Your competitor for the input is in another industry

1. Two companies are a quarter of the chip industry

Most people's intuition about AI demand is several years out of date, so start with the scale.

Per their SEC filings, Nvidia's data-centre revenue was $193.7bn in FY2026 and AMD's $16.6bn in FY2025. Together that is $210.3bn against a $791.7bn global semiconductor market, or 26.6%.

Two companies' data-centre lines alone equal more than a quarter of an industry that also supplies every phone, every car, every appliance and every industrial controller on earth.

The trajectory is steeper than the level. Per IDC, data-centre semiconductors reach $477.1bn in 2026 and $843.2bn by 2030, approaching half the total market.

AI compute is no longer a segment of semiconductor demand. It is the largest single pole in it. Every other buyer of silicon, whether automotive, industrial, medical or consumer, is now negotiating for capacity against a customer with more money and more urgency. That is a different competitive position from five years ago, and most of them have not repriced it.

2. The magnet problem, and a comparison that needs care

Rare-earth magnets are the clearest case of cross-industry competition, because two unrelated energy transitions want the same material at the same time.

Magnets account for 41.0% of processed rare-earth oxide consumption, driven by NdFeB production for EV traction motors and direct-drive wind generators. Per Adamas Intelligence, global rare-earth magnet demand was around 260,000 tonnes in 2024 and is forecast at roughly 600,000 tonnes by 2035, a 2.31 times increase, or about 7.9% a year.

That growth rate is corroborated twice over. Adamas publishes a headline NdFeB demand CAGR of 8.6%, and separately forecasts a 206,000-tonne annual shortage by 2035 which it describes as nearly a third of the market, implying a 2035 market between 620,000 and 690,000 tonnes. Both agree with the numbers above.

The intensity comparison needs more care, because the usual version of it is wrong.

EV traction motors use 1–3 kg of rare-earth oxides per vehicle. Direct-drive wind generators use 100–500 kg per MW. This gets quoted very widely as wind using "two orders of magnitude more".

It is not a like-for-like ratio. One figure is per vehicle and the other is per megawatt, different denominators measuring different things. Pair the ends differently and the implied ratio runs anywhere between 33 times and 500 times.

The direction survives comfortably. Wind is far more magnet-intensive than automotive, and a wind build-out consumes material at a rate that dwarfs a vehicle programme. The precision does not survive, and anyone quoting a single multiple has silently chosen one pairing out of several. We give the range and say why it is a range.

3. Supply is not being built to meet it

Against a demand path that more than triples, the supply response is insufficient. Forty-two new magnet plants are needed by 2030, and eleven are in development. That is coverage of 26.2% of stated need, a shortfall of 31 plants.

The figure comes from a single industry analyst rather than an independently verified count, and plant numbers are a crude proxy for capacity, since plants vary enormously in size. We would not build a model on it.

The direction is unambiguous, though, and the more telling fact is what is missing: nobody has published a number that closes the gap. In a market this closely watched, the absence of a credible counter-estimate is itself informative.

4. China closed the door in three deliberate stages

The rare-earth chokepoint was not a single event, and reading it as one announcement misses how carefully the pressure was applied. It ran over two years, and each stage closed an escape route the previous stage had left open.

December 2023 brought the technology ban. China prohibited the export of rare-earth extraction and separation technology. Note what that does not restrict: the material itself. It restricts the ability to build the industry anywhere else. A country with deposits and capital could still buy Chinese material. It could no longer buy the means to stop needing to.

April 2025 added the elements that matter most. Seven medium and heavy rare earths went onto the export control list: terbium, dysprosium, samarium, gadolinium, lutetium, scandium and yttrium.

Those are not arbitrary. They are the elements that let a magnet hold its strength at high temperature, which is exactly what wind generators, advanced traction motors and defence systems require. The light rare earths that make up magnet bulk stayed available. The additives that make magnets work in demanding applications did not.

December 2025 reached beyond the border, with extraterritorial controls in the style of the US foreign direct product rule, covering products manufactured outside China that contain Chinese-origin inputs.

Technology first, so alternatives cannot be built. Then the specific elements those alternatives would most need. Then jurisdiction over what other people make with Chinese material.

Each stage answered the workaround the previous stage created. Anyone reading the third announcement without the first two will misjudge both the intent and the likely next move.

5. Defence gets paid to jump the queue

This is the part commercial buyers most often miss, and it changes the allocation question rather than merely the price.

In July 2025 the Pentagon took a $400 million equity stake in MP Materials, becoming its largest shareholder. The package included price floors at $110 per kilogram, guaranteed government offtake, Project Vault, and new bilateral frameworks with Australia, Japan, Malaysia and Saudi Arabia.

Look at what a price floor plus guaranteed offtake actually is. Calling it a subsidy obscures the mechanism. It is a claim on output that clears before the market does.

A commercial buyer competing for that material is not bidding against another company on comparable terms. It is bidding against a counterparty that has removed price risk from its own side of the trade, one that will still be buying when prices are high enough to make every commercial buyer walk away, and that has contracted for volume regardless.

Our read, not a sourced finding: any magnet-supply model that treats defence demand as one buyer among many has mis-specified the queue. The defence claim is senior, everything else is subordinate to it, and the seniority was purchased explicitly.

6. The battery pool has already switched

The same pattern is now visible in battery cells, and the swing buyer has changed without much notice being taken.

Battery demand from stationary storage rose 51% in 2025, against 26% growth in EV-related demand. Storage is growing 1.96 times as fast as the application everyone models.

Manufacturers have already responded. More than 50 GWh of capacity from LG Energy Solution and Ford was reallocated toward LFP production, largely targeting grid storage rather than vehicles.

For anyone modelling EV battery costs this inverts the standard assumption. Grid storage and electric vehicles draw on one pool of cells, not two separate markets. They compete for the same factories, the same cathode material and the same lithium.

So a grid-storage surprise is directly an EV cost event. A data-centre operator or a utility signing a large storage contract is, in supply-chain terms, a competitor to a carmaker, and the capacity reallocation above shows manufacturers already treating the two as interchangeable demand.

7. How to run the census

The method is short, and it is the single thing here worth copying.

For any critical input, list every industry that consumes it, not every company in your own industry. Then ask three questions about each.

How much do they use per unit of their output? Intensity, not size. A small industry with high intensity can dominate demand growth for a specific material.

Is their demand growing faster than yours? Growth rate determines who sets the marginal price in five years, regardless of who is larger today.

Can any of them pay to clear before you do? Government offtake, price floors, strategic reserves, national-security designations. Anything that makes a buyer price-insensitive makes them senior in the queue.

Run those three tests on magnets from an automaker's position and it fails all three. Wind uses far more per unit of output. Storage is growing faster than EVs. Defence has bought priority outright.

That census takes an afternoon. It needs no proprietary data and no forecasting. Not running it is how a supply chain looks comfortably diversified right up to the moment it is not, because every supplier was mapped and none of the other buyers were.