Introduction
The effective closure of the Strait of Hormuz this week has concentrated minds on energy security. Oil prices have surged past $80 a barrel. Natural gas markets are in turmoil after Qatar halted LNG production. Shipping lines have suspended bookings across the Gulf, rerouting vessels around the Cape of Good Hope and adding weeks to delivery schedules. Approximately 170 vessels are stranded, waiting for circumstances beyond their control to resolve.
But the implications extend well beyond hydrocarbons, the conflict is stress-testing global supply chains in ways that should prompt some serious reflection among technology leaders, particularly those responsible for AI operations. Enterprise AI infrastructure exhibits concentration risks that rival, and in some respects exceed, those now disrupting energy markets.
The Vulnerability of the “Supply Thread”
Consider the foundations upon which most corporate AI initiatives rest. Between 80 and 90 per cent of AI workloads run on chips from a single vendor. The vast majority of those chips are manufactured by one foundry in Taiwan. Three cloud providers control roughly two-thirds of the compute capacity enterprises rely upon for training and inference, and most organisations have built their machine learning pipelines around the tooling and architecture of whichever provider they selected first – often years ago, before AI became strategically critical. “This is not a supply chain, it is a supply thread.”
Geopolitics and the Rise of Sovereign AI
The current conflict illustrates how quickly geopolitical conditions can shift. Forty-eight hours before US and Israeli strikes began, the Strait of Hormuz was functioning normally. Today, insurers have effectively withdrawn coverage for transit, and the mere credible threat of interdiction has achieved a de facto closure without requiring an actual blockade. The same dynamic applies to AI infrastructure. Export controls on advanced semiconductors have already demonstrated how policy decisions in Washington can reshape hardware availability overnight, with the H20 chip restrictions catching thousands of organisations off guard. Meanwhile, nations from Canada to Saudi Arabia are investing tens of billions in “sovereign AI” capabilities precisely because they recognise that infrastructure controlled by others is infrastructure that can be priced up, restricted, or withdrawn.
The Hidden Culprit: Deployment Failure and Infrastructure Rigidity
The geopolitical dimension is only part of the exposure. There is a more immediate fragility built into how most enterprises have constructed their AI operations. Industry data suggests that approximately 90 per cent of AI projects fail to deliver return on investment. The usual explanation points to model quality or data issues, but the more common culprit is deployment failure – the inability to move a working prototype into production at acceptable cost and speed. This happens because most AI pipelines are tightly coupled to specific infrastructure. Models trained in one environment require substantial rework to deploy in another. Hardware changes necessitate code changes. Compliance requirements in one jurisdiction demand architectural modifications that break assumptions made elsewhere. When conditions are stable, this complexity is merely expensive, but when conditions shift (a cloud provider changes pricing, a hardware allocation is delayed, a regulatory requirement forces data localisation) the rigidity becomes operationally dangerous.
The Energy Tax on AI Workloads
The current energy price shock offers a preview of the pressure enterprises may soon face. AI workloads are extraordinarily energy-intensive. Data centre operators in Europe are already warning that sustained elevation in natural gas prices will force difficult decisions about capacity allocation. Cloud providers will pass higher energy costs through to customers, though the timing and magnitude remain uncertain. Organisations with workloads locked to a single provider or region have limited options when pricing changes. Those with the flexibility to migrate (to shift inference to more cost-effective hardware, to rebalance workloads across geographies, to move processing closer to where data resides) will be better positioned to absorb the shock.
The same logic applies to hardware availability. If the Taiwan Strait were to experience the kind of disruption now affecting the Persian Gulf, the impact on semiconductor supply would dwarf anything the energy markets are currently experiencing. Enterprises whose AI operations can run across multiple hardware architectures would have options. Those without such flexibility would face extended outages.
From Performance to Resilience: A New Mandate
For the past several years, the conversation around AI infrastructure has focused primarily on performance and cost and these remain important considerations, but the events of the past week suggest that resilience deserves equal weight. Resilience in this context means the capability to, Build, Run, Operate AI workloads across diverse environments without fundamental re-architecture.
It means hardware flexibility – the ability to run inference on whatever silicon is available and cost-effective, rather than being locked to a single vendor’s roadmap and pricing. It means systems that function without mandatory connectivity to external services that may become unavailable or unacceptable.
This is not an argument for abandoning current infrastructure investments. Most organisations have made substantial commitments to particular cloud providers and will continue to derive value from those relationships. The argument, rather, is for ensuring that those commitments remain choices rather than constraints.
Conclusion: Building Optionality into Architecture
The Iran conflict will eventually be resolved, oil prices will stabilise, and shipping routes will reopen, but the underlying lesson will remain: concentration risk is real, and the consequences of ignoring it can materialise with remarkable speed. According to recent reports, even OpenAI has been seeking alternatives to its primary chip supplier, and finding that its considerable leverage cannot guarantee optionality when a dominant vendor moves to acquire potential alternatives. If the organisation at the centre of the current AI moment cannot secure infrastructure independence through market power alone, most enterprises have little chance of doing so.
The alternative is architectural. Organisations that build abstraction into their AI operations – that treat hardware, cloud, and deployment environment as configuration rather than architecture – these are the organizations that will have options when conditions change. Those that remain tightly coupled to today’s assumptions will find themselves, like the vessels now stranded outside the Strait of Hormuz, waiting for circumstances beyond their control to resolve. The disruption in the Middle East is a reminder that such circumstances have a way of arriving without warning. The time to build resilience is before it is needed.
