Intelligence is becoming a flow. Atoms are still a stock. Over the next decade, money moves from whatever AI makes abundant to whatever AI cannot make faster.
The idea
AI keeps making thinking cheaper. Writing, analysis, code and review get less expensive with every model release. Every unit of that cheap intelligence runs on things that do not scale like software: power, memory, chip packaging, water, metals, explosives for defence, fertiliser for food. Those are stocks. A stock gets dearer as you draw it down, and adding more takes years and billions.
When one input becomes abundant, the scarce input next to it takes the margin. We look for that point and position on both sides of it: long the firm that owns the scarce step, short the firm that has to pay for it.
Why now
- AI is limited by physics, not by demand. Moving data costs far more energy than computing on it. The bill sits in memory, packaging and power.
- Lithography is running out of road. In September 2026 Kepler Computing raised USD 468M to build dense AI memory without EUV. Serious capital now pays for density that does not come from shrinking transistors.
- Material demand keeps compounding. The world used 30 billion tonnes of material in 1970 and uses about 106 billion now. Roughly 7% comes back into use.
- 2026 showed how thin the chains are. The Strait of Hormuz has been effectively closed for most of the year. One sea lane hit helium for chip fabs, desalination in the Gulf, sulphur for fertiliser and European gas, all at once.
How we pick
We do not screen for themes. We start from physics.
- Define the unit the buyer needs, such as one firm megawatt-hour for a data centre.
- Write the physical chain from raw input to that unit.
- Score every link: time to add capacity, number of qualified suppliers, substitutes, capital needed, exposure to hostile states, and share of the end cost.
- Find the binding link: long lead time, few suppliers, no substitute.
- Find who owns it, then check whether the market already pays for it.
- Write down the physical event that would remove the bottleneck. We watch that, not the share price.
We add one more test. Is the link a stock, whose cost rises as it is used up, or a flow, whose cost falls on a learning curve, like sunlight or engineered biology? Bottlenecks on stocks are what we own today. Flow routes near the point where they become cheaper are what we own next.
Six pillars
| Pillar | The binding link we found | What would prove us wrong |
|---|---|---|
| Compute | Substrate materials. One supplier makes over 95% of the build-up film inside every AI chip package, and its next plant opens in 2032. | A second film source qualifies, or glass-core substrates scale early. |
| Energy | Large power transformers and the electrical steel inside them, with lead times of two to three years. Also turbine blade castings: four foundries at scale, all full. | Transformer lead times fall below a year. |
| Water | A permitted coastal site, not the membrane. Equipment takes one to two years. A permit takes many years, or never comes. | Permitting speeds up in California, Chile and southern Europe. |
| Materials | Copper mines, which now take close to eighteen years from discovery to output, and heavy rare-earth separation, where China holds about nine tenths of refining. | China drops its export controls for good, or new separation capacity outside China arrives early. |
| Food | Sulphur for phosphate, much of it shipped through Hormuz, and cheap gas for nitrogen. US producers pay a fraction of what European plants pay for the same gas. | Hormuz reopens and sulphur prices fall back. |
| Sovereignty | Energetics, not shells. Europe can press shells faster than it can make the nitrocellulose and explosives to fill them. Missiles depend on very few rocket-motor suppliers. | A lasting ceasefire combined with fiscal austerity in Europe. |
Where we think the market is wrong
- The obvious bottlenecks are priced. Memory makers and copper miners have already re-rated sharply. The edge is one layer down.
- The real chokepoints are small and boring. Build-up film, transformer steel, turbine castings, rocket-fuel oxidiser, nitrocellulose. Each is a tiny part of the end cost, so buyers cannot refuse a price rise. Several sit inside companies the market files under food, lubricants or general industry.
- By-products cannot answer price. Sulphur and helium come out of oil and gas. A higher price does not create more of them. Shocks there last until the source reopens.
- Many sell-offs are company news, not thesis failure. Defence and grid names fell on guidance cuts and project delays while the physical shortage stayed in place.
Rules
- USD 50M notional. A liquid core of listed equities, pairs where possible (long the bottleneck, short the firm that pays for it), and cash for new ideas.
- Positions of 1% to 4%. Nothing we cannot explain from first principles in five minutes.
- Every decision is logged before it trades, filled at the next market open, and never edited afterwards. The log is timestamped on the Bitcoin blockchain and in git.
- Each position carries the physical kill signal that would close it.
What could make the whole thesis wrong
- A sharp cut in AI capital spending removes the demand behind compute, energy and water at once.
- Model efficiency improves so fast that energy per useful operation drops without new hardware.
- A broad peace, open sea lanes and a global slowdown together. That would loosen almost every link we track.