The Oil Must Flow: A Global Map of Energy
On 15 May, eleven weeks into the closure of the Strait of Hormuz, I started building a map. After my departure from APMC in mid-August I had the time, and I needed to work on something other than sports, streaming and ad-tech. So I went back to it, and this week it is finished enough to publish.
The Global Energy Map is an open atlas of the world's oil and gas system: reserves, extraction sites, pipelines, refineries, LNG terminals and voyages, storage and ports, from 1990 to 2024. It runs disruption scenarios — close the Strait of Hormuz for crude or for LNG, cut Druzhba, Baku–Tbilisi–Ceyhan or the Caspian Pipeline Consortium line — and shades every importing country by the share of its imports that had been travelling that way. Every number on it traces to a cited public source.
The report that goes with it is The Oil Must Flow, and it is the second of these quarterly reports.
What the map says
On 2024 trade data, 27.9% of the world's recorded crude imports came from Gulf exporters along routes through the strait. By share of their own imports, Pakistan (78%), the Philippines (77%), Japan (73%) and South Korea (62%) depend on it most. By volume it is China, at about 3.5 million barrels a day.
Run the same calculation back through twenty years and the interesting thing is not the total, which barely moves. It is whose dependence it is. The United States falls from 19% Hormuz-dependent in 2005 to 7.9% in 2024. China's at-risk volume more than quadruples. Japan sits at about three-quarters throughout, on a total that shrinks by nearly half. On the gas side the world's exposure genuinely falls — 30% of recorded LNG imports in 2015, 19% in 2024 — while Pakistan goes the other way, to 88% dependent on LNG that comes through the strait.
What happened when the strait actually closed
The map was built during the disruption it models, which makes it a test rather than a thought experiment.
As a first sort of who gets hurt, it holds up. The countries it ranks highest are the ones that acted: Pakistan moved its government to a four-day week, the Philippines declared a year-long energy emergency, Japan released 80 million barrels, South Korea capped fuel prices for the first time in nearly thirty years, Sri Lanka rationed fuel by QR code. The LNG buyers Qatar named in its long-term force majeure — Italy, Belgium, South Korea, China — sit near the top of the map's LNG ranking.
As a guide to how much, it is weak, and the reasons are the interesting part. It holds three things fixed that turned out to be where the action was. Bypass capacity: the map gives Saudi Arabia the 12% of its crude that normally goes west, but by late March the East–West pipeline was reportedly running at its full 7 million barrels a day. Substitution: India's exposure was 1.9 million barrels a day on paper, and its actual imports fell by about 760,000, because it bought Russian crude instead. Stocks: the largest coordinated release in the IEA's history was adding 2.5 million barrels a day by May.
And some of the worst-hit countries are not in the data at all. Bangladesh shut its universities and idled four of five state fertiliser plants over gas, and the trade statistics record almost no LNG reaching it in 2024. Iran, the country the war is about, has been close to invisible in open trade data since 2019 — though the barrels leave a trace: the same dataset records "Malaysian" crude arriving in China rising from 33,000 barrels a day in 2016 to 833,000 in 2024, which is more than Malaysia produces.
That is the honest summary of what a first-order exposure model is for. It tells you who is standing in the road. It does not tell you how fast the car is going.
How it was built
The other half of the report is a record of the build, because I think it is the more useful half for anyone doing this kind of work.
The map was built by one person and five Claude models across two working periods four months apart: Opus 4.7 and Sonnet 4.6 in May, then Fable 5.1, Sonnet 5 and Opus 5 in September. The last four phases — about 23,000 lines added — ran in roughly twenty-two hours.
The decisive step was not a faster model. After six phases of test-driven work with a green suite, I asked a different model to review the whole thing with one instruction: use it the way a researcher would. It found that the map's headline layer had been painting the wrong quantity since the first day, that the basemap had never rendered in production, and that the newest feature was reachable only by the tests. Its diagnosis was that every phase had been "done" without anyone ever using the product. I have written before about automation that works perfectly for nobody. This was the same failure, and the missing user was me.
The report documents the rest: what the agents caught, what only a human in a browser caught, why commit metadata is an unreliable record of which model wrote what, and what I actually did, which was mostly decide and mostly agree with the recommendation put in front of me.
The ask
The map is meant to be used and argued with. The code is MIT, the openly licensed data files are downloadable with checksums, the whole build reruns from pinned sources with one command, and the report text is CC BY 4.0. There is a research agenda at the end of the report — validating exposure rankings against the 2026 closure, adding the resilience half (bypass and stocks), replacing the static route shares with observed ones, filling the gaps where the trade data fails.
If you work on energy security, or you are affected by it, I would rather this became a shared instrument than stayed a personal one. Corrections are as welcome as extensions, and I will publish them.