I started my career as a climate scientist, working on global climate models for my Doctorate — large Fortran systems running on NASA, NOAA, and university supercomputers. We scenario-tested the planet for a living.

Back then, in the early 2000s, AI sat at the edges: we used simple statistical machine learning to help analyse output, not to model the system itself.

Over the past two decades, the compute available to model the Earth has grown by orders of magnitude. The planet has been wired with satellites and sensors at a scale no earlier generation could imagine. And AI has made the sharpest leap of all. In two decades, it has gone from a collection of specialised tools to a pervasive general-purpose technology that learns its own representations from data rather than following rules we write: the substrate the next era of science and industry will run on.

No longer a tool for analysing the system, it is increasingly becoming a means of modelling it.

What I argued in 2018 — and what I missed

In 2018, as a partner at PwC, I co-authored two early reports with my team on how AI could be applied to the Earth system: one with the World Economic Forum, one with Lucas Joppa, then Microsoft’s chief environmental officer. Both came before the current wave of foundation models and the skilful neural simulators that followed.

We were optimistic. The argument was simple. Alongside steadily rising computing power, the planet was being instrumented: commercial constellations imaging the entire land surface every day at a few metres’ resolution, open satellites at ten, and expanding observation of the oceans, ice, and forests. AI, we argued, would help us turn that data into better predictions and early warnings, better connected and optimised systems, and better decisions.

A lot of that has happened. The progress has been staggering. Machine learning is now routine in satellite image analysis, in crop and commodity monitoring, in wildfire and methane-leak detection, in grid optimisation, and in tropical-storm forecasting and earthquake early warning.

In Earth observation, scientists grapple with incompatible archives: different satellites, different resolutions, snapshots interrupted by cloud and orbit. DeepMind’s AlphaEarth Foundations learns a single representation across them, ten-metre cells, annual, globally consistent. It shows how the land surface has changed, though not what governs the change.

Weather forecasting is unrecognisably better, and climate models have advanced substantially. Google’s flood forecasting service now runs in more than eighty countries. GraphCast, Aurora, NeuralGCM — neural emulators from DeepMind, Microsoft, and Google — have matched or beaten the world’s best physics-based weather models on standard medium-range scores, at a fraction of the compute cost. NVIDIA’s Earth-2 now bundles a competing family of them into an open, end-to-end stack that, on the company’s own benchmarks, edges ahead.

Whether these learned models handle extremes and tail risk as well is a more open question — and for high-impact events, the one that matters most. The tail is where the stakes are highest: planners, insurers, and farmers all plan for the extreme, not the average. A model strong on the median but weak in the tail is not a safer tool but a more dangerous one: it invites a confidence its tail behaviour has not earned. Closing that gap is a live frontier. These are real gains, concentrated in the parts of the Earth system that are well-observed, well-theorised, and short-horizon.

In our 2018 predictions, we were also reaching toward something deeper — that AI might discover patterns in complex physical systems humans had missed. The instinct was directionally right, but it assumed the system itself was already defined.

For the hardest Earth-system problems, it isn’t. The ones people most need answered are exactly the ones that haven’t moved: what a hurricane will do when it arrives; when the next drought will develop, and how severe; food security under a changing monsoon; subseasonal forecasts, the two-to-eight-week window so important to farmers; the stability of ocean circulation under freshwater forcing as ice sheets melt.

The same failure shows up inside the models. The envelope on 21st-century sea-level rise is nearly as wide today as it was fifteen years ago. The terrestrial carbon sink — the part of nature that absorbs about a third of what we emit — remains the least resolved term in the carbon budget. Tropical convection, the process that governs rainfall for billions of people, is still represented in climate models by approximations every working scientist knows are unreliable.

These have barely moved. More data hasn’t solved it. Bigger models haven’t solved it. Better physics at higher resolution hasn’t solved it.

We thought the bottleneck was compute. It wasn’t.

It was mostly representation. By which I mean: not just the variables we chose to track, but how we carve the system up.

That gap has sat in the middle of Earth-system science for a long time.

The turn: World models

It’s easy to see the recent wave of AI-in-science results as incremental improvements on a familiar trajectory.

They’re not.

They’re the low-hanging fruit of a much bigger shift. A quieter class of AI system is evolving fast, moving beyond language to do something fundamentally different: learn how the world behaves.

These are called world models, though the term is used somewhat differently across research groups to describe a family of related architectures rather than a single agreed design. Here I use it in its broad functional sense.

The researchers most associated with the shift are pursuing it along different lines. Demis Hassabis and DeepMind are advancing systems that simulate and plan over complex environments. Fei-Fei Li’s World Labs is building toward spatial intelligence, with a less settled architectural commitment. Yann LeCun left Meta at the end of 2025, after twelve years as its Chief AI Scientist, and in early 2026 launched AMI Labs, built around his JEPA architecture: a new approach to learning predictive representations of the physical world from observation rather than language. The direction is shared: from modelling text to modelling the world.

A large language model (LLM) predicts what comes next in text. It can reason over language, code, and increasingly images. It is extremely powerful at working with information — but it does not maintain a grounded state of a physical system. It models descriptions of the world, not the world itself.

Reinforcement learning is about learning a policy: what action to take inside a defined environment with a clear goal. Some forms also learn an internal model of that environment, but the focus remains on behaviour rather than the system itself.

A world model is different. It learns directly from data how a system behaves over time, building an internal representation: a compressed picture of the system, organised not around the variables humans chose but around whatever best explains how the system evolves. From that representation it can simulate forward and explore how the system responds to change. It is a model of dynamics.

Language models learn representations of what humans say about the world. World models learn how the world changes.

That is the distinction that matters, and it goes beyond architecture. Traditional simulation fixes the state variables in advance—temperature, pressure, moisture—and writes down the equations that govern them. A world model does something categorically different: it learns its own representation of the system and its dynamics directly from data. The structure of the system is discovered, not specified. Data assimilation has long inferred hidden variables inside prescribed models; learning what the representation itself should be is new. That is a different mode of scientific inference.

And, critically for what follows, the most useful versions of these models do not try to predict the world in full sensory detail. They predict it in a compressed, abstract space, learning what matters for the system’s evolution and discarding what does not. For a planet’s worth of data, that distinction is not optional. It is the difference between a model that can scale and a model that drowns in pixels.

Two further properties are worth naming, because they sharpen what the term means. First, a world model is something you can ask questions of: not just “what happens next” but “what happens next if we do X.” It learns to predict the system’s behaviour conditional on actions or interventions — emissions trajectories, land-use change, a policy choice. That is what separates it from a passive forecaster, and it is what makes it useful for decisions, not only for analysis.

Second, for a system as large as the Earth, a world model has to work at multiple time scales at once. Convection runs in minutes; weather in days; ocean overturning in decades; ice sheets in centuries. A model that captures only one of those is the wrong object. The architecture has to be hierarchical — coarser representations of slow dynamics, finer ones of fast, coupled in both directions — or the planet does not fit inside it.

And finally, deeper than the other two: for a chaotic system like the Earth, a useful model cannot produce a single trajectory. It has to describe a range of possible futures; uncertainty built in. That is an architectural demand: the dynamics themselves must be probabilistic, so the spread of outcomes emerges from the model rather than being estimated around it. Ensemble forecasting was invented for this reason, and the physics-based systems meet it at supercomputer cost. What is new is that a learned model now meets it too.

GenCast matters here for one reason: it shows a learned system can meet the probabilistic demand, generating a full ensemble by sampling a distribution over weather states, and on DeepMind’s evaluation it outperformed ECMWF’s ensemble on roughly 97 percent of targets, in minutes. But GenCast is not itself a world model in the sense this piece means. It predicts in weather-map space over the standard variables, not in a learned, compressed representation of its own — an emulator with probabilistic output, not learned latent dynamics. It proves the principle that a learned model can be probabilistic; it does not yet prove the architecture.

That is the deeper step change world models point to: not faster simulations of systems we already know how to specify, but learning latent state and dynamics directly from data, and enabling continuous estimation, prediction, and intervention in systems that refuse to hold still.

A note on the examples in this piece, GenCast among them. These share the world model’s DNA: they learn how a physical system behaves, not label images or spot patterns. Emulators reproducing known physics, a few controllers of bounded systems like a fusion chamber: cousins and early ancestors, not the thing itself. What has not yet been realised for the Earth—the focus of this piece—is a single integrated world model that learns a unified representation of the ocean, atmosphere and cryosphere as one evolving system, rather than modelling them as distinct components that exchange fluxes across prescribed interfaces.

Why now

Why is this conversation happening in 2026 and not a decade ago? Four things have changed, and they have changed together.

The first is architectural. The class of model this piece is about has been theorised for years. What is new is that we can now train predictive models in latent space at scale, learning the structure of dynamics directly, without reconstructing observations. That is the technical breakthrough.

The lineage that delivered it is real and converging: the JEPA family, MuZero and the model-based reinforcement-learning family, latent-dynamics architectures across robotics and physics.

Capital is following: the labs that built the LLM era are reallocating toward systems that simulate, plan, and reason about the physical world. The scale of that reallocation is visible in the capital: Jeff Bezos launched Prometheus, a physical-AI venture, from stealth in 2026 at a reported $41 billion valuation. It’s a bet, in effect, that models of the physical world are the platform after language.

The second is data. The Earth is now instrumented at planetary scale. ERA5 provides 80 years of hourly global atmospheric reanalysis; Argo gives near-continuous reporting from the ocean interior; Sentinel-2 covers the Earth’s surface weekly. The volume and density required to learn joint dynamics across atmosphere, ocean, land, and ice did not exist a decade ago. They do now.

The third is proof of tractability. The neural weather models already discussed — GraphCast, Aurora, NeuralGCM, Prithvi-EO and their kin — settle a narrower but essential question: neural systems can ingest planetary-scale data without breaking. The question has shifted from whether to where.

The fourth — and least obvious — is that for this class of problem, the binding constraint has moved. Compute is no longer the only gate. Weather-ML has proved that learned Earth-system emulators are tractable. The harder bottleneck is increasingly data access: observational records scattered across agencies, formats and access regimes, held by institutions never built to share them for system-level inference. That constraint is institutional, not technical.

It is no longer whether we can simulate the systems we understand. It is whether we can represent the systems we don’t.

That is why this is happening now.

It’s a representation problem, not a prediction problem

The hardest Earth-system problems are not just prediction problems. They are representation problems. They fall in the gap of too poorly understood to write down in full, too sparsely observed to learn from data alone.

Most of our models are organised around variables scientists chose, like temperature, pressure, humidity, wind. The equations are written for those variables. The data is collected for those variables. The models predict those variables forward. When the prediction fails, we assume we need finer resolution, more data, or a slightly better equation.

But what if the variables that matter for the system’s future aren’t the ones we chose to measure? What if what governs the next year of monsoon rainfall is a coupling between ocean and atmosphere that the models still represent in pieces? What if the couplings between soil moisture, vegetation, cloud formation, and heatwave intensity don’t live cleanly inside any one subsystem — but in a joint trajectory that no current model represents at all?

The representation gap isn’t just that we chose the wrong variables inside each subsystem. It’s that we chose to model subsystems in the first place.

Higher resolution, more data, or machine learning bolted onto existing physics cannot close a representation that is incomplete to begin with.

A world model, in principle, offers something new: the possibility of learning the system’s state directly from data and, where the observational record is rich enough, letting the dynamics emerge from it.

Why science needs world models

The case for world models in science is not that they are smarter than what came before. It is that they learn how systems behave. That is what makes them uniquely valuable for scientific discovery.

First, they can infer states we cannot observe. Most scientific systems are chronically underobserved. Satellites map the ocean surface, but the kilometres of water beneath are sampled only sparsely. We see the top few centimetres of soil, not the root-zone moisture that governs drought. We see the Sun’s surface, not the interior magnetic field that drives space weather.

A world model, trained on how observable variables evolve together, can infer what we cannot measure directly. An LLM cannot: it has no grounded representation of the physical world. A numerical simulator can infer only the hidden variables defined within a representation we specified in advance. A world model can instead learn that representation itself from data.

Second, they can learn dynamics we do not fully understand. Large parts of the Earth system are governed by processes whose equations are either unknown or known to be provisional. Cloud microphysics. Turbulent mixing. Ice-cliff fracture. The feedback between soil moisture and convection. For these, simulation depends on parameterisation (a fitted guess at the thing we do not know). World models learn dynamics directly from data, without requiring us to write the equations first.

Third, they can represent coupled systems as a single object. Science has made huge progress by decomposition: separating atmosphere from ocean, fast processes from slow, physics from biology. But many of the most important phenomena live in the couplings between subsystems, which today’s models miss. A world model, trained on the full joint record, can learn these. That is not an incremental improvement on existing tools. It is a different kind of model of the system.

We also need to be honest about where confidence degrades. Decisions depend on asking what happens under conditions the Earth has not yet entered: higher warming, a weakened Atlantic overturning, a regime change in Antarctic ice. Neither pure approach handles these well. Physics-based simulators rely on parameterisations tuned to the present climate, which may or may not hold in a new regime. World models also extrapolate beyond the data they have seen. The promise is not certainty but honesty: a model that flags where its confidence should fall rather than failing silently. Whether they can do that reliably is still open. It is the difference between a scientific instrument and a confident guess.

A natural architecture for these regime-change questions is a hybrid built around a learned core: hard physics enforced as constraints within the model, combined with high-resolution simulation and observations informing the learned dynamics.

The deliverable is not prediction. It is a tighter envelope of plausible futures, produced by an architecture that combines what each approach can guarantee with where each must admit uncertainty. This is a narrower claim than the one usually made about AI and climate, and it is the one the evidence supports.

No existing tool, whether LLM, reinforcement learning, or numerical simulation, does all of these together. World models applied to the Earth is not a marginal improvement; it is a different way of doing science, and if it works, it changes not just what we can answer but what we can ask.

The shape shows up clearly in specific cases.

Take subseasonal forecasting — the two-weeks-to-two-months window. Weather skill degrades at two weeks because the atmosphere is chaotic. Seasonal skill returns only slowly, from the ocean, the land surface, and the stratosphere. What lives in between is a coupling no decomposed model holds as a single object. A world model trained on the joint record can.

Take land–atmosphere coupling. Soil moisture amplifies heatwaves; transpiration sustains rainfall recycling; root-zone water triggers convection. The variable that actually controls these feedbacks sits half a metre to two metres below the surface — almost entirely unobserved at global scale. A world model can, in principle, infer it from its imprint on the joint behaviour of everything we observe.

Take ice sheets and sea level. The assessed likely range for 21st-century sea-level rise is roughly 0.3 to 1 metre, with a low-likelihood, high-impact tail reaching toward 2 metres that cannot be ruled out — wide enough to make coastal adaptation ambiguous for almost every major coastal city. The high end of that range is dominated by the deep uncertainty in how Antarctic ice cracks and melts, not by which emissions pathway the world follows. It will not be narrowed by running a higher-resolution version of an ice-sheet equation that is itself provisional. It will be narrowed, not resolved, by a representation consistent with the full observational record at once.

The ocean interior — where more than 90% of the excess heat of the warming climate now sits, and where circulation is inferred rather than measured — is another frontier with the same structure.

Different systems, same shape: hidden state, partial dynamics, signal in the couplings. The argument does not rest on any single one of them. It rests on what they share — and each is treated at greater length in the annex.

The problems no one owns

The hardest problems are not inside any one system. They sit between them.

Subseasonal forecasting sits between meteorology, oceanography and stratospheric dynamics. Land–atmosphere coupling sits between hydrology, vegetation science and atmospheric physics. Heatwaves are not atmospheric events alone but land–atmosphere feedback loops that amplify themselves. Ice sheets sit between glaciology, ocean physics and the solid Earth. Extend the list and the pattern repeats: grid resilience under compound extremes, space weather from solar wind through the magnetosphere to satellite drag.

The highest-value Earth-system problems increasingly live at these interfaces. This is not a coincidence. It is the structure of what is left: the problems that fit cleanly inside a single subsystem have already yielded many of their biggest advances, and what remains is what no single model, discipline or dataset owns.

That is also where existing tools begin to struggle.

The scientific apparatus we built for the twentieth century is organised by decomposition. Atmosphere, ocean, ice and biosphere are modelled largely as separate systems and then coupled. Coupled models are among the field’s great achievements, and for much of climate science they work remarkably well. But the coupling remains a negotiation at the seam, not a single joint representation. Where the most important behaviour lives in how those systems co-evolve, a handshake at the boundary cannot hold it.

The same fault line runs through the data. Much of what we call climate data is not raw observation but reanalysis: observations combined with physics-based models to produce a consistent picture of the Earth system. For the atmosphere, dense observations make it extraordinarily powerful, and it is what today’s best AI weather models learn from. But it is only as good as the observations beneath it. For the deep ocean, the metres below the soil, or the base of the great ice sheets, observations remain sparse, and the model is necessarily filling in much of the picture itself.

It is worth being precise about what it means to say the data are not ready, because the clearest success story in AI for science shows why that matters. AlphaFold did not emerge from raw sequencing output. It was built on the Protein Data Bank: half a century of experimentally determined structures, curated, standardised and made openly available as a single machine-readable corpus. The breakthrough was as much an achievement of that data infrastructure as of the model.

Earth-system science has no equivalent. It has extraordinary observational assets, but they remain scattered across reanalyses, satellite archives, ocean floats, climate model output and station records that were never designed to become a unified training corpus. Unlike protein structures, the Earth system is continuously evolving, its observations are constantly revised, and much of the record is itself model-assisted. Building such a corpus is therefore harder. But the missing ingredient is not permission to download files. It is the machine-ready scientific corpus itself.

That reframes the highest-leverage public investment. The instinct is to fund compute and models; the more decisive move is to build the Earth-system equivalent of the Protein Data Bank, and to decide who governs it. Whoever assembles and curates that corpus shapes which questions get asked, which models get trained, and which regions and risks are represented at all. In an era when the same data underpins climate resilience, food security, energy security and critical infrastructure, that is not technical housekeeping. It is strategic infrastructure — and no institution today owns the task of building it.

Where world models are now

Early versions are already reaching real hardware. In fusion research, a team from MIT, EPFL and Commonwealth Fusion Systems built a model that learns the plasma’s dynamics directly from a few hundred experiments, disciplined by known physics, and used it to design a safe shutdown. It builds on an earlier DeepMind and EPFL system that learned to control a plasma through reinforcement learning. This is the primitive the piece is about: dynamics learned from data, not written down by hand, running on real hardware.

That result matters, but it should not be over-read. A fusion reactor is a contained experimental chamber with dense diagnostics and a clearly defined set of control actions. The Earth system is none of those things. It is open, sparsely observed, non-stationary, and vastly higher-dimensional. Whether the architectural choices that worked for a bounded environment carry across to an open one remains a genuinely open research question.

The same shift is now visible across fields whose architectures share almost nothing else. AlphaFold predicts protein structure with attention-based networks. Materials-discovery models operate over graphs of crystal structures. Medium-range weather increasingly relies on transformers and diffusion. Different problems, different architectures, one underlying move: learning the structure of a system directly from data rather than specifying it in advance. Whether it scales to systems that refuse to hold still is the frontier the rest of this piece is about.

Almost all commercial attention on world models today points at enterprise: factories, warehouses, autonomous vehicles, game engines. That is not an accident. It is where the data is cleanest, the reward signal clearest, the payback fastest, and it is where every major AI wave has landed first. Chatbots reached customer service before science, computer vision retail before medicine, reinforcement learning games before anything real.

World models are following the same path. IBM has already declared world models the next frontier for enterprise.

Which raises the question that actually matters: not whether these systems work in principle, but what they get trained on — because what a world model learns from shapes what it learns to model. Trained on warehouse logistics and driving footage, it learns the physics of the bounded and the ownable. Trained on planetary observation, cell-biology time-series, grid dynamics, it learns the systems we live inside — the ones that are not bounded, not clean, and not optional.

The world model as organising layer

A learned, compact representation of joint dynamics is the right organising layer for the kinds of systems we cannot fully specify in advance, such as planetary science. Around that layer sits everything else the field has built: physics-based simulators for the components we understand, conservation laws as hard constraints on the learned dynamics, data assimilation across the full observational record. The world model does not replace these, but reorganises how they fit together. What changes is not the role of physics but the structure of the model: from separate subsystems to a single learned representation of their interaction.

This is also why the strongest architectures today are hybrid and will remain so for the foreseeable future. The world models are not yet good enough to stand alone, and the physics we do understand is too valuable to push to the margins. The point is not that hybrids disappear. Hybrids built around a world-model core, with physics inside the loop, are a different object from hybrids built around a physics-based core with machine learning added on. The first reorganises the stack: the second accelerates it. For the coupling problems in this piece, the promising path runs toward the first.

World models have their limits and the honest picture requires a distinction the hype usually skips. Some regimes are sparsely observed but physically continuous with what we have seen — a monsoon a degree and a half warmer, an under-sampled ocean interior. On these, a learned representation can extrapolate along the same underlying manifold, and this is most of the value. But the Earth system also has genuine regime shifts where the dynamics themselves change, like the AMOC collapse, or marine ice-cliff instability at scale. No volume of historical data contains them, because the behaviour on the far side of the bifurcation was never sampled. This is the deepest limit on the paradigm, and no larger model removes it.

What the approach can still do is more useful than confident prediction, which no method offers. It changes what can be bounded. The firm part is physical. Conservation laws hold whether or not the circulation has collapsed, so hard constraints can keep the model within physically admissible states even in regimes it has never observed. That bounds what is dynamically possible, not merely what has been observed, and lets the model identify where it is extrapolating beyond the evidence.

The softer possibility is early detection. In theory, systems approaching a bifurcation leave statistical signatures, critical slowing-down, rising variance, that a model might learn to recognise. But the evidence is contested. Recent claims that the Atlantic overturning is already exhibiting such signals have been challenged because the records are short, the proxies uncertain, and the bifurcation assumed rather than demonstrated.

So the deliverable for genuine tail risks is neither a narrower forecast nor, at least for now, a reliable alarm. It is a physically constrained envelope of what the system could do, bounded by conservation laws rather than the historical record: the decision-relevant quantity for an insurer pricing coastal exposure or a central bank stress-testing climate risk. You do not need the year the Atlantic overturning collapses. You need the boundary of what is physically possible, and an honest flag where the model’s sight ends.

Nor is this the story of a single monolithic model of the Earth. That framing is marketing. The architecture will be federated, and the “true world model” is a reference capability, not an engineering target. What exists today are fragments; what is missing is the system that connects them.

This is not a product. It is a way of modelling. And eventually, a way of deciding.

What is actually at stake

Much of today’s AI conversation is about agents, enterprise automation and productivity. Those are real shifts.

But if you work on the systems we actually live inside our — earth systems, our biology — there is a quieter shift happening, and it is starting to matter.

For the first time, we have a class of AI system that could allow us to represent Earth processes whose structure we don’t currently know how to write down. Not better predictions of the variables we chose. Different variables. Different dynamics. Different inferences.

A world model is two things at once: a scientific tool for systems we cannot specify in advance, and the capability AI needs to reason about, plan in, and act on the physical world.

The same capability that could narrow the sea-level envelope is what moves AI beyond the limits of language models — toward systems with grounded understanding.

Earth-system science is not just a beneficiary of this shift. It is one of the hardest test environments — partially observed, non-stationary, multi-scale, irreducibly probabilistic, and without the ability to reset. Progress here is a real signal of progress in the broader paradigm: the test is not just whether AI can predict, but whether it can know what it cannot predict.

The implications extend beyond climate. Cell biology, disease systems, immune dynamics, ecosystem responses — these are also partially observed, evolving, governed by interactions we cannot fully specify in advance. Researchers in those fields are beginning to reach for the same architectural answer, which is itself a signal: the convergence is not confined to Earth science. In each, the frontier is not just better prediction, but better representation. The choice of architecture is therefore also a choice about what science the next decade of AI is built to serve.

But Earth systems are the hardest test. They are open, global, and consequential. If this approach works here, it will be a serious signal for the wider paradigm. If this direction holds, it will look like a generational project to build a learned, observationally grounded representation of the planet — woven through physics, governance, and data access.

Who pays for the frontier

The money is not distributed evenly across this frontier, and where it falls decides what gets built. Enterprise applications have customers, contracts and quarterly results, and the capital following world models in 2026 is aimed overwhelmingly there: factories, autonomous systems and robotics. It is how frontier technology has always been built, each generation of billion-dollar training runs funded by the generation of revenue before it. The enterprise wave is not a rival to the science; it is what pays for the capability the science will inherit.

Where that inheritance has reached Earth science first is where a commercial market already exists: weather services and catastrophe risk modelling. And here the pattern turns interesting, because much of it is being shared rather than sold. NVIDIA’s Earth-2 releases its forecasting models as an open stack, built for governments and public institutions to run their own systems. It is not an isolated case. Across Earth AI, the leading neural forecasting models from DeepMind, Microsoft, NVIDIA and the IBM–NASA collaboration have largely been released openly, with GraphCast available for research, if not yet for commercial use, and these systems increasingly running alongside physics-based models at major forecasting centres.

NVIDIA frames Earth-2 as infrastructure for sovereign AI, so that any government—not only the handful of states able to build frontier models themselves—can run its own forecasting on its own compute. It is both public-good and commercial strategy, and the two are not in tension: open software expands the ecosystem, frontier compute remains the economic engine. The same logic extends beyond applications to the foundational science they depend on.

Step back, and the opportunity is extraordinary. Between open weights and the falling cost of access to frontier compute, a research group today can reach for capabilities that, a single generation ago, existed only inside a national laboratory or a frontier lab’s private cluster. As Dario Amodei has argued, frontier AI could soon amount to “a country of geniuses in a datacenter”: tens of millions of expert-level minds running in parallel. No generation of scientists has ever had access to this much capability this fast.

You cannot fine-tune your way to a representation that has not been built. Adapting GraphCast gives you a better forecast; it does not give you a unified model of the coupled Earth system. The problems that have barely shifted in a generation—the range of sea-level rise, the strength of the land carbon sink and skill at longer-range forecasting—will not move because someone tuned an open model. They require training a foundational model on Earth-system data itself: a unified learned representation of the coupled Earth, built on the kind of curated scientific corpus that transformed protein structure prediction. And that is precisely the kind of model commercial incentives struggle to produce.

The challenge is that many of the outputs are public goods: a firmer estimate of the carbon cycle, a narrower range of future sea-level rise, or an earlier warning of how the next drug-resistant pathogen will spread. There is no direct customer; the benefit is shared, while the payoff may lie decades away.

The market does not underfund these problems because they are unimportant. It underfunds them because importance is not the variable it responds to. Appropriability is. That is why the deepest problems, the ones whose uncertainty has barely moved in fifteen years, are the last the commercial flywheel reaches, if it reaches them at all.

A field that meets the frontier early helps shape what it is built to do; a field that meets it late can only adapt it after the fact. The fix is not only more funding. The foundational Earth-system model, learned dynamics and physics-based simulation together, with the curated corpus beneath it, is sovereign AI for science: public infrastructure that governments and institutions can run on their own compute, just as Earth-2 is designed to let governments run frontier forecasting themselves. It is one of the highest-leverage public investments of the decade, and one that frontier labs, research institutions and infrastructure companies will need to build together.

There is a further, harder-edged reason this matters now. The AI build-out is only beginning, and the social licence it depends on, public consent for the energy, capital and physical infrastructure it requires, is not guaranteed. It is not a brake on AI. It is a condition for building it at the speed the moment demands.

Every great infrastructure era, from railways to electrification, earned its mandate by delivering benefits people could see. The strongest case for AI is the same: not that it makes firms more productive, but that it helps solve problems people can feel, the climate, disease and the systems everyone depends on. Pointing AI at the public good is not a detour from the commercial project. It is what earns it.

The question, then, is no longer whether we can build world models. It is what we choose to model, and whether humanity’s hardest problems shape these architectures from the start, or inherit them later.

The next frontier for world models may not be enterprise software.

It may be the world itself.

A final note. There is one question this essay has deliberately set aside: what running world models at scale does to AI’s energy demand. It is not a footnote. The field is already forking between architectures with very different computational appetites: models that predict the world in compressed abstractions, and models that generate it frame by frame, in full sensory detail. However fast the chips improve, those choices set fundamentally different floors under energy demand. Which path we build is being decided now, in research labs and procurement meetings, by people who do not think of themselves as setting energy policy. That argument deserves its own piece because the choice between these architectures is also a choice about AI’s future energy demand.

Annex: Four worked examples

Four Earth-system areas where the argument above is especially concrete. Each has the same structural shape: observed state incomplete, dynamics only partially understood, signal living in couplings that no decomposed model represents as a single object. I have worked on all of them.

Subseasonal forecasting. The two-weeks-to-two-months window is the structural gap in global forecasting. Weather skill falls off a cliff at two weeks because the atmosphere is chaotic. Seasonal skill returns — slowly — from the ocean, the land surface, the stratosphere. The window in between depends on how those systems interact. The signal lives in the coupling. Our models represent the subsystems; they do not represent the coupling as an object. A world model, trained on decades of joint reanalysis data, could hold the atmosphere, ocean, land, and stratosphere as a single evolving state and let the coupling emerge. No decomposed model does this. The critical point is not better forecast accuracy at the margin. It is representing a coupled trajectory that nothing currently represents. Early evidence that this is feasible is now emerging, and moving fast. Ola, a coupled ocean-atmosphere model from Princeton and NVIDIA, internally generates El Niño cycles with realistic amplitude and the correct sub-surface wave dynamics, with the oscillation emerging from the learned dynamics rather than being prescribed, though on a short validation record. Newer systems are extending the approach across atmosphere, ocean, land, and sea ice together, closer to holding the whole coupled system as one object. None is finished, but the direction is clear: the dynamics that decomposed models simplify away can be learned from data.

Land–atmosphere coupling and the carbon sink. Soil moisture amplifies heatwaves. Vegetation transpiration sustains rainfall recycling. Root-zone water triggers convection. These feedbacks are at the heart of how a warming world gets more extreme. They also govern the land carbon sink: the terrestrial biosphere absorbs roughly a third of the CO₂ we emit, and a single hot, dry year can flip a whole region from sink to source, as Europe did in 2003 and Amazonia in 2010. Extreme heat and the largest uncertainty in the carbon budget converge on the same blind spot — the water held half a metre to two metres down. But global models handle it through sub-grid parameterisations their own developers acknowledge are provisional and a weak point. Satellites retrieve only the top five centimetres of soil moisture. The water that actually matters sits half a metre to two metres down, unretrieved, at scale, anywhere in the world. A learned coupling model can, in principle, infer that hidden state from its imprint on the joint record, but the variable that matters is not observed. No direct measurement will arrive in this decade.

Ice sheets and sea level. The assessed likely range for 21st-century sea-level rise is roughly 0.3 to 1 metre, with a low-likelihood tail reaching toward 2 metres that cannot be ruled out. That span is wide enough to make coastal adaptation ambiguous for almost every major coastal city. The uncertainty concentrates in processes — grounding-line retreat, sub-shelf melt, potentially ice-cliff instability — whose governing equations are still provisional. Refining the resolution of a model whose parameterisation may be structurally wrong does not resolve this. Observation helps but cannot see beneath the ice or into sub-shelf cavities. A world model here does not replace ice-sheet physics. It integrates satellite, airborne, and sparse in-situ observation into a representation whose inferred dynamics are consistent with the full record at once, rather than with any single parameterisation. The deliverable isn’t a prediction of the year of a regime change. It’s a narrower, more observationally disciplined envelope. For trillion-dollar coastal planning, that is the scientific quantity that matters.

The ocean interior. The ocean holds more than 90% of the excess heat the climate has accumulated since industrialisation. The meridional overturning circulation redistributes heat and carbon at planetary scale. Yet the deep ocean is the least observed component of the Earth system. Below two thousand metres, coverage is sparse. Circulation is inferred, not measured. Different reasonable model choices yield different histories of the Atlantic overturning. Eddy-resolving simulation is an impressive achievement but it does not resolve whether its parameterisations of sub-mesoscale dynamics and ice-ocean coupling are right. A learned interior-dynamics representation — trained jointly on Argo, altimetry, hydrography, and eddy-resolving model output — can maintain a consistent inferred state that no single data product delivers on its own. That state is actionable for decadal prediction and for the range of plausible futures the centennial projections must span.

Whether an initial peace handshake is a week or a month away, a measure of damage is already baked in. A political handshake is not a full deal, and a full deal is not an operationally open Hormuz — add shipping times and constraints on flows, as earlier posts have covered, and longer still. With those lags built in, the world is close to a tipping point: the threshold where the subdued economic response we have seen so far gives way to something materially more recessionary.

This piece works through where that tipping point may sit, the three structural shifts that will define the energy security system being written underneath, the cost ledger emerging, and the choice it forces on capitals.

Phase one has ended

So far, the pain has been concentrated in emerging economies in Asia and Africa most reliant on Gulf oil and gas imports — headlines vary from fuel queues, rationing and blackouts, to four-day working weeks and industrial curtailment.

The pain has now moved West, while deepening in Asia. European diesel is around €2 a litre, up more than 20 per cent since the war began. US gasoline is up more than 50 per cent, and back-to-back US inflation reports have pointed to the ongoing conflict with Iran. The IEA warns of jet fuel shortages within weeks in Europe. Pakistan, Indonesia, Vietnam and the Philippines are approaching critical inventory levels. Roughly 15 million barrels a day are off the market. The buffers — strategic reserves, commercial inventories, refining run cuts, adjusting buying cycles — are running out. The United States is partly insulated by domestic supply but not from prices.

Approaching the tipping point

How traumatic this becomes depends on how much longer the strait stays closed. Even in the optimistic scenario — a peace framework agreed by end of May — full terms could take another month, and if shipping stays at a standstill, inventories drop to critical levels by mid-to-late June. With supply-side levers nearly exhausted, the system nears its tipping point: the threshold where shortages would drive prices high enough to constrain demand, turning a supply shock into a recessionary force. Oil prices would have to rise high enough to reduce 10 to 15 per cent of world consumption. That is the trajectory the world is staring down.

In 2008, Brent peaked at $147 and demand fell, but much of the destruction came from the credit crisis arriving alongside, not the price alone. The urgency for success to emerge from any peace talks is clear: a Hormuz closure into the summer would push prices well above 2008 levels. Oil-price-driven demand destruction at this level would be unprecedented in the history of oil markets.

Globally, prices could turn powerfully recessionary, sharpest in the import-dependent economies of Europe, Asia and Africa. Central banks would have no clean response: every monetary choice trades inflation against output. Europe absorbed 2022 without sustained recession at a cost of €700bn in fiscal cushioning. The fiscal headroom is now structurally narrower. Bond markets and reactivated EU fiscal rules will discipline any €700bn-scale repeat.

Even after the strait reopens, a multi-month normalisation will further stress inventories. The US military has begun clearing Iranian mines, but the operation will take months. Escort capacity caps throughput at perhaps ten to twenty vessels a day, against pre-conflict levels of 135. Insurance must normalise; crews must be willing. When cargoes sail, the first relief from a reopened Hormuz reaches Pakistan in 6-7 days; Vietnam in 15-16; Europe — routed around the Cape of Good Hope — in 38-39. The paper market adjusts with political news; it is not reflecting the physical stress. Neither market is pricing what comes next: a system rewired by baked-in stress that no agreement will reverse.

The system that emerges

Every major energy crisis has reshaped the energy security paradigm. The 1973 Arab oil crisis built the International Energy Agency, coordinated stockpiles, and decades of efficiency policy. 2022 forced Europe’s pivot from Russian pipelines to seaborne LNG. Hormuz is writing the third.

Three features will define it.

First, a new bill for redundancy. Crude oil could always be stockpiled: the Strategic Petroleum Reserve, the IEA’s 90-day cover, all built for a commodity that stockpiles cleanly. In many countries, the physical market is most stressed not in crude, but in oil products and LNG. Gas storage exists for seasonal balancing, not strategic reserve. Diesel and jet fuel degrade in storage and cannot move easily between markets with different specifications. Where stockpiling fails at strategic scale, substitution becomes the only durable response.

Even where stockpiling does work, every importing economy now needs more of it. The fossil system that survives this crisis will be structurally more expensive. Strategic reserves restocked above pre-war levels in IEA economies. Gulf-exposed non-IEA importers building reserves at scale for the first time. Refineries holding more crude stock after a decade of running lean. Pipeline duplication and bypass capacity projects moving forward, proposals that might have been politically impossible six months ago. War-risk insurance repriced for chokepoint exposure, unlikely to unwind. If a toll on Hormuz transit emerges from the negotiations, it becomes a permanent line on every cargo bill. This is a redundancy tax on the integrated fossil energy system — insurance against geopolitics, paid by every importing economy. It does not unwind when the strait reopens. Each of these costs was not there before March. Each is now structural.

Second, demand permanently redrawn — beginning in Asia. Beneath the visible crisis responses is a forced acceleration of electrification, clean energy and resilient supply: capital-intensive and distributed, rather than dependent on continuous imported flows. South Korea is tripling renewable capacity to 100 gigawatts by 2030, with its energy minister calling for a system “not shaken by external shocks such as the Middle East war.” The Philippines declared a national energy emergency in March, with 250 MW of solar-and-storage online within weeks; its energy secretary called renewables and storage “a strategic necessity and a national imperative.” Indonesia is moving on 100 gigawatts of solar that political inertia had stalled for a decade, with President Prabowo calling it “urgent, given the current situation.”

The momentum reaches beyond Asia. Kenya and Ethiopia are among the African economies most exposed to Gulf supply: in March, Ethiopia’s solar imports rose 391 per cent and Kenya’s 207 per cent. Fifty countries set all-time-high solar imports. April saw the largest monthly inflow into global renewable energy ETFs since January 2021. EV sales in the EU jumped nearly 50 per cent in March. These are not policy responses. They are population responses. Policy responses can be reversed. Population responses, once wired into homes and habits, are stickier.

Europe is the second wave, accelerating from May onwards. The post-2022 pivot exchanged a pipeline dependency for a seaborne one, Qatari and American LNG, and half of that has just broken. As effective closure persists, Europe and Asia compete for the same Atlantic cargoes in the refill window. Chemicals, fertiliser and primary steel will be first in line for curtailment. We saw this in 2022. European heavy industry began moving to the US and Asia, from BASF closures at Ludwigshafen to ammonia capacity moving to the US Gulf. Hormuz accelerates the offshoring: capacity that closes does not reopen, because capital moves to where feedstock is cheaper and stays. This is not just gas. Europe is also weeks from running out of jet fuel. Middle East imports, the IEA’s Birol said, are “basically now almost zero.” On 3 May, the UK government relaxed slot rules to let airlines consolidate or cancel flights amid jet fuel concerns.

The choice for Europe’s import-dependent governments is now binary in a way it was not in 2022. Either accelerate electrification, clean energy, efficiency and sustainable fuels at a pace that compresses a decade into three years — or accept the loss of energy-intensive industry as a permanent feature of European economic geography. The Commission’s AccelerateEU package, adopted on 22 April, names electrification, grids and renewables as the structural answer. Either path produces a smaller European fossil demand base.

Third, a segmented geopolitical re-sorting. Alongside fossil supplier diversification, substituting oil and gas creates new dependencies. Where electrification is fastest — solar, EVs, grid-scale batteries, and the critical-minerals that underpin them — China dominates. Across the emerging world, China reads as the reliable commercial partner; the United States as a source of volatility. In other areas — heat pumps, nuclear, grid technologies, long-duration storage, industrial decarbonisation, sustainable fuels — supply is more diversified, and Western industrial competitiveness is contested. Where China leads, dependence consolidates. But this is not Gulf or Russian dependence on continuous flows. The Chinese import is a manufactured item — a panel, a battery, a transformer. Once installed, it generates power for decades. The dependency is real, but the overnight risk is not. Critical minerals are the exception, and strategic stockpiles and diversification are required.

This week’s Trump-Xi summit in Beijing showed the pattern in real time: G2 cooperation on the immediate crisis (open Hormuz, no toll, China redirecting oil purchases toward the US), structural competition on the longer term (electrification supply chains, the manufactured-asset dependency map). The political theatre doesn’t change the geography being drawn underneath.

The crisis presents a narrow political window. Memories are short, and urgency fades when prices fall. Europe’s 2022 response accelerated structural change, but the €700bn was spent on universal subsidies and long-term LNG contracts, reinforcing the old paradigm. What happens next depends on whether governments commit to the full portfolio — renewables, efficiency, grid investment, long-cycle industrial capacity including nuclear — while protecting vulnerable households through targeted support, not broad fossil subsidies.

The cost ledger is already visible

At the global scale, today’s Energy Transitions Commission report on the Hormuz crisis estimates the additional cost to the world economy this year at $1–2 trillion, around 1–1.5% of global GDP. The world spends roughly $4 trillion a year on fossil fuels; the crisis is adding a quarter to a half of that bill in shock costs alone.

In Europe, the numbers are already concrete. On 29 April, Ursula von der Leyen put Europe’s additional oil and gas bill at almost €500mn a day: €27bn in 60 days (by that point) that buys, in her words, “not a single additional molecule of energy.” If Hormuz stays closed, the daily bill could plausibly rise by 20-30 per cent through May and June as inventories deplete and physical products price higher. By the end of June, Europe will have paid roughly €60bn more for oil and gas. And the bill will continue.

Whenever the strait reopens, a multi-month normalisation will follow. Demining, reinsurance repricing, refinery restarts, the return of shut-in production, the rebuilding of strategic reserves and the honouring of bilateral fuel-supply agreements all delay the adjustment. Prices do not snap back; the bill stays elevated for longer. Back-of-the-envelope arithmetic puts the additional annual cost to Europe at €100-150bn on a mid-year reopening. A delayed reopening — closure persisting through summer — would push the fossil bill beyond €200bn and could follow the trajectory familiar from 2008: a sharp spike to prices well above 2008 highs, demand destruction following, then a sharp price collapse. Lost output, industrial offshoring, and financial sector stress would together dwarf the €700bn Europe spent cushioning 2022.

At €100-150bn, Europe’s additional bill would run at roughly a third of the €450bn that Mario Draghi and the European Commission estimate Europe needs to invest each year to reach its 2030 climate goals. For the same cost, a meaningful share of that gap closes.

That comparison matters. Spent on fossil fuels at elevated prices, the money buys the same commodity at a higher cost, a recurring premium. Spent on redundancy — pipeline duplication, higher inventories, repriced insurance — it buys partial protection. But redundancy cannot cover what stockpiling cannot reach. Spent on clean energy and flexibility — efficiency, storage, long-lead-time investment in grids and industry — the money builds assets that reduce exposure to the next disruption. Once built, these systems produce power for decades with far less exposure to imported fuel flows. All three costs will be paid. For importers, the third offers both output and resilience. Redundancy buys protection for the next crisis. Substitution reduces the bill for every crisis that comes.

The strait will reopen. The energy security architecture that survives won’t be the one we entered the crisis with. The question for capitals is not whether the system has changed — it has — but how much they will pay for redundancy versus substitution. The first protects against the next crisis. The second reduces the bill for every crisis that comes.

By way of background, I have recently been delighted to contribute my time as a Senior Advisor to the excellent team at the Energy Transitions Commission (ETC). For the ETC’s newly published and detailed analysis of emerging lessons from the ongoing Hormuz crisis, see the Commission’s new report out May 15th 2026: https://www.energy-transitions.org/hormuz-crisis-clean-energy-key-to-reducing-economic-cost-of-fossil-fuel/

New here? This is the fourth post in a series tracking the Hormuz disruption, and the first on gas.

The largest oil and gas supply disruption in the history of global energy markets is still running. The oil story has dominated. The gas story, much less so. That is what this piece is about.

India is rationing gas. Taiwan has just over a week of reserves. Bangladesh can no longer afford what is available on the market.

At the same time, Asian LNG benchmarks have retraced from crisis highs. European prices are well below their March peaks. Cargoes trade. Available supply finds buyers.

The Market is Clearing. The system is straining anyway.

This is not a contradiction. It is how modern energy systems fail under stress: balancing globally while failing locally, pricing smoothly while rationing physically.

The pain is already there. It shows up in different places, in different ways, and at different times — compounding with every week the disruption persists. That distribution, easy to miss when prices look calm, is what most commentary is missing about this phase of the crisis.

One fact matters for everything that follows: March 1 was, seasonally, the most favourable moment a Hormuz shock could have occurred. In any other season, the break point would arrive sooner. We will come back to it.

And the direction of travel for many importers after the Hormuz crisis? For most, it is not to secure more LNG, but to need less.

Not a Shortage. A Flow Problem.

Most commentary is still reading this as a supply shock. It isn’t.

Energy crises are usually defined by supply shocks. The 1973 Arab oil embargo, the 1979 Iranian revolution, and the European gas shock that followed Russia’s 2022 invasion of Ukraine all involved major disruptions to physical supply. In 2022, the defining feature was structural loss — pipeline gas from Russia to Europe was removed, and the system had to adjust by replacing it, largely with LNG from the United States.

This crisis is different. The gas exists. Global LNG production has not collapsed. What has collapsed is the system’s ability to move gas where it is needed, when it is needed.

In a market built on highly specialised infrastructure — liquefaction plants, purpose-built vessels, regasification terminals, long-term contracts — what matters is not aggregate production. It is whether the right molecule reaches the right buyer at the right time.

Hormuz has broken that chain. Qatari and UAE cargoes cannot sail freely. Part of Qatar’s export capacity is offline following infrastructure damage, with two of the country’s fourteen liquefaction trains estimated to remain out for three to five years. Atlantic Basin supply cannot substitute instantly — it requires ships, insurance, berths, and time.

The starting point is not reserves. It is flow.

Prices tell you what the market is pricing. Constraints tell you what the system can actually do. The distance between the two is where this crisis lives.

Three Constraints

Strip away the headlines and the LNG market reduces to three constraints. Every country-level outcome follows from them.

Effective supply. Not what is produced, but what can actually move. Cargoes stranded behind Hormuz. Capacity offline at Ras Laffan. Combined, a large share of globally traded LNG is functionally unavailable to the buyers who need it. It is an enormous number, but it does not show up cleanly in prices because it is offset, incompletely, by displaced flows finding new routes.

Spot market constraint. Roughly three-quarters of LNG trade is locked into long-term contracts. The marginal price is set by a much smaller pool — mostly flexible US volumes. When disruption hits, long-term supply does not reallocate; it simply stops delivering. Displaced demand piles into the remaining tight spot market, which clears, but only by pushing weaker buyers out.

In effect, 5 to 10 percent of tradeable LNG is setting the price for the other 90 percent.

Time. Everything in LNG takes time. Insurance clearance. Loading. Transit. Regasification. Even a clean restart takes weeks to translate into delivered supply. For some countries, particularly in Asia, that is longer than their entire gas buffer. The gap between days of cover and days to arrival is the critical variable, not because gas does not exist, but because it cannot arrive in time.

The system is operating at the knife-edge. Small changes do not smooth outcomes. They flip them.

Can the United States fill the gap?

That is the obvious question, and the one I am regularly asked. The answer is quieter than the politics suggests.

US LNG exports are running at near-peak capacity: 17.9 Bcf/d in March against 18.3 Bcf/d of nameplate. Spreads to Europe are up over 80 percent in a month; to Asia, nearly doubled. The price signal is the strongest on record. And the system cannot respond. The only near-term flexibility in the entire US export complex is the deferral of scheduled maintenance windows. New capacity is coming — Corpus Christi Stage 3, Golden Pass, Port Arthur — but in increments of 0.6 to 1.6 Bcf/d, against a 10 Bcf/d Qatari hole, and on timelines measured in quarters, not weeks.

Nor can the major traders reallocate around the gap. The combined LNG portfolios of Shell, TotalEnergies, BP, ExxonMobil, Vitol, Trafigura and a handful of others total around 200 MTPA, roughly half of global trade. But headline portfolio size is not the relevant number. Most of those volumes are bound by long-term offtake contracts, tolling commitments, and shipping positioning. Actual day-of-diversion capacity is 10 to 15 MTPA.

Hormuz has effectively removed around 80–85 MTPA from the addressable pool — close to the entire annual volume that trades on a flexible, spot or short-term basis (~100 Mtpa). Portfolio players can reallocate molecules; they cannot create them.

The US is the supposed replacement. The molecules exist, but the infrastructure to move them does not.

At the heart of the crisis: flow, not volume; timing, not tonnage.

A System That Splits

These constraints do not hit evenly. That is the point most easily missed.

The gradient runs from countries already past the line to countries barely affected.

Bangladesh is already rationing, not because gas does not exist globally, but because it cannot secure it. With 70 to 80 percent of LNG supply at risk, affordability and availability have effectively merged. India has moved to formal allocation, protecting households and transport while industry absorbs the cut. Taiwan is running a clock: its buffer sits below the time it takes for replacement cargoes to arrive. South Korea is acting before constraints bind, reducing LNG generation, accelerating nuclear restarts, building optionality. Europe is not short today but is behind on refill, facing a competitive winter in a tighter market. Japan sits at the price-exposed end of the insulated group — the system holds, but at rising cost. China is structurally insulated: scale, pipeline anchors, and domestic substitution absorb the shock.

Exposure is not the same as vulnerability. A country can lose more Qatari gas than its neighbour and still be less stressed because it has deeper buffers, wider substitution pathways, or a stronger pipeline anchor. The binding constraint is not the flow figure. It is the interaction between exposure and the system’s ability to absorb it.

The exhibit below sets out ten importers across the gradient, with binding constraint, specific exposure, and trajectory for each. Fuller country-level detail is in the annex at the bottom of this post, alongside the methodology for the LNG disruption analysis.

A single-metric view of this crisis, i.e. “how much gas is at risk?” misleads. The crisis is not synchronous. The countries are not in the same moment. They are on different clocks, arriving at different points in the same sequence, each absorbing the shock through whichever dimension its system is weakest on.

Price Transmission, and What it is Hiding

So why are prices telling a softer story?

Because this is a system clearing at the margin, and markets are better at pricing short-term risk than sustained constraint. When 5 to 10 percent of flexible LNG sets the price for a much larger contracted market, prices can look contained even as underlying conditions tighten.

The divergence shows up regionally. Asian LNG (JKM) spiked into the high $20s/MMBtu after the initial disruption. European TTF followed, less sharply. US Henry Hub barely moved at all, anchored by abundant domestic supply and the limited near-term ability to expand exports. Gas is not a single global market — it is a set of regional ones, stitched imperfectly together by LNG infrastructure. Constraints can be localised even when the system as a whole appears balanced.

The ceasefire pattern reinforced the point. Prices softened on the announcement. Physical flows did not meaningfully recover. Over two weeks of ceasefire produced almost no real restoration of LNG movement, Qatari tankers approached the strait and turned back rather than risk passage. Prices moved on narrative. Physical reality did not move at all.

The price transmission that is occurring is landing more in bills than in benchmarks. Wholesale gas prices in Europe doubled at their March peak and remain well above pre-crisis levels. Even if they stabilise, the requirement to refill storage into a tighter, contested market keeps the effective floor well above pre-crisis levels.

For import-dependent economies absorbing a stronger dollar, local energy costs continue rising even as headline benchmarks drift down. Households and firms on fixed contracts do not feel the full effect immediately — they feel it at renewal. That extends the economic impact well beyond the initial shock and keeps the political pressure accumulating long after the crisis looks, from the screens, like it has passed.

Stress is building in the gap between paper markets and physical flow. The gap between headline prices and delivered bills is where the politics is.

Even when the strait eventually reopens, the bill does not end with it. Stockpile drawdown and contract renewals carry the inflationary pressure forward for quarters.

Constraint Now, Glut Later

This is where, in my view, most commentary breaks.

The striking thing is, a sharp price collapse from here would not disprove the crisis. It would confirm it.

Flows do not resume smoothly when the strait reopens. They restart in steps, gated by insurance underwriters, by crew willingness, by loading queues, by congested shipping. Even when officials briefly declared the strait open to commercial traffic, Qatari tankers approached and turned back. Uncertainty itself sustains the constraint, long after the physical bottleneck eases.

With the replacement source already at its ceiling, the adjustment comes from demand, not supply. By the time supply properly returns, the system will already have adjusted. Demand will have been reduced. Industrial activity curtailed. Procurement deferred. Cargoes that were delayed do not disappear. They simply arrive later, into a market that has already cut consumption.

The result: temporary oversupply. A sharp downward move in prices.

A price collapse, if it comes, would not signal that the system was never tight. It would reflect how thoroughly that tightness has already been absorbed, through destroyed demand, lost output, and industrial capacity that does not instantly reboot. Some of that demand will not return at all. Crises accelerate electrification and efficiency. Structural shifts already under way get pulled forward, and policy amplifies them. The European Commission’s push toward formal electrification targets as part of its energy security response is the early signal of exactly this.

The LNG market was probably heading toward structural oversupply in the late 2020s anyway, as new US and Qatari capacity comes online. This disruption accelerates the whiplash: acute constraint now, sharper oversupply later, as delayed cargoes hit a market that has both shrunk and was already trending toward glut.

This asymmetry is sharper in LNG than in oil. Delivery depends on tightly coordinated infrastructure and scheduling. Disruption cascades through the schedule, not just the price. When it reverses, it reverses the same way.

Constraint now, glut later. Both true. The sequencing is the story.

The 2022 analogue: where it holds, where it breaks

This is not the first time the system has been stressed at this scale, and the cleanest analogue is the most recent one.

Russia’s removal of well over 100 bcm of pipeline gas from European supply over 12–18 months was comparable in magnitude to a Hormuz shock, though slower in onset. Europe absorbed it through demand reduction, substantial additional LNG, incremental Norwegian and Algerian pipeline gas, and storage management.

The market did not break. It tightened, transmitted price violently, redistributed pain inequitably, and forced political intervention. That is what stressed energy markets do.

The consequences extended well beyond gas. Prices fed into electricity, industrial costs and inflation. Emerging-market buyers were priced out. European industry curtailed. Fertiliser and chemicals were among the hardest hit.

A 2026 Hormuz disruption would be comparable in scale but different in structure: more concentrated in the globally tradeable LNG pool, faster in onset, and subject to additional deliverability frictions (shipping, insurance), even as the global LNG base and portfolio flexibility going into this crisis are larger than they were in 2022. Europe’s cushion is also different: not smaller in absolute storage terms, but thinner relative to the gap that must be absorbed. Storage is now the planned baseline, not a strategic surplus.

That is what we are seeing now: degraded clearing at prices that force multi-region demand destruction and buyer-tier exclusion, but cargoes that continue to move. The market will function, but with the strain showing up in the economies and politics it serves.

That is the comparison if Hormuz reopens within weeks. It changes if closure persists into Q3 — through the European refill season.

The Hormuz crisis remains within 2022-type strain for as long as the LNG market can rebalance through reallocation, pricing out marginal buyers and redirecting cargoes while core economies maintain their storage and contractual buffers. It moves beyond 2022 once reallocation can no longer restore those buffers ahead of winter, most clearly when European storage refill falls persistently off the trajectory required for a functional heating season.

Under a March 2026 closure that persists into Q3, that transition is likely to become visible through mid-to-late Q3 2026, as the marginal adjustment shifts from emerging-market exclusion to industrial curtailment inside OECD economies, increasingly an allocation outcome rather than a pure price response. Winter 2026–27 is then not the trigger, but the point at which that earlier shift is exposed: a system entering peak demand without the storage cushion that ultimately stabilised 2022.

March 1 was, in seasonal terms, the most favourable moment a Hormuz closure could have occurred. In any other season the break point would arrive sooner — a winter closure compresses the timeline by half. Winter 2026–27 is the late case, not the central one. By break point I mean the moment when reallocation can no longer rebuild buffers ahead of winter demand and the marginal balancing tool shifts to industrial curtailment inside OECD economies. Not collapse. Allocation, by force, where price alone can no longer clear.

What Hormuz Has Already Revealed

What is worth sitting with, nearly two months in, is this: a system can appear to be functioning while already failing for those least able to absorb it. Aggregate prices can drift down while specific buyers are priced out. A market can clear mathematically while breaking operationally.

This is not just a disruption of energy flows. It is a divergence in how states can respond to them. Countries with flexibility and institutional capacity to act early — to reroute supply, manage demand, absorb shocks — are operating on a different footing from those that cannot. Responses that work for one constraint do not work for another. Releasing strategic stocks does not help a country that cannot afford the spot market. Subsidising consumers does not solve a timing problem.

The deeper shift is in what energy security actually means. For half a century, the frame was supply: enough of it, at tolerable prices, reliably delivered. That framing is no longer sufficient. In an interconnected energy system, resilience is a function of flexibility (storage depth, procurement diversity, demand-side optionality), and the institutional capacity to act before constraints bind rather than after.

Flexibility is now the scarce resource. And not all countries can buy it.

This matters beyond gas. The logic — constraint expressed through flow rather than volume, through timing rather than tonnage, through the ability to move supply through infrastructure under stress — applies to any tightly coupled system the global economy depends on. Semiconductors. Critical minerals. Batteries. All of it moves through a handful of physical chokepoints, and Hormuz has just reminded us how exposed those chokepoints remain. The next chokepoint crisis will not be about running out. It will be about being unable to move what exists.

When I look at benchmarks, it’s clear they show some of the regional picture — Asia harder than Europe, Europe harder than the US. But that is surface only. What they do not show is the damage absorbed (or soon to be absorbed) in fiscal cost, industrial margin, and deferred procurement. Or the damage still to come at renewal, refill, and refinancing that may fuel inflation. The pain already baked in is far less visible than the market’s reaction to diplomatic headlines.

The acute phase of Hormuz is not yet over. But its shape is now clear.

The countries that emerge from this intact will not be the ones with the most gas. They will be the ones that built the systems to need less of it. Electrified systems are front-loaded. Once built, they produce locally, without fuel passing through anyone’s waters.

The long answer to Hormuz is not more gas. It is a different architecture.

That is the question this crisis has forced open: : not how to secure the next cargo, but whether the system that makes cargoes necessary at this scale is the one worth defending.

For readers who want the full country breakdown, the annex below sets out the detail behind the model.

Annex: country-level analysis

Data reflects the model baseline as of late April 2026. Figures are modelled estimates, directional rather than precise. See methodological note at the end for full sourcing.

Bangladesh

Already past the line. Roughly 35 to 45 percent of Bangladesh’s LNG comes from Qatar under long-term contract — cargoes stranded inside a closed strait. Spot replacement exists globally but not at prices Dhaka can pay, and not for a sovereign already carrying hundreds of millions in overdue LNG bills. Domestic production still covers most gas supply but is declining and inflexible; LNG is the marginal fuel behind the power grid. The result is a financial shortage that manifests as a physical one — at which point ‘too expensive’ and ‘not available’ become the same thing. Load-shedding widens, industrial gas is rationed first, fertiliser output falls. The correct comparator is Pakistan, not India. The fix sits in the IMF and the currency market, not the regas terminal.

India

India has moved to formal allocation. Qatar and the UAE together supply roughly fifty percent of India’s LNG. March delivered just 10 cargoes, barely a third of normal supply. The government is reserving 100 percent of supply for household cooking gas and transport fuel, while cutting industrial deliveries by up to 40 percent in affected segments. Curtailment is significant, but administered rather than market-driven. India can still procure gas. It has chosen where the pain lands. Unlike Bangladesh, the state has the policy and fiscal headroom to steer the shock rather than be overrun by it.

Taiwan

A timing case. Taiwan’s LNG buffer sits below the time it takes for replacement cargoes to load, sail, and regasify. Supply is secured through May; beyond that, procurement becomes sharply harder. Active emergency buying is already visible. When reassuring government statements coincide with urgent procurement, the urgent buying is the leading indicator. The system has not failed. It is running out of runway. Taiwan’s constraint is not affordability and not allocation — it is the clock.

South Korea

Pre-emptive adjustment. Direct Hormuz exposure is around 14 percent — material, but cushioned by diversified long-term contracts and deeper procurement flexibility than peers. What distinguishes Korea is institutional behaviour: reducing LNG generation, accelerating nuclear restarts, emergency market stabilisation measures. Korea is acting before the constraint binds. The cost is consumed optionality — flexibility spent now that will not be available later if the disruption extends.

Japan

Insulated physically, exposed financially. Qatar is only 4.4 percent of Japan’s LNG intake; the UAE is a rounding error. Combined Hormuz exposure is around 5 to 6 percent. The pressure on Tokyo runs through price, not cargoes. Japan is the world’s second-largest LNG buyer and the market’s most price-sensitive swing importer — absorbing the cost as Europe and the rest of Asia compete for replacement volumes. Nuclear restarts, standing coal capacity, and oil-fired dispatch absorb the residual physical stress. Household bills are up around ¥15,000 a month. The binding constraint is industrial competitiveness and the fiscal cost of subsidising power — not keeping the lights on. Japan’s vulnerability in this shock is macroeconomic and political, not physical.

Europe (aggregate)

Not short today. Scheduled for shortage. EU storage entered the 2026 injection season at around 28 percent, against roughly 50 percent a year earlier — a gap of around 60 percentage points to the ~90 percent November target under scenario assumptions where Qatari volumes remain absent. Injection is running at around one-third of normal April pace. That gap must now be closed while competing directly with Asia for marginal Atlantic cargoes, and Atlantic supply is already stretched, replacing what Russia used to provide rather than adding new volumes on top. The adjustment is binary: either the European–Asian spread narrows, or European industrial demand is forced out. No third option at scale.

United Kingdom

Europe’s most structurally fragile major market. Working gas storage of only around 3.0 to 3.2 bcm — roughly 10 to 12 days of peak winter cover, the thinnest in developed Europe. UK Continental Shelf production is declining at around 11 percent per year, deepening LNG dependence even as the buffer shrinks. Its direct LNG share of demand is modest at around 11 percent — the exposure is not in the flow, it is in the structural cushion, which is almost non-existent. Its interconnectors can flip to drain UK stock under tightness, and it sits outside EU solidarity. The system is holding — adjusting through prices rather than volumes — but the real risk sits in the spring-summer 2026 procurement window, when volumes must be secured into a tight and largely pre-contracted global market. Winter 2026-27 is where the consequences land — on industrial costs first, then, with a lag, on household bills and inflation.

Italy

Italy carries Europe’s clearest direct-flow exposure. Around 6–8 bcm a year of Qatari LNG — roughly 10–13 percent of total gas supply — is concentrated at Adriatic LNG, the offshore terminal that ran almost entirely on Qatari cargoes and has now gone quiet under QatarEnergy’s force majeure. Pipeline alternatives from Algeria and Azerbaijan are already near capacity, limiting how much can be replaced. The constraint emerges early, in the spring-summer procurement window, as replacement volumes prove hard to secure. It lands next in industrial demand, where ceramics, glass, chemicals, and steel feel the price signal first.

Germany

Germany has almost no direct Qatari LNG flow through direct terminals, but its exposure runs through two channels the headlines miss. The first is the refill gap: Europe’s deepest storage system, around 24 bcm, entered the injection season some twenty percentage points below last year’s pace, and that deficit now has to be closed in a Qatar-limited, Asia-bidding market. The second is the Belgian transit node at Zeebrugge, where QatarEnergy has booked full regas capacity through 2044 — a contracted channel into the Rhine industrial corridor that is now sharply constrained. Scale amplifies both, as when German companies like BASF, Covestro, and Thyssenkrupp pay more for gas, the cost transmits across European manufacturing.

Spain

Europe’s most secure major gas market, and structurally different. It has among the lowest gas-for-power shares of any major European economy. Only France is lower, and the two arrive there by different routes: nuclear in France, renewables in Spain. Spain’s LNG supply is also diversified — fourteen origins and the Algerian Medgaz pipeline, running near capacity, leave Hormuz exposure in the single digits. The winter-peak channel that binds the UK and Italy — where gas prices propagate into power prices and marginal LNG cargoes are pulled — is structurally weaker here. Spain relies less on marginal LNG to keep the power system balanced; industrial demand remains exposed, but the power-sector pull is narrower. Yet Spanish security does not transmit northward: the Pirineos interconnector caps exportable flexibility at roughly 7 bcm a year. It is a national outcome, not a European one.

China

Structurally insulated. Diversified sourcing through Russian pipeline gas, long-term Australian LNG contracts, and Central Asian pipelines materially reduces Hormuz dependence. China has reduced LNG imports during the disruption rather than competing for scarce cargoes — evidence of flexibility, not vulnerability. More structurally, as the world’s first true Electrostate, China has reorganised its economy around electrification and can redispatch coal-fired generation when oil and gas tighten, without the political constraints that bind European governments. The flexibility itself becomes a form of power: while other major importers face procurement urgency, Beijing shapes how it responds rather than simply enduring it.

Methodological note

This piece draws on a structured, scenario-driven model of global LNG disruption, designed to separate system mechanics from price signals.

The model is built around three constraints: effective supply availability (combining transit disruption, infrastructure outages, and operational constraints); spot market constraint (the relationship between displaced demand and the available uncontracted LNG pool); and buffer-to-delivery-time dynamics (country-level storage and flexibility against the time required for replacement cargoes to arrive). These are assessed independently and then combined to determine country-level outcomes.

Delivery timelines incorporate loading delays, voyage times, restart frictions, and destination-side handling. Demand adjusts through a combination of price response, administrative rationing, industrial curtailment, and deferred consumption. Country outcomes are determined by binding constraints — affordability, procurement access, timing, or storage trajectory — rather than by a single global shortage metric.

The framework has been extended through successive runs. Country exposure is now assessed across three distinct rankings — gross flow exposure, net replacement difficulty after substitution, and structural system fragility — which can rank countries differently. Systemic risk thresholds have been operationalised separately at country, regional and global level with explicit timing; calibration is anchored to the most recent official statistics, overlaid with live 2026 operational data and adjusted for re-export flows and interconnector flow direction; and a parallel contractual-transmission layer tracks force majeure cascades alongside the physical disruption.

Inputs draw on a tiered hierarchy of sources: primary market data (vessel tracking and cargo flows via Kpler, LNG market activity via ICIS, confirmed operator statements and force majeure declarations); structural system data (IEA country-level gas balances, BloombergNEF trade flow modelling, Eurostat and national TSO disclosures); infrastructure intelligence (confirmed liquefaction and regasification outages, cross-validated across sources); and market signals (JKM, TTF, NBP, and Henry Hub benchmarks, storage injection rates, policy responses, and procurement behaviour).

All numerical values are modelled estimates designed to capture relationships between supply, demand, and time, not exact point forecasts. A key distinction is made between absolute physical shortage — where supply is unavailable at any price — and supply-constrained rationing, where gas exists globally but does not reach end-users in sufficient volume. In practice, the latter produces outcomes that function as shortages at the point of consumption.

New here? This is the third post tracking the Strait of Hormuz crisis. The first — When the Buffers Run Out — asked which countries could bridge the disruption and for how long. The second — How the New Blockade Could Deepen the Crisis — examined how escalation pathways could remove the alternative supply routes keeping the crisis manageable. This post carries the full restoration analysis promised in Post 2: what happens when the strait reopens, the assumptions that determine recovery timelines, and why some countries bridge and others don’t. Start with Post 1 if you haven’t. The context matters.

The Strait of Hormuz is “open.” But almost no one is using it yet.

A handful of oil tankers raced toward the waterway on Friday after Iran declared it operational. Most stopped short of actually transiting. At least 135 million barrels of crude and refined products remain stuck on tankers in the Persian Gulf. Bloomberg spoke to over a dozen ship owners, and most indicated a “wait-and-see approach.” Industry groups urged continued caution.

This crisis is not primarily about supply. It’s about timing.

The question is no longer whether oil exists, but whether it arrives before countries exhaust their buffers.

The gap between political announcement and operational reality creates exactly this timing problem. President Trump claimed Iran agreed to “never close the strait again.” Iranian Foreign Minister Abbas Araghchi confirmed only “passage for all commercial vessels through the Strait of Hormuz is declared completely open for the remaining period of the ceasefire”—but requires coordination with the Islamic Revolutionary Guard Corps Navy.

Bloomberg delivered the reality check: “There is nothing to that effect from the Iranian side” on Trump’s broader claims. The shipping industry revealed which version they believe. “The market will treat it with cautious optimism rather than full confidence,” said Captain Farhad Patel of Sharaf Shipping Agency in Dubai. Lloyd’s of London hasn’t reinstated coverage despite the announcements. War risk premiums remain at crisis levels—100 times normal rates.

The question now becomes: which countries have enough buffer to outlast the time it takes for supply to return?

The Day 49 Timing Arithmetic: When Every Day Counts

This analysis captures the system at the moment of announcement—before commercial flows resume.

The key question is simple: does a country have enough buffer to outlast the time it takes for supply to return? The bridge calculation measures whether remaining reserves can bridge the full gap between announcement and usable supply.

Starting from Day 49, with 49 days of crisis drawdown already behind us before the Hormuz reopening was announced, the model tests whether countries can survive the complete restart sequence—insurance coverage resumption (one week assumed), plus voyage times from Hormuz ranging from 9 days (India) to 38 days (UK via Cape route).

Positive bridge values indicate countries can wait out the delays; negative values show reserves depleted before cargo from Hormuz arrives. The methodology inclorporates supply offsets from the IEA emergency release and floating cargo, but does not incorporate demand destruction, emergency alternative sourcing, or other adaptive crisis measures that governments typically deploy to extend timelines.

The paper market prices the reopening announcement. The physical market waits for the first cargo to dock.

These figures are modelled outcomes reflecting orders of magnitude rather than precise day-counts. Faster restart sequences would improve all results; slower processes or additional delays would worsen them. Detailed methodology, data sources, and key assumptions are provided below.

EXHIBIT: Country Bridge Status — Day 49 Operational Restart

Note: This analysis focuses on countries with remaining buffers at Day 49 including average voyage time to man country port, and an insurance lag of 7 days. Several countries had already exhausted reserves before the reopening announcement: Sri Lanka, Pakistan, and Bangladesh face severe shortage, while the Philippines shows modelled shortage. The Day 49 timing arithmetic applies only to countries that reached the announcement with reserves intact.

India LPG: LPG constraint not quantified due to limited stock transparency, but represents critical vulnerability (90% of imports via Hormuz, 60% import dependency). See main analysis for details.

UK crude: Aggregate result reflects strategic reserve coverage for Hormuz-dependent flows only (~6-8% of total crude imports). Calculation does not credit North Sea production (~1M bpd) capacity to substitute for lost Gulf imports (~150-200k bpd), which would significantly improve aggregate crude security. Products constraints represent the binding limitation.

The exhibit shows a simple truth: survival depends not only on how much you have, but on how fast supply returns.

The results reveal three distinct tiers. Countries with massive buffers—Japan and China—have time regardless of diplomatic uncertainty. Countries with moderate buffers bridge, but with margins measured in single-digit days—where one failed transit or one additional week of insurance hesitation flips the verdict. And countries where the relief timeline exceeds the buffer face deficits that deepen with every day the system takes to restart.

The first cargo matters more than the final agreement. Three patterns emerge from the timing arithmetic that reshape how we understand energy security when every day counts.

The India Reality: When Sophisticated Planning Hits Timing Arithmetic

Under current restart timelines and modelled assumptions, India moves into a position where remaining oil reserves, though not yet exhausted, are unlikely to outlast the time until meaningful volumes begin moving through Hormuz and reaching port.

This doesn’t mean shortage—it means stress. India continues day-to-day sourcing from Russian crude, Atlantic Basin arbitrage, and emergency measures, but each option carries economic, diplomatic, or domestic costs that extend far beyond energy markets.

India doesn’t run out of oil. It is stressed because it runs out of time. With around a week of effective petroleum cover on modelled assumptions—crude and products combined—against a roughly two week timeline for supply to move, arrive, and be usable once flows resume, India falls short even in a best-case restart.

And beneath that, the system is even more fragile. LPG (the cooking fuel for 1.4 billion people) tightens earlier and harder, exposing a different kind of vulnerability. Industrial and commercial allocations have tightened, household delivery times extended, and LPG-reliant sectors reported disruption. Unlike crude, LPG cannot easily be substituted or rerouted at scale.

Emergency procurement from the US and Russia, alongside naval deployments, has partially offset supply loss, but at significant cost. The buffer was bought through diplomacy and deployment.

Announcements do not change that underlying timing constraint.

Europe’s Product Problem

IEA Executive Director Fatih Birol confirmed this week exactly what Post 2 predicted: “Europe has only six weeks of jet fuel left before shortages will hit because of the Iran war.” KLM announced 160 flight cancellations due to “rising kerosene costs”, but this not just a price shock. It’s a timing shock, and the two are now tightly linked.

In practice, insurance, not naval announcements, determines when oil flows. While EU crude exposure to Hormuz runs only 6-8%, refined products tell a different story. Approximately 42-50% of European jet fuel imports and over 25% of diesel imports normally transit the strait. No IEA strategic reserve requirements exist for refined products: emergency releases are crude-denominated and cannot solve jet fuel shortages that hit on a different timeline.

The European pattern under the Day 49 restart scenario—where the Strait is declared open but commercial flows have not yet resumed, and timelines include insurance and voyage delays—is clear: crude buffers may bridge, but refined oil products fail first. Those are the ones consumers care about the most.

Germany shows this starkly. On an aggregate basis (total petroleum stocks across crude and refined products) it still bridges with just over three weeks of margin. But on jet fuel, that margin collapses to just three days. That gap between total fuel resilience and the point at which jet fuel becomes the constraint, is where the system actually breaks. Even a short delay in restarting flows would eliminate the remaining jet fuel margin, and turn a narrow bridge into a deficit.

France demonstrates a different vulnerability: technical success that remains fragile. It bridges with around 21 days of aggregate margin, but like Germany, that collapses to three days on jet fuel. One additional week of insurance hesitation flips the verdict. This is bridging, but barely.

The UK demonstrates how products constraints hit differently than crude constraints. Despite North Sea crude production that could offset most Hormuz supply losses, the country's dependence on Gulf-sourced jet fuel creates systematic shortfalls extending more than a month before relief arrives. Domestic crude production cannot easily substitute for specialized refined products—the constraint that actually binds

Italy shows how this plays out in practice. On paper, the system still bridges—both at the aggregate level and on jet fuel. But at the point of delivery, it is already breaking. In early April, airports including Milan Linate, Venice, Treviso and Bologna began imposing refuelling caps as jet fuel tightened—not because the country has run out, but because the system that moves fuel to the aircraft has.

The country hasn’t run out.

Parts of the system already have.

That distinction matters. Jet fuel is not a central pool. It is a chain, from refinery to depot to pipeline or truck to airport tank to aircraft. When supply is disrupted, that chain breaks unevenly. National stocks can hold while airport tanks run dry.

A country that bridges at the national level can still fail at the point of use. The timing calculation tells you whether a country runs out in aggregate. It does not tell you where the system fails first.

This is a structural shift. Strategic reserves still largely designed around crude are not sufficient on their own. Product availability and distribution, not crude supply, are now the binding constraints.

When Margins Collapse: The Regional Cascade

South Korea retains a small margin. But once exports are included, that margin collapses to just days—forcing a choice between maintaining regional supply and preserving domestic security.

That decision cascades. South Korea normally exports refined products to Australia, Japan, and Southeast Asia. If Seoul restricts those exports—the logical response to tight margins—Australia’s diesel shortfall deepens. The Philippines and Thailand face wider gaps. One country’s prudent policy becomes another’s crisis.

This isn’t theoretical interdependence. The countries most exposed aren’t spread randomly—they’re connected through refining networks that amplify rather than absorb shocks.

Australia’s position demonstrates the difference between having energy resources and having energy security. The country exports more energy than any nation except Russia and Saudi Arabia. It also imports most of its transport fuels through Asian refining chains it cannot control.

Australia’s deficit isn’t just about Australian policy. It depends on decisions made in Seoul, Tehran, and London insurance markets. Energy abundance without energy security becomes the defining characteristic of countries that assumed supply chain reliability rather than building it.

Why Timing Trumps Everything Else

The analysis focuses on the right question: when does your country’s first relief cargo arrive? Two factors that dominate commentary about Hormuz reopening turn out to be noise for that specific timing question.

Production restart doesn’t delay the first cargo. Gulf producers curtailed output for 49 days, and ramping back to full capacity takes months. Kuwait’s CEO has said three to four months, while the IEA projects pre-conflict production levels only in late 2026. It matters for sustained recovery. It does not delay the first cargo.

The key question asks when your country’s first tanker loads and sails, and that is answered by storage availability, not production capacity. The IEA April OMR confirms Gulf export losses of over 13 million barrels per day against bypass route flows of approximately 7.2 million barrels per day—implying roughly 12-13 million barrels per day accumulating onshore for 49 days. Terminal storage at Ras Tanura, Fujairah, and Juaymah has been at or near capacity for weeks. The first cargoes are likely to draw heavily from accumulated onshore inventories—though not all of that stock is immediately exportable.

Exit queue clearance is the same constraint as insurance lag, not additional. With approximately 120 loaded tankers trapped inside the Gulf as of Day 40, it appears that your shipment must wait for the queue to clear. It does not—those 120 tankers are distributed across every destination. The first India-bound vessel does not wait for all 120 to move.

What the queue does create is a specific shape of recovery: a front-loaded surge as stored inventory and queued tankers hit markets simultaneously, a plateau as that inventory depletes faster than restarted production refills it, then a gradual climb.

The roughly week-long insurance lag—potentially longer if early transits fail to build confidence—brackets the realistic range. Everything else is noise.

The Two Scenarios That Determine Everything

The gap between what was announced and what was actually confirmed is what creates the two futures.

Day 49 now presents two distinct paths.

Scenario A assumes the US version proves correct: Iran’s concessions hold, transit becomes predictable, and insurance markets normalise within weeks. In that world, the crisis becomes a temporary shock. Strategic reserves function broadly as intended, and countries return to planning around stable voyage times and restored security guarantees.

Scenario B assumes Iran’s position remains definitive: access stays conditional, tied to ceasefire dynamics and enforced through coordination with the IRGC. In that world, disruption becomes structural. Iran determines who can pass — and may begin to price that access. Insurance markets determine whether energy flows move, and at what cost. Transit remains contingent, and countries must plan for repeated interruptions rather than a single event.

The implications are immediate.

Countries with scale and flexibility — Japan and China— remain resilient under either scenario. Countries with tight buffers or complex supply chains do not.

India’s timing deficit, Britain’s jet fuel exposure, and Australia’s reliance on external refining chains leave them vulnerable—not to the blockade itself, but to the gap between political announcement and operational reality.

The system does not fail evenly. It fails at the most constrained links.

At the start of a restart, the first cargo matters more than any statement. Over time, the agreement determines the cost — and who pays it.

What Day 49 Actually Means

Yesterday’s announcements—ceasefire declaration and Iran’s commitment to reopen the strait ‘completely’ to commercial shipping—didn’t end the Hormuz crisis. They began the timing test. The question isn’t whether the strait is theoretically open but whether relief arrives before buffers exhaust.

Commercial flows will eventually test Iran’s confirmed position: conditional passage with IRGC oversight during the ceasefire period. That test will establish operational reality over political claims. Insurance markets will price accordingly. Countries dependent on Gulf imports must plan around verified access rather than announced intentions.

The countries adapting fastest to timing constraints gain relative advantage. Japan’s massive reserves and China’s selective passage arrangements provide temporal cushion regardless of which version of Iranian cooperation proves accurate. Countries assuming simple restoration face systematic disadvantage that compounds with each lost day.

The mathematical countdown continues regardless of political statements. India’s timing gap widens daily. South Korea approaches the point where export restriction decisions become unavoidable. Australia’s cascade faces stress whenever any operational link encounters delay.

When ships actually start flowing again—and the insurance verification period begins—the timing arithmetic will reveal who has energy power and who’s been borrowing it. The framework matters because the assumptions underlying recovery planning determine strategic choices that extend far beyond this crisis.

This is the shift that matters. Energy security today is no longer defined only by whether supply exists. It is defined by whether it arrives in time.

Fast forward to a more electrified future, and timing matters less again: with clean electrified power systems, the power comes from home.

For readers interested in the underlying assumptions and methodology, see below.

Methodological Note

The bridge analysis combines IEA stock data, shipping flows, and insurance lag assumptions to estimate whether countries can outlast the time required for supply to return.

Bridge Analysis Foundation

The core question asks whether remaining strategic reserves exceed the time needed to bring relief: country buffer minus (voyage time plus insurance restart lag). Positive results suggest successful bridging; negative results indicate potential deficit periods before relief arrives.

Stock Coverage Methodology

Base figures use IEA standard methodology combining government-controlled stocks and obligated industry stocks while excluding free commercial inventory above mandatory minimums. Post-IEA emergency release figures reflect the 400 million barrel coordinated release completed in March—the largest in IEA history.

Insurance Lag Integration

The roughly week-long commercial restart lag reflects realistic insurance market restoration requirements. Lloyd’s war risk premiums jumped 100-fold during the crisis, from 0.05% to over 5% of vessel value. Coverage restoration requires demonstrated Iranian cooperation, established IRGC coordination protocols, and actuarial confidence in new transit arrangements. This lag is built into the primary analysis, not treated as sensitivity, though the actual duration could vary significantly depending on early transit outcomes.

Hormuz Exposure Weighting

Not all crude imports transit Hormuz. This analysis applies exposure weightings based on actual shipping flow data from Kpler and Vortexa: India 40% crude exposure, Japan 77% exposure, UK 6-8% crude exposure but significant product exposure through refined imports. Product exposure figures use seaborne flow data because no strategic reserve exists specifically for refined products.

Voyage Time Assumptions

Direct shipping times from Gulf terminals include: India (Vadinar) 9 days, South Korea (Ulsan) 19 days, Japan (Chiba) 20 days, UK (Southampton via Cape route) 38 days. Australia product figures use 34-day cascade timing: 11 days Gulf-to-Singapore crude delivery, 6 days refinery processing, 17 days Singapore-to-Sydney product delivery.

Refining Hub Constraints

South Korea calculations separate domestic consumption from total refinery feedstock requirements, with the latter representing the binding constraint given Seoul’s role as regional refining hub supplying refined products to Australia, Japan, and Southeast Asia. Export restriction decisions cascade through regional supply chains.

Production vs. Inventory Distinction

Gulf producer output was curtailed for 49 days, but first cargo availability depends on terminal storage, not production capacity. IEA data confirms 12-13 million barrels daily accumulated onshore during the disruption. First cargoes draw from existing inventory while production ramp determines sustained recovery phases.

Data Sources and Limitations

Analysis combines IEA Oil Market Reports, official government stock disclosures, Kpler and Vortexa shipping databases, and Bloomberg terminal tracking. These figures are directional—what matters is the gap between buffer and time. China’s 59-day estimate—measured against total petroleum demand—incorporates selective passage arrangements and estimated strategic reserve levels given limited official disclosure. The IEA reports 120 days on a net seaborne crude basis, which uses a narrower denominator; this analysis applies total demand for consistency across all countries.

New here? The first post in this series — When the Buffers Run Out — asked one question: when the world’s most critical energy corridor closes, how long before countries start running out of oil? The analysis was directional and worst-case. In a crisis that keeps escalating, worst cases have a way of becoming the baseline. This post builds on that analysis. Start there if you haven’t.

One social media post. One naval announcement. And a question most of the coverage is not asking: if the United States blockades the Strait of Hormuz, what happens to the one supply route that was actually working?

The Strait of Hormuz has been effectively closed for 44 days — the waterway through which roughly a fifth of the world’s oil normally flows. But not entirely. A trickle has kept moving: Iranian crude to China under a sanctions waiver, selective passage for some vessels, residual flows under negotiated terms. A temporary ceasefire last week briefly steadied markets. It did not reopen the strait. Peace talks in Pakistan collapsed on Sunday, and the United States announced a naval blockade within hours. The ceasefire formally remains in place. The blockade is already being enforced. What remains of that trickle is now at risk.

The most important seaborne supply-side check on the crisis — Saudi Arabia’s East-West pipeline, rerouting approximately 3.8 to 4.4 million barrels per day to its Red Sea port of Yanbu — is now exposed to a second threat. Those tankers loading at Yanbu exit south through the Bab el-Mandeb Strait. Straight through Houthi territory.

The new blockade does three things. It removes the approximately 1.7 million barrels per day of Iranian crude still flowing through the strait. It ends any residual managed passage that had begun to move under negotiated terms. And if it triggers a Houthi response at Bab el-Mandeb, it threatens the Saudi rerouting that has been the primary check on the disruption’s scale. Each accelerates physical shortage and price pressure in a market that has no slack left. The effective net daily supply loss does not stay at 11 million barrels per day. It could move back toward 15 million — the gross disruption figure IEA Director Fatih Birol cited early on in the war, which until this weekend looked like a ceiling, not a destination.

What 44 Days Has Already Done

All figures in this series are directional and illustrative, drawing on primary sources where available — IEA March Oil Market Report for crude exposure by country; Kpler and Vortexa for refined product flows. They indicate orders of magnitude and relative vulnerability, not precise forecasts.

The analysis published on 5 April (Day 36 of the disruption) showed when countries would face physical shortage from the crisis based on total stock cover. That was the right directional first question. The more precise one is: how much of each country’s supply actually transits Hormuz? Aggregate stock cover is a ceiling. Hormuz exposure tells you how much of that ceiling is actually at risk.

Europe looks less alarming on crude than the Day 36 numbers implied, and more alarming on oil products. UK and EU crude exposure to Hormuz is approximately 6 to 8%. On crude alone, European buffers are not the acute story. The acute story is jet fuel: approximately 42 to 50% of European seaborne jet fuel imports transit Hormuz (Kpler). Diesel exceeds 25% of seaborne diesel imports (Vortexa). No IEA reserve is designed specifically for refined oil products — the emergency system is overwhelmingly crude-based. Crude in storage does not fix a jet fuel shortage. Jet fuel prices have increased by approximately 95% since the conflict began (Euronews). The product market is already under direct structural pressure.

Asia is where Hormuz exposure is both severe and differentiated. Japan sources approximately 77% of its crude via Hormuz but holds sufficient strategic stocks. The Prime Minister’s public statement that supplies last into next year is directionally consistent with the analysis. South Korea is not just an oil consumer, it is one of Asia’s largest refining hubs, processing Gulf crude and exporting refined products across the region. That creates two distinct readings of its buffer.

On domestic supply (i.e. what South Korean households and businesses actually consume) the buffer at Day 44 is approximately 40 days once adjusted for Hormuz exposure. Meaningful. On the binding constraint — the total crude feedstock South Korea’s refineries need to keep running, including the product flows they supply to Australia, Japan and South-East Asia — the figure drops to approximately 32 days. And when Hormuz reopens, shipping time alone can eat close to 20 days of that buffer before the first cargo arrives. A disruption to South Korean refining is not a domestic issue. It is a regional one.

Australia illustrates the cascade directly. It produces crude but exports almost all of it, with domestic refining covers only 17% of demand. The diesel and jet fuel that keep it running come primarily from South Korean and Singaporean refineries processing Gulf crude. Based on data from IEEFA and Kpler, Australia’s effective product exposure to Hormuz disruption runs to approximately 42 to 55%. This is not because tankers travel directly from the Gulf to Sydney, but because the refineries that supply it do. As of early April, diesel stocks stand at approximately 29 days. The relief timeline is considerably longer than a direct voyage would suggest, once refining lag is factored in. Post 3 carries the full analysis.

India’s crude Hormuz exposure is approximately 40% — lower than the Day 36 post aggregate number implied, because India has deliberately diversified toward Russian and West African crude since 2020. India’s crude buffer is therefore longer than the headline stock cover figure suggests. The acute Indian story hits households and is about LPG: approximately 90% Gulf-dependent, and the cooking fuel shortage arrived ahead of the crude shortage.

The Philippines — approximately 98% Hormuz-dependent, no strategic reserve and no IEA access to fall back on — faces a structural supply shortfall from its Hormuz exposure, sustained only through emergency replenishment from alternative sources. It was the first country in the world to declare a national energy emergency, on Day 24. Vietnam, at approximately 85% crude exposure and minimal reserve, was already at or near physical shortage conditions at the time of last week’s ceasefire announcement.

China, which holds an estimated 120 days of cover, is managing its position by negotiating access with Iran for selective passage for Chinese-flagged vessels since late March. It is not a passive victim of this crisis. Those arrangements are now directly targeted by the US blockade announcement.

The countries most exposed to an extended disruption once a Hormuz weighting is applied rather than aggregate stock cover, are Vietnam, the Philippines, South Korea, India on LPG specifically, and Europe on jet fuel and diesel. The aggregate numbers from Day 36 were directional, but incomplete. The Hormuz-adjusted picture sharpens them, as does another eight days of drawdown on buffers that were already under pressure.

If Iran’s Barrels Are Also Removed

Since the strait closed, the effective net loss to global oil supply so far has been running at approximately 11 million barrels per day — after all interventions: Saudi Arabia’s Yanbu reroute and UAE’s crude diversions, sanctioned oil from floating storage, and partial demand destruction. It was already the largest peacetime supply disruption ever recorded: more than double the 1973 Arab oil embargo at a fraction of the lead time.

It was not the worst case. And it was not complete.

One reason the loss stayed at 11 million rather than rising to 15 million is that Iran never fully stopped exporting. Bloomberg tanker tracking data shows Iran shipped approximately 1.7 million barrels per day of crude and condensate in March — at pre-war levels — primarily to China. A US sanctions waiver, quietly issued last month, authorised continued Iranian crude sales as a pressure valve for Asian markets.

That waiver expires April 19. The blockade was announced before it expires.

If the blockade takes effect and Iranian exports stop, effective net loss rises from approximately 11 million barrels per day to approximately 12.7 million — call it 13 million. That figure lands between the current baseline and the IEA’s 15 million upper bound, which until this weekend was treated as a stress test rather than a base case. The buffer figures that made India, South Korea, and the European countries look manageable at Day 36 were built on 11 million barrels per day. At 13 million, those margins compress. The countries already at the margin move into deficit sooner.

There is also a dimension the blockade announcement does not resolve: China is the primary buyer of Iranian crude. A blockade that intercepts Chinese-chartered or Chinese-flagged vessels is not an Iran policy. It is a direct US-China maritime confrontation. A blockade that exempts them has large holes in it and markets will identify those holes within days of enforcement beginning. These are not close to the same price outcome. Former US Ambassador to Saudi Arabia Michael Ratney, speaking on Bloomberg, flagged the China dimension as the blockade's most consequential open question: whether the US Navy would intercept Chinese-flagged vessels, and what that would mean for US-China relations.

The Houthi Escalation: Why Yanbu Is Now The Exposure

This is the dimension that current coverage is most seriously underweighting.

When Hormuz closed, Saudi Arabia rerouted approximately 3.8 to 4.4 million barrels per day through its East-West pipeline to the Red Sea port of Yanbu. For Asian buyers — China, India, South Korea, Thailand and Pakistan — this became the primary alternative supply corridor for crude they could no longer receive through Hormuz. The IEA named it the most significant supply-side response of the crisis. It is the single most important reason the net supply loss has stayed at approximately 11 million barrels per day rather than 15 million.

The Very Large Crude Carriers loading at Yanbu are too large for the Suez Canal. The Asian-bound cargoes exit south through the Bab el-Mandeb Strait — at the southern end of the Red Sea — into the Gulf of Aden and onward to Asia. The Houthis spent 2023 and 2024 demonstrating their capacity to disrupt exactly this route, reducing flows through Bab el-Mandeb materially during their campaign. Bloomberg Economics and CSIS have both flagged a return to Bab el-Mandeb as a material escalation risk in response to a naval blockade. Iran has already struck the Yanbu refinery, a joint Aramco-ExxonMobil facility, once during this conflict. The idea that the Yanbu bypass remains outside the war is no longer a safe assumption.

If Houthis activate at Bab el-Mandeb, the effect is both additive and structural. It adds new disruption while removing the mitigation that has been limiting it. The roughly 4 million barrels per day transiting the strait includes the Saudi crude rerouted from Hormuz, now a critical supply lifeline for the Asian economies most exposed to the closure, as well as Russian crude flows to Asia that remained stable through the earlier Houthi campaign.

The mitigation and the threat converge on the same chokepoint. The volumes are smaller than Hormuz but disproportionately important, because they are the volumes replacing it.

This would tighten not just crude availability, but the already strained product markets — particularly jet fuel and diesel — where substitution is limited and no IEA emergency mechanism exists specifically for refined products.

These pathways are conditional, not additive. 11 million barrels is approximately where we are. 13 million is where we go if the blockade removes Iranian flows. 15 million is where we go if the blockade then triggers Houthi activation and Yanbu is lost. Each step requires a further decision, whether in Washington, in Tehran, or in Sanaa. The directional logic is clear, even if timing is not. None of them is inevitable. All of them are now live.

The countries most exposed to a Yanbu disruption are those currently receiving Saudi crude via the Red Sea route: China, India, South Korea, Thailand, and Pakistan. The same countries. A second chokepoint failure does not spread the disruption across a different set of nations. It concentrates it further.

The Number Nobody Is Pricing: Shipping Days

None of the above yet accounts for the most basic physical constraint of all.

When Hormuz eventually reopens, whatever the terms, the barrels do not arrive the next morning. They arrive when the ships do. First relief cargo reaches India in approximately 9 days from Hormuz reopening. South Korea: 17 to 18 days. Japan: 20 days. The UK and Europe, routed around the Cape of Good Hope: 38 to 39 days.

Every additional day of closure — from blockade enforcement, from ceasefire failure, from Houthi escalation, from the insurance market’s own lag in certifying routes as safe — extends those voyage gaps. The buffer consumed today cannot be restored overnight. The physical system does not respond to announcements. It responds to cargo arrivals.

Markets priced last Tuesday’s ceasefire as if oil would flow freely within days. Dated Brent — the price for actual physical delivery — closed nearly $30 above the June futures price on the same afternoon. The paper market priced the announcement. The physical market priced the reality.

Today, oil rose more than 4% on the blockade announcement, before pulling back after Trump said Iran had reached out to discuss a deal. The paper market moved on both signals. The physical market will, again, take longer to agree.

Neither move fully reflects the underlying physical supply picture outlined in Post 1. At around $100 a barrel, markets are still not fully pricing 44 days of sustained drawdown, or what comes next.

The deficit already locked into the global supply chain was substantial before Sunday night. It is larger now.

There is a longer frame here, beyond the immediate fuel shortage arithmetic.

This is not just a supply shock. It is a stress test of the system the energy transition is replacing.

Moments like this do not derail the transition — they reshape it. The immediate response is not just clean energy, but whatever is available: coal, diesel, emergency cargoes from wherever they can be sourced. But the strategic response moves in the opposite direction. Energy security, affordability and decarbonisation stop being separate goals and collapse into one. Electrification, domestic generation and supply chain control are no longer climate policies. They are geopolitical ones.

The next post starts where this one ends. Not with the closure, but with what comes after it. Country by country. Product by product.

The Power of Power publishes at the intersection of energy, geopolitics, and the contest for global advantage.

Footnote: stock cover figures in this series use IEA standard methodology — government stocks plus obligated industry stocks — and exclude free commercial inventory held above the mandatory minimum. Hormuz-adjusted figures apply a weighting to each country’s buffer based on their actual crude and product exposure to Hormuz flows, drawn from IEA, Kpler and Vortexa primary data. South Korea’s two figures reflect two distinct analytical questions: domestic consumption cover versus total refinery feedstock requirement, the latter being the binding constraint given South Korea’s role as a regional refining hub. Product exposure figures — particularly for European jet fuel and diesel — draw on Kpler and Vortexa seaborne flow data rather than IEA obligated reserve figures, because no strategic reserve exists specifically for refined products; the IEA emergency mechanism is crude-denominated. Full methodology and more detailed workings and analysis will be published with Post 3.

This is not an oil shock. It is a broad energy crisis — the largest supply disruption in the history of global energy markets.

Thirty days ago, the question was whether the Hormuz disruption would last. It has. Now the question changes: what happens when the buffers run out?

The scale is worth stating plainly. Roughly 10 million barrels per day of crude (close to 10 percent of global oil supply) has been removed from accessible flows. Nearly one-fifth of global LNG trade, over 110 bcm a year, is now effectively disrupted, with no alternative export routes to market. Qatar’s Ras Laffan, the largest LNG facility in the world, was attacked on 2 March; two of its fourteen trains are gone for three to five years. Over 3 million barrels per day of refined products — jet fuel, diesel, gasoline — and a further 1.5 million barrels per day of LPG are blocked at source.

The system’s shock absorbers were deployed immediately. Strategic oil reserves were tapped: 400 million barrels, around a third of IEA member governments’ emergency stocks. Alternative routes were taken where physically possible. Price rationed demand where it could.

For the first 30 days, it worked. Not painlessly, since the crisis began Brent has traded in a volatile $90–$115 range. But it worked.

As of today — day 36 — the first phase is over. Now the arithmetic takes over.

Buffers deplete. Once they do, the shortage gets managed by a sharper price, but it also becomes rationed by scarcity. This is the phase we are entering.

The Buffer Maths

What follows is my own back-of-the-envelope arithmetic, built from IEA, Bloomberg and other published data. It looks only at oil buffers as gas has no equivalent backstop. It’s meant to be illustrative rather than precise, and I’d genuinely welcome challenge or refinement from anyone with better data or a different read. To my knowledge, no official agency has yet published a consolidated country-level stock depletion view like this. If that changes, I’ll update.

The question that matters most is not when prices spike. It is when the world crosses from an expensive energy crisis into a physical shortage — the point at which oil is simply not available at any price a buyer can afford to pay. Where fuel terminals run dry. Where rationing is no longer a policy choice but a physical necessity. The market cannot clear. Governments cannot print oil.

The net supply loss

Hormuz carried roughly 20 mb/d before the disruption. Bypass pipelines through Saudi Arabia and the UAE cover 3.5–5.5 mb/d. Temporary sanctions waivers on Russian and Iranian oil already at sea provided additional cushioning, but the Russian waiver expires on April 11, and the Iranian waiver expires April 19. As those cushions run out, Bloomberg estimates the effective current net loss at around 11 mb/d, rising toward the 15 mb/d that IEA Director Fatih Birol cites as the combined loss of crude and products.

Let’s model these two scenarios as an illustration of what could ensue:

Scenario A — ~11 mb/d: the current effective loss, after all interventions, consistent with Bloomberg’s estimate.

Scenario B — 15 mb/d: IEA Director Fatih Birol’s stated figure — the level we are heading toward as the waivers expire.

Upfront: this is an illustrative scenario under sustained disruption at current rates. Demand rationing and alternative supply are already underway — they can slow depletion, but not fully close the gap. Demand will adjust, and some spare capacity exists, but both are constrained — by timing, logistics, and the physical limits of the system. A near-term diplomatic or military resolution changes everything.

What the emergency 400mb release actually covers

Four hundred million barrels sounds like a lot. It is, as the largest emergency release in the IEA’s fifty-year history. But volume is not the binding constraint. Daily drawdown rate is.

Strategic reserves cannot be released overnight. They flow through pipelines, terminals and distribution networks that have physical limits. The US’ contribution of 172 million barrels, the largest single share, releases at roughly 1.4 mb/d over a 120-day period. Across all IEA members, the practical combined maximum release rate is around 2–2.5 mb/d.

Against a daily supply loss of 11–15 mb/d, that covers only 17–23% of the gap. It slows depletion. It does not close it.

The remaining 77–83% of daily shortfall has to be absorbed elsewhere — and that somewhere else is drawdowns in consuming country stocks, alongside some demand destruction and limited supply substitution. That is what the table below shows.

———

The sequential cliff edges

Stock cover figures calculated post-IEA release, from 5 April 2026 — day 37 of the disruption.

A few footnotes:

The US — special case. Post-release the Strategic Petroleum Reserve (SPR) holds 243 million barrels (its lowest level since 1983). But producing 13.6 mb/d domestically, the US faces price pain, not physical shortage.

Canada — the only G7 nation with no strategic petroleum reserve, faces no physical shortage either. Energy self-sufficiency is its own kind of power.

India: The ~45 days includes commercial stocks and cargoes at sea. The SPR alone covers only ~9.5 days. In practice, India's effective crude buffer is longer: only around 33% of crude imports transited Hormuz before the crisis, down from 61% in 2020. But around 90% of LPG imports transit the strait, exposing household energy supply.

Stock cover figures use IEA methodology — total reserves against total net import needs. Countries with significant non-Hormuz supply have greater effective buffer than these figures suggest; those with near-total Hormuz dependence face a reality closer to what the figures show. In practice, adjustment will be uneven: stocks are not drawn down at the same rate, refined product shortages tend to emerge before crude is exhausted, and supply is allocated rather than shared.

What this does not capture: China’s estimated 1.3 billion barrels outside IEA coordination could extend its own timeline materially, but those barrels are not easily available to others. Further IEA releases are under consideration. A diplomatic or military resolution would materially alter the outlook. This is the base case.

The system does not have one cliff edge. It has a sequence of cliff edges, by country, by reserve depth, by import dependency.

The most vulnerable have already crossed theirs. The question is how fast the cascade moves up the chain. That is not a forecast. It is arithmetic.

Even in the extreme, if IEA members released their entire combined government and industry stocks (i.e. roughly 1.4 billion barrels), the binding constraint is not volume, but flow. In practice, they would not. But even if they did, the answer is the same.

The infrastructure required to move oil from storage to market, like pipelines, terminals, and distribution networks, caps the release rate at around 2–2.5 mb/d across all IEA members combined. At that rate, even a full drawdown would take well over a year. Stress-tested against the Hormuz gap directly rather than the release rate, that same full IEA buffer is volumetrically overwhelmed by mid-August at 11 mb/d, early July at 15 mb/d. Either way, far longer than most countries can sustain before encountering physical shortage.

More reserves do not mean faster relief. They mean the same flow, sustained for longer.

The IEA’s Executive Director has called this the greatest threat to global energy security in history. He is not given to exaggeration.

What That Scale Actually Means

There is an uncomfortable parallel with February 2020. COVID was present, spreading, already devastating parts of Asia. Markets looked broadly fine. Then the full implications landed, all at once.

Senior energy analysts are now drawing the same comparison explicitly. We may be in the February 2020 moment of this crisis. The cliff is there. The market hasn’t looked down yet.

The reference point for understanding the scale is 2020. Not because governments pulled the demand lever, but because they didn’t. The largest demand destruction event in oil market history was a byproduct of government pandemic policy, not energy policy. The collapse (roughly 9 million barrels per day for the full year, peaking at 25 to 29 million barrels per day in April) was a consequence. An accidental stress test of what demand destruction at scale actually looks like.

It required grounding much of aviation, driving restrictions, and confining billions of people to their homes.

A sustained 10–15 million barrel-a-day supply disruption sits in the same order of magnitude, except this time no government chose it, no virus produced it, and no fiscal transfer can substitute for energy that simply isn’t moving.

When buffers run out, demand has to fall. Part happens through the market — prices rise, activity slows, consumption is forced down. But price alone cannot close a gap this large. Governments step in too, to dampen demand by rationing, restricting, curtailing industrial output, and accelerating electrification as emergency demand response. Either way, the result is the same. Less gets used.

That lever is already being pulled in Asia. The Philippines and Pakistan have moved to four-day work weeks. Thailand has ordered civil servants to work from home and capped diesel prices. Sri Lanka has introduced fuel rationing — 15 litres per week per driver. Bangladesh has shut universities early.

The question is whether it is anywhere near large enough.

Managed demand destruction at 2 or 3 million barrels per day is policy. Forced demand destruction at 10–15 million barrels per day is the system pricing economies into contraction, country by country, industry by industry, at a pace no government can administer or cushion.

History offers a clear pattern: sustained physical fuel shortages do not produce slowdowns. They can bring sharp economic contractions.

The 1973 Arab oil embargo (roughly 4 mb/d removed) was followed by recession across the US, Western Europe and Japan. The 1979 Iranian revolution (a further ~4 mb/d disruption) brought a second wave of economic damage.

Both disruptions were significantly smaller than what is now being modelled. The global economy has no modern peacetime precedent for a supply loss of this magnitude.

The last time the world absorbed demand destruction of this magnitude, it shut itself down to get there.

This time, no one chose it. And the lever, pulled hard enough, breaks the hand holding it.

Why Markets Haven’t Caught Up

So why don’t prices reflect it?

Since the crisis began, Brent has held in a $90–$115 range. That is lower, remarkably, than the $125 it briefly touched after Russia’s invasion of Ukraine, when Russia never stopped producing a single barrel. The market ran up on the risk of losing supply that never actually left. Today, supply has genuinely been stranded — and prices are lower. For now.

The explanation is not irrational. It is a failure of historical memory. Few if any traders active in today's markets have experienced a genuine physical disruption of this magnitude. What traders know how to price is risk. In 2022 they priced Russian supply risk, got it wrong, and lost money. The lesson: don’t overreact to geopolitical risk.

That lesson is now working against them. This is not a risk environment. It is a disruption environment. The supply is not at risk. It is gone.

Markets in March still believed Trump would end this quickly. That belief creates a circular trap. The president has less urgency to act because markets aren’t screaming. Markets aren’t screaming because they think he will act. In the meantime, in Asia airlines are beginning to cancel flights due to jet fuel shortages, not because it is too expensive, but because it is not there.

The most important price signal is not Brent. It is the physical price to deliver a barrel to Asia, which is already $130–$150. The paper price is a financial signal. The physical price is the real one. Most of the world’s major trading desks sit in cities that are not yet feeling physical pain. That insulation is part of why the market hasn’t caught up.

When the market finally looks down, the adjustment will not be gradual.

The Cascade

Even framing this as an oil and gas shock understates it.

The LNG disruption alone is without precedent. Over 110 bcm of LNG transited Hormuz in 2025 — almost one-fifth of global trade, with no alternative export routes. QatarEnergy has declared force majeure. Two of its fourteen trains are damaged beyond repair for three to five years. Restarting the rest from cold shut-in will take eight weeks minimum after the war ends, according to industry estimates.

Gas prices follow. Electricity prices follow gas. Europe entered this crisis with gas storage at just 30% capacity, already structurally exposed before a single Qatari cargo was stranded. Economists are warning of technical recession in energy-intensive economies if the disruption persists through the summer refill season.

The chain does not stop at energy. Gas feedstocks make fertiliser. Fertiliser grows food. More than 30% of global urea trade moves through Hormuz. Urea prices have risen 50% since the war began, hitting during the Northern Hemisphere spring planting season.

The food price consequences will extend into 2027.

The energy shock has a hardware problem embedded in it: refineries built to process heavy Gulf crude cannot simply switch to alternative grades. Atlantic Basin crude is lighter and sweeter, and chemically incompatible with refinery configurations across much of Asia. The modifications are costly and the lead times are months.

Who Gets Hurt, and How Badly

The disruption does not land equally.

The United States — producing over 13 million barrels per day domestically, with deep gas production and growing LNG export capacity — faces significant price pain but not physical constraint. It remains the only actor capable of reopening the strait, and every week that passes without intervention is itself a geopolitical statement: about leverage, about alliance dependency, about who needs whom.

Asia bears the weight. Japan holds government and industry stocks equivalent to roughly 200 days of net imports, making it the deepest strategic reserve relative to consumption of any major importer in the world. It was a deliberate policy choice shaped by the trauma of 1973. South Korea, the world’s fourth-largest LNG importer, has considerably less depth. For both, reserves are a bridge, not a solution. At 60 days, even Japan’s stockpile starts to matter less than what comes after it.

China’s position is more complex, and more consequential. The world’s largest oil importer, is exposed on both Gulf oil and LNG channels. But it has diversified import routes with the ESPO pipeline and the Central Asian corridors as a deliberate Hormuz hedge. More structurally, as the world’s first true Electrostate, it has reorganised its economy around electrification, can deploy new solar, wind and batteries faster than any other nation, and can redispatch baseload coal-fired generation when oil and gas tighten. That floor matters.

China may be one of the few major importers that does not need this crisis resolved urgently. That structural absence of urgency can translate into a degree of influence over how, when, and on whose terms the strait eventually reopens, whether deliberate or not. With 1.3 billion barrels of stocks outside IEA coordination, it also holds something rarer still: the ability to offer selective relief to desperate smaller importers, and the political capital that comes with it.

Europe is less dependent on Middle Eastern oil than it was a decade ago, particularly crude, but its gas exposure, elevated since the Ukraine war and now compounded by the loss of Qatari LNG, makes it acutely vulnerable. Storage entered this crisis at just 30% capacity. European gas benchmark prices have nearly doubled. The ECB has already revised growth forecasts downward. And several major EU economies face elections in 2026. Energy rationing in Germany or France is not just an economic problem. It is a government-ending one.

The UK sits in a more exposed position still with limited storage, inflation already heading toward 5%, and outside the EU’s formal solidarity mechanisms that give continental neighbours at least the option of coordinated response. Its crude import base is more resilient, drawing from Norway and the North Sea, but refined product exposure is significant with at least half its jet fuel from the Middle East.

The correct framing for UK and Europe is at this stage in the crisis is: largely insulated on crude supply, genuinely exposed on shortages of specific refined products, and hit by the crude price shock regardless.

When G7 economies start hitting their own cliff edges, and the arithmetic suggests that is weeks away for some, the political pressure on Washington to act will intensify sharply. Whether that pressure translates into coordinated alliance response or fragmented national crisis management is the more uncomfortable question.

Smaller importers across South and Southeast Asia are no longer facing the market as it might become. They are living it now in fuel queues, school closures, and rationing cards.

The Known Risk

This should not be framed as an unforeseeable event. It wasn’t.

The Strait of Hormuz has been identified as the world’s single most critical energy chokepoint for decades. Every serious scenario planning exercise (e.g. IEA, US Department of Energy, the major oil companies) has modelled something like this. Governments knew. The risk sat, clearly labelled, in the literature.

It is not the first time. Europe’s dependence on Russian gas followed the same pattern — warnings published, presented, and set aside. Then 2022 arrived and the bill came due: emergency LNG contracts, energy rationing, deindustrialisation pressure, and a scramble to build infrastructure that should have existed a decade earlier.

Hormuz was not just modelled by energy agencies. Closure scenarios were actively war-gamed by the Pentagon, repeatedly, and over decades. The consistent findings from US Central Command (CENTCOM) pointed to oil price spikes of $150 to $200 per barrel or more under sustained closure, severe economic stress on Asian allies, and a significant risk of alliance fracture, with Japan, South Korea, and India facing impossible choices between economic self-preservation and geopolitical alignment with Washington. The military conclusion was consistent: the US Navy could reopen the strait, but it would take weeks, not days, and come at meaningful cost.

What is playing out may not have been in those scenarios: a prolonged disruption with the United States choosing not to act quickly. An energy superpower with domestic oil and gas self-sufficiency has more strategic patience than many allies — and that patience is itself a form of power. Each week the strait remains closed without US action makes the world’s dependency map clearer.

But patience is not the only factor. Reopening Hormuz by force is not a simple operation. Iran’s drone capabilities which have been developed and deployed at scale since 2019, have significantly extended the threat perimeter well beyond the strait itself. The cost asymmetry is striking: disruption can be sustained at relatively low cost, while the military and economic cost of intervention is significant and real. Each week of non-intervention reflects both strategic calculation and a genuine assessment of operational risk.

What failed was not intelligence. It was political economy. Building genuine resilience — deeper storage, more diverse supply infrastructure, grid flexibility, fuel-switching capacity, alongside deeper security around the chokepoint — was expensive, and the system hadn’t broken yet. The 400 million barrel release has left most IEA members below the 90-day standard they set for themselves. Only Japan appears to still meet it. Reserves were designed for short-duration shocks. Now the bill arrives, again.

Iran has now demonstrated something that changes the strategic calculus permanently: it has shown it can and will close the Strait of Hormuz if backed against a wall. Every energy security framework built on the assumption that Hormuz would remain open must now be rebuilt from scratch.

After Hormuz: What Comes Next

The conventional definition of energy security focused on supply: enough of it, at affordable prices, reliably delivered. That framing shaped policy for half a century.

It is no longer sufficient.

A 60-day Hormuz disruption does not test whether countries have supply. It tests whether they have systems (i.e. the storage depth, grid flexibility, fuel-switching capacity, electrification, and supply chain redundancy) to absorb sustained stress without fracture. The countries that emerge least damaged treated resilience as infrastructure rather than insurance.

But here is the forward problem. The world closing Hormuz in 2026 is still substantially a hydrocarbon world. By the mid-2030s, that changes. The IEA has called it directly: the world is entering the Age of Electricity. Electrification accelerates. Clean energy scales. AI and data centres become a structurally growing load on every major grid.

In that world, strategic vulnerability does not disappear. It migrates toward the Strait of Malacca — the narrow waterway between Malaysia and Indonesia through which roughly 40 percent of global trade already passes.

In an electrified, AI-driven economy, the flows that matter most are no longer primarily fossil fuels. They are semiconductors, the hardware on which AI runs. Battery materials and critical minerals — cobalt, lithium, nickel, copper, graphite — converging on Asia from Australia, Indonesia and the Congo. Rare earth elements (the backbone of electric motors, turbines, and defence systems) overwhelmingly processed in China. Solar panels, EV components, wind turbines manufactured overwhelmingly in China and shipped globally.

The irony is not lost on strategists: the country that manufactures most of what flows through Malacca is also the one with the greatest capacity to influence access to it.

Close Malacca in 2036, and you do not disrupt a fuel supply. You disrupt the physical infrastructure of the electrified economy itself. There is no coordinated strategic battery reserve. There is no SPR equivalent for semiconductors. For critical minerals and rare earths, national stockpiles exist but are modest and fragmented. The electrified economy has nothing comparable to the IEA’s emergency oil reserve system.

But the vulnerability is different. Energy from oil and gas depends on continuous flow. Interrupt supply and the impact is felt quickly, often within weeks.. Clean and electrified energy systems are built: grids, batteries, solar panels, data centres, EVs embed their critical materials into infrastructure that keeps running even when supply tightens. The risk is not an immediate shortage of energy supply, but a slower constraint: a creeping inability to build, expand or replace. That still confers profound economic leverage to those who control access, whether through physical chokepoints, export controls, or technology bans.

The next challenge is ensuring the critical dependencies of the electrified economy don't simply replace one chokepoint with another.

The chokepoints shift. The dependency changes.

Energy security is no longer just about supply.

It is about maintaining power when the system — whatever system — can no longer be managed.

The Power of Power publishes at the intersection of energy, geopolitics, and the contest for global advantage.

Footnote — No official agency has published a consolidated country-level stock depletion analysis of this kind since the disruption began. The methodology follows IEA standard practice: total stocks against total net import needs. More precise calculations would weight each country's cover against its specific Hormuz exposure, but that data is not publicly available in consolidated form. Countries with significant non-Hormuz supply appear more vulnerable than they are; those with high Hormuz dependence face a reality closer to what the figures show. Official agencies will hopefully publish this with fuller data. Until they do, this is a robust public estimate.