The Intelligence Chokepoint: Why AI Concentration Is the Geopolitical Risk Nobody Is Pricing

On the morning of March 3, 2026, a senior commander of Iran's Islamic Revolutionary Guard Corps announced that the Strait of Hormuz had been closed to commercial traffic.¹ The declaration was brief. The market response was not.

Within hours, Brent crude rose nearly 10 per cent. Maritime insurance underwriters issued cancellation notices for war risk coverage, effective within 48 hours. Iraq and Kuwait began curtailing oil production — not for lack of supply, but because storage was filling and export routes had vanished. Dubai International Airport, which processed 95 million passengers in 2025, suspended operations.

The mechanism behind this cascade is straightforward. The Strait of Hormuz — a 21-mile passage between Iran and Oman — carries approximately 20 million barrels of crude oil and petroleum products per day, representing around 20 per cent of global petroleum consumption and 25 per cent of all seaborne oil trade, according to the International Energy Agency.² Five nations — Iran, Iraq, Kuwait, Qatar, and Bahrain — have no meaningful alternative export route. When the strait closes, so does their economic lifeline.

The relevance of this episode extends well beyond energy markets. It is, in structural terms, a precise analogue for a dependency now being constructed at far greater speed in a different domain entirely: artificial intelligence.

A market more concentrated than oil

In 2023, the world's five largest large language model providers collectively accounted for approximately 88 per cent of global LLM market revenue, according to market research published by Hostinger and corroborated by Springs Apps.³ That figure likely understates the true concentration of capability, since market revenue does not capture the degree to which frontier model performance is even more narrowly held.

The providers in question are OpenAI, Google DeepMind, Anthropic, Meta, and a cluster of Chinese state-adjacent laboratories including Baidu, Alibaba, and DeepSeek's parent entity. Two jurisdictions. A handful of corporate entities. And a global market for AI services projected by Precedence Research to grow from $638 billion in 2024 to $3.68 trillion by 2034.⁴

The parallel with Hormuz understates the AI concentration problem in one critical respect. Oil has strategic reserves. It has alternative producers. It has futures markets that allow importers to hedge. The AI dependency being constructed has none of these circuit-breakers. There is no International AI Agency maintaining emergency capacity. When access is removed, it is removed immediately and completely.

What makes this concentration more precarious still is that the software layer rests on a physical infrastructure that is itself alarmingly concentrated. Approximately 90 per cent of the world's advanced semiconductors are manufactured by a single company — TSMC — on a single island that sits at the centre of the most consequential geopolitical dispute of our era. Nvidia's H100 and B200 chips remain so scarce that access to them has become an instrument of US foreign policy, with export controls used to deny compute capacity to adversaries and allies alike.

The access problem is no longer theoretical

Iran, a nation of 89 million people with a substantial engineering and scientific workforce, cannot access ChatGPT, Claude, or Google Gemini.⁵ The restriction is dual-layered: domestic government policy prohibits use of Western internet services, while OpenAI, Google, and Anthropic have independently blocked traffic from Iranian IP addresses.

Russia's situation is more nuanced and instructive. OpenAI began actively removing access for Russian users in mid-2024. As of March 2026, the Russian government is advancing legislation that would require any foreign AI service with more than 500,000 daily users to store Russian data on Russian territory — a condition that major Western providers are unlikely to meet.

China observed this dynamic early and acted accordingly. Its domestic AI ecosystem was not constructed primarily for commercial reasons. It was constructed because Chinese policymakers understood that dependency on foreign intelligence infrastructure was a strategic vulnerability of the first order.

The corporate dependency: a structural trap

The business model of firms such as Infosys, Accenture, Wipro, TCS, and Capgemini was constructed on a durable proposition: that enterprises required human intermediaries to implement, integrate, and maintain technology systems. Infosys today employs approximately 337,000 people globally.⁶ Its billing model rests fundamentally on the monetisation of human hours.

Generative AI attacks that model at its foundation. The McKinsey Global Institute estimated in its July 2023 report⁷ that activities accounting for up to 30 per cent of hours currently worked across the US economy could be automated by 2030, a proportion accelerated materially by generative AI. IT services and business process outsourcing are among the sectors most directly exposed.

The strategic trap runs deeper than automation risk alone. OpenAI has established partnerships with Infosys, TCS, Cognizant, Accenture, and Capgemini to distribute its Codex coding tools through enterprise channels.⁸ Presented as partnerships, these arrangements are more accurately characterised as a distribution strategy: OpenAI accesses enterprise client relationships these firms spent decades cultivating, while retaining ownership of the intelligence layer. The services firms function as sales channels rather than technology providers.

On the question of building sovereign AI

The response of both governments and corporations to this dependency has been, almost uniformly, to announce investment in AI capability. The announcements are real. Their strategic adequacy is more questionable.

Infosys has committed $2 billion over five years to AI.⁹ Accenture has pledged $3 billion.¹⁰ France has invested in national GPU infrastructure.¹¹ Singapore has launched SEA-LION. Malaysia has ILMUchat. Switzerland, Apertus.¹²

France's national GPU deployment represents less than 6 per cent of the compute infrastructure Microsoft alone has committed to installing in French data centres.¹³ The aspiration toward sovereignty runs well ahead of the capacity to realise it.

This does not mean the only alternative is passive dependency. A more productive framing distinguishes between frontier model development — which requires capital and talent concentrations available only to a handful of actors globally — and the deployment and customisation layer, which is significantly more accessible. Open-weight models — Meta's Llama, Mistral's open releases, Alibaba's Qwen — make substantial AI capability available without frontier-scale investment. The meaningful question for most governments is not whether they can build a GPT-5 competitor. It is whether they can run capable models on sovereign infrastructure, fine-tuned on proprietary data, governed by domestic policy.

The structural risk remains unpriced

The Hormuz blockade of March 2026 lasted, in its most acute phase, approximately 72 hours. The global economic cost — in oil price volatility, insurance disruption, logistics paralysis, and production curtailment — ran to tens of billions of dollars. The cost of decades of underinvestment in bypass infrastructure is now visible against those consequences.

The AI dependency now being constructed is accumulating at a comparable rate, with less visibility and no equivalent of an energy security framework to monitor or mitigate it. Companies are integrating frontier AI into core workflows — legal, financial, operational, strategic — without systematic assessment of what happens if access is revoked, pricing is restructured, or model behaviour changes in ways that conflict with their requirements.

Governments are adopting AI systems for public services, defence planning, and economic management while outsourcing the underlying infrastructure to foreign commercial entities subject to foreign law and foreign policy pressure.

The intelligence chokepoint is being built now. The question for policymakers, boardrooms, and investors is whether they are treating it with the seriousness that a dependency of this magnitude warrants. The evidence, at present, suggests they are not.