPU capacity and developer access, rather than only megawatts, floor space, and announced timelines.
  • 19The emergence of a third Russian foundation-model developer, or further concentration of data, status, and state demand around Sber and Yandex.

India & Gulf

  • 5Actual utilization of India’s shared compute infrastructure and the competitiveness of its national models.
  • 6Commissioning of the announced Gulf AI capacity.

Chokepoints & rules

  • 7The emergence of binding international rules for frontier and military AI.
  • 12A major new bottleneck shock involving the Strait of Hormuz, semiconductor-fabrication equipment, HBM, rare-earth materials, power grids, or subsea cables.
  • 13The commissioning outside China of sufficient capacity to process critical minerals and produce magnets, anode materials, and other critical intermediate goods—not only for the defense-industrial base but also for the broader economy.

Belarus–Poland–Baltics

  • 9A concrete reciprocal proposal for theater-level restraints encompassing Belarus, Poland, and the Baltic states.
  • 10Recognition of Belarus by both Russia and the Western coalition as a separate party to any future agreement.
  • 11A change in the Russian or joint military presence in Belarus.

Hinge states

  • 15A shift by major hinge states toward exclusive, long-term commitments to a single stack—or, conversely, the institutionalization of their right to retain mixed supply chains.
XIX

Conclusion

In brief The new world will be multipolar in its flags, dual-core in its foundational technologies, and hinge-based in its resources, markets and routes.

By 2035, the world will be defined not by a single victorious power center, but by competition over the configuration of interdependence.

Its principal features will be:

  • political multipolarity;
  • technological oligopoly;
  • state-corporate AI architectures in which models are easier to replace than data and operational workflows;
  • partial mobilization autarky without equal economic efficiency;
  • hinge states acting as gatekeepers to resources, markets, and routes;
  • state control of critical infrastructure;
  • persistent militarization without necessarily escalating to total war;
  • modular sovereignty;
  • transactional coalitions;
  • selective application of universal norms;
  • the return of verification and arms control;
  • the convergence of energy, compute, and data into a single source of power.

The new world will be multipolar in its flags, dual-core in its foundational technologies, and hinge-based in its resources, markets, and routes.

The central question of the decade is whether the technology cores can turn mutual vulnerability into verifiable constraints—and whether hinge states can preserve the right to make non-exclusive choices—before regional wars, technological blockades, and infrastructure shocks converge into a single global crisis.

XX

Limitations of the Evidence Base

In brief The forecast rests on an uneven evidence base: the AI race and security are covered better than climate, finance, Africa and Latin America.

  • The Focus Intermarium event base provides its most comprehensive coverage of Belarus, Russian–European security, infrastructure warfare, the Iran crisis, and the state-led race for AI.
  • Climate, demographics, biotechnology, the global monetary system, Africa, and Latin America are not covered in sufficient depth to support equally detailed forecasts.
  • The Focus Intermarium event base does not cover flows of critical minerals, food, and energy as systematically as it covers the state-led race for AI; the list of potential hinge states is an analytical typology, not a definitive ranking.
  • The AI-race records used here primarily capture state and supranational actions, not the AI market in its entirety. For Russia, publicly available data provide particularly limited visibility into actual compute capacity, model-training expenditure, circumvention channels for accelerator procurement, and the scale of military adoption. For China and the EU, official-source evidence provides a strong picture of rules, grants, tenders, and selected workflows, but offers less insight into utilization intensity, measured effects, and classified state systems.
  • An announcement, budget allocation, contract, construction, commissioning, and operation represent distinct states. An announced amount or capacity is not treated as an operational resource.
  • This forecast is an analytical extrapolation; Focus Intermarium records substantiate the underlying actions and hypotheses, but do not guarantee the outcome described.

Strategic forecast as of September 22, 2026. Based on open official sources; the forecast is an analytical extrapolation, not a guarantee of the outcome described.