How fast can atoms keep up?

In short. Ideas move fast but also spread fast; chips, power and minerals are slow to build and hard to copy. At today’s balance between the two, middle powers, countries that cannot match the United States or China alone but still carry real weight, hold much of the chip supply chain and with it the most leverage. If AI’s growth shifts towards software, a lead becomes worth more just as that leverage fades, so now is the time to write lasting rights into law.

The mix and its direction

Some of AI’s progress comes from better methods that squeeze more out of the same chips. The rest needs new physical things: more chips, more electricity, and eventually robots and factories. Recent progress has come from both⁸²⁸³.

How this mix changes as AI starts doing AI research is the biggest open question behind this post. In 2026, AI already leads about a quarter of the research tasks at one major lab, under human supervision⁸⁵, and at another, AI agents put in about three working days of effort for every day of human research work⁸⁶. If AI researchers can stand in for the computing power that experiments need, progress could become almost entirely a matter of ideas; the evidence on that is split⁸⁰⁸⁷. Part of what looked like pure software progress in the past turned out to need ever larger computers to pay off⁸¹, and by one estimate AI’s contribution to its own research is still somewhat below the level needed to keep speeding itself up⁸⁴⁸⁹.

The chart below makes the mix a slider, and the interactive charts further down follow it. Capability here means effective computing power: the chips’ raw power multiplied by how well the methods of the day use it. Ten times more of it does not mean ten times more useful work.

Four things need separating: technical capability, deployment, economy-wide growth and who gets the gains. Better models still need useful applications, changed workflows and people with the skills to implement them. Research on the productivity J-curve explains why those complementary investments take time and can initially obscure gains¹⁰⁵. Demand matters too: saving labour per service need not cut employment if lower prices bring enough new customers, but that is a conditional mechanism, not a promise that new demand will absorb everyone¹⁰⁶.

Cunningham and colleagues’ model separates narrow improvements at AI research from broader economically useful capabilities. Their rough calibration falls short of self-sustaining acceleration; it is not a measured ceiling, and automating research alone does not establish an economy-wide growth explosion⁸⁴. These charts explore technical and physical scenarios, not forecasts of GDP or wages.

Software onlyChips, power and robots ▲ Today (rough)

AI capability physical build-out robot production

Change the assumptions

How fast ideas alone raise capability: Fastest doubling time of physical build-out:

Model: AI capability grows only as fast as the slower of its two ingredients, better ideas and more physical things. The speeds of ideas and of building are adjustable assumptions; copper and magnet inputs match the race chart further down. Treat the dates as illustrations.

Ideas can, in principle, raise capability ten-fold a year or more, while physical build-out at its fastest doubles every year or two. So once the physical share passes a small threshold (about a sixth, with the chart’s starting assumptions), the speed of building decides the speed of AI.

Which way is the slider moving? By my rough reconstruction from published trends, for about a decade roughly 55–60% of AI’s progress came from new hardware. From late 2024 the share fell to around 40%: AI systems that reason step by step got much better while running on chips that labs already had, and measured capability started improving more than twice as fast as before⁹¹⁹². That dip probably came from a one-off source of easy gains. The four largest cloud companies expect to spend $720–745 billion in 2026, much of it on AI data centres and chips⁹³, which could push the share back up, while AI doing more of its own research would push it down⁸⁵⁸⁶. My best guess is a share that holds steady or edges up over the next two years, with a wide range either way.

Who holds leverage, from mines to models

A country has leverage when it controls something the rest of the world needs and cannot quickly replace. AI depends on a long chain of such things, from mines and refineries at one end to computers and trained models at the other. The chart below shows who holds each stage. Each row is shaded by how much that stage matters at the mix you choose, and the bottom row adds the stages up.

Software onlyChips, power and robots

Stage	USUS	ChinaCN	JapanJP	TaiwanTW	KoreaKR	NetherlandsNL	GermanyDE	UKUK	Gulf statesGulf	Australia & CanadaAU/CA	DRC, Indonesia, Chile, PeruProducers	OthersOther
Mining	<5%	32% ●								<5%	41% ●	19%
Mineral refining	<5%	70% ●	<5%							<5%	10% ●	13%
Magnets & robot parts	<5%	50% ●	40%		<5%		5%					<5%
Power & data-centre hosting	45%	25%	5%		<5%	<5%	<5%	<5%	<5%	<5%		14%
Grid equipment	15%	40%	10%		5%		10%					20%
Chip materials	15%	5%	60% ●	5%	5%		5%					5%
Lithography & chipmaking tools	40% ●	5%	20% ●		<5%	25% ●	<5%					5%
Leading-edge chips & packaging	5%	<5%	<5%	80%	10%							<5%
AI memory (HBM)	20% ●				80%
AI chip design	95% ●	<5%										<5%
Installed AI computing power	75%	15%	<5%		<5%		<5%	<5%	<5%			<5%
Notable AI models	55%	35%			5%			<5%		<5%		<5%
Weighted total

Shading = share of world supply × how much that stage matters at the chosen mix. ● = supply restricted 2019–26. Measures differ by row; hover a cell for detail.

Slide to the software end and two rows take over: the computers that already exist and the leading models. In a large sample of AI supercomputers from early 2025, about three quarters of the computing power was in the United States and about a sixth in China⁸⁸, and more than half of the notable models released in 2025 were American¹⁰⁴. Slide to the physical end and refining, magnets, robots and grid equipment take over, where China leads: in 2024 it refined about 91% of the rare earths used in magnets and made about 94% of the strongest magnets²³, and it produces about 99% of new gallium⁴². Mining is more spread out: the Congo mines about three quarters of the world’s cobalt, Indonesia two thirds of its nickel, and Chile, the Congo and Peru half of its copper¹¹³, though Chinese firms own many of the mines in the Congo and Indonesia¹. In the middle, around today’s mix, chips and the machines and materials that make them matter most, and that is where Japan, Taiwan, South Korea and the Netherlands are strong.

China uses its levers, and presents them as normal. Its commerce ministry says “export controls are not export bans”⁶¹, and its controls have usually followed US tariffs or US limits on chip exports⁶². Since March, Chinese rules also let Beijing retaliate against foreign companies that cut off Chinese suppliers⁶³. China is exposed too: it imports over 80% of the copper ore it smelts⁵² and over 45% of its sulfur, more than half of it through the Strait of Hormuz⁶⁴, and its controls leak as provinces protect local businesses and smugglers follow price gaps⁶⁵.

Japan is the main counterweight in chip materials. Japanese companies make about 44% of the world’s silicon wafers, around three quarters of the light-sensitive chemicals used to print chips, most of the glass cloth and insulating film in chip packages, and most precision robot gears^6667175728. When China banned exports of dual-use goods, items with both civilian and military uses, to Japanese military users in January, Japan’s imports of dysprosium ore fell 82% in the first half of the year²⁴; Japanese tungsten users switched to scrap and recycling, and some analysts think the controls are backfiring⁶⁸.

Mining countries have levers of their own. Since 2020, countries have imposed 98 export restrictions on critical minerals, nearly half of them by the Democratic Republic of the Congo, China and Indonesia⁶⁹. The Congo limits cobalt exports to support prices⁵⁸, and Indonesia insists on processing its nickel at home⁷⁰. Korean defence firms buy rare earths through Russia and Arab states⁷¹, and Chile’s parliament has warned that a US minerals deal must leave its trade with China untouched⁷². None of the producer governments I looked at talks about AI or robots as its opportunity.

Fear itself can be a lever. China’s 2010 rare-earth “embargo” against Japan barely shows in the trade data, yet it shaped Japanese policy for years afterwards⁷³: working with its industry, Japan cut its use of rare earths to about 30% below 2010 levels²³.

Importers are building levers of their own, slowly. The EU’s Critical Raw Materials Act, in force since May 2024, sets non-binding goals for 2030: mine at least 10% of the strategic raw materials the EU uses within the EU, process 40% and recycle 25%, and take no more than 65% of any one of them from a single non-EU country¹⁰⁸. In February, the EU’s auditors found that its partnerships with supplier countries had so far contributed little to secure supply. For processed lithium, magnesium, gallium and rare earths, the EU still gets more than 65% from a single country, and 10 of the 26 critical raw materials needed for clean energy are not recycled at all¹⁰⁹. Of the 60 projects the EU has selected for faster permits and support, only two target materials for chips; where buyers have been announced, about as much output is promised outside the EU as inside it, including half of one French project’s heavy rare earths to Japan¹¹⁰. Middle powers cooperate and compete for the same supply.

What a lead is worth, and who can bargain

Bottlenecks also affect who gets paid. In Jones and Tonetti’s model, essential tasks that remain human can slow aggregate growth while sustaining labour’s share of income¹⁰⁷. My inference for physical bottlenecks is less reassuring: if the scarce input is a privately owned grid connection, chip plant or mineral deposit, the gains may instead concentrate with its owners. Slower growth is not automatically fairer growth, nor does a supply share by itself show that a government can use it to bargain.

In an economy that doubles every year, a two-year head start means four times the output⁶. The faster AI grows, the more a head start is worth. But a lead is worth only what you can keep, and that depends on how quickly others can catch up.

Ideas move fastest, and they also travel fastest. They spread through published papers, through staff who change jobs, and through training a smaller model on a larger model’s answers. The best small, freely available models have improved about as fast as the leading ones, so the gap between them has stayed roughly fixed⁹⁰⁹¹, and models that run on a single high-end gaming chip trail the best ones by roughly six months to a year on several tests⁸³. A lead in factories, mines and power stations travels much more slowly: a rival has to build its own.

So the lead you can keep is largest a little way in from the software end. There, AI grows very fast, because a little building adds to what ideas alone achieve, and part of each year’s gain is tied to equipment that rivals lack. At the pure software end, AI still grows very fast, but rivals copy most of each year’s gain, compress it onto smaller computers and compete with it. Towards the physical end, little leaks, but AI grows slowly, so a year’s lead is small. How much leaks is uncertain: the chart below assumes rivals copy about 70% of a year’s gain from ideas and 10% of a gain from building, and you can change both.

Even at the software end, the United States’ position rests on owning hardware. Ideas that leak still need computers to run on, and the country with the most computers gets the most out of every idea, including its rivals’. That is why the US share stays high at that end of the chart above.

The chart below puts the value of a lead next to the leverage of middle powers, here Japan, Taiwan, South Korea, the Netherlands, Germany, the UK, the Gulf states, and Australia and Canada together.

Software onlyChips, power and robots

The lead you can keep (share of its highest value) Middle powers' combined share US share China's share

Change the assumptions

How much of a lead from ideas rivals copy within a year: How much of a lead from building rivals copy within a year:

The lead you can keep: a one-year leader's extra capability after rivals copy part of it, on a log scale (formula in the notes). Leverage: each group's share of the supply-chain stages in the chart above, weighted by how much each stage matters at the chosen mix. Middle powers: Japan, Taiwan, South Korea, the Netherlands, Germany, the UK, the Gulf states, Australia and Canada.

The two curves peak in different places. With the chart’s starting assumptions, middle powers together hold about 40% of the weighted supply chain around today’s mix and under 20% at the software end, where the United States holds nearly 60%. Where a lead is worth most, a little way in from the software end, their share is about 30%. These numbers move with the weights, but the direction holds: the closer to the software end, the smaller the middle powers’ share. The 40% is a combined total: no single middle power comes close to the US share, so it counts only if they act together. An AI boom that also automates factories may favour countries with large industrial bases and light regulation, such as China and the Gulf states⁷⁴; a boom that runs mostly in software favours whoever owns the computers and the leading models.

So if AI’s growth shifts towards software, middle powers lose leverage just as a lead becomes worth more. The levers that hold up in that shift are legal ones: rights over computers on their soil, rights to test the models that run there, and rules for the devices sold in their markets. I argue in a separate post that the time to write those rights into law is while the supply-chain leverage still exists.

The physical end: why building is slow

A process can only go as fast as its slowest step. Training a new AI model takes months. Building a data centre takes a year or two. A new copper mine takes about seventeen years on average, from discovery to first production¹. Much of that waiting is queueing: a large transformer takes a few months to make, and the two-year wait is the line of orders ahead of it¹⁵.

When demand doubles on a fixed schedule, supply that grows a steady few percent a year falls further behind every year. Imagine robot production doubling every eighteen months while the supply of the magnets inside them grows 8% a year, a healthy pace for a mining industry: 100,000 robots become a million in five years and ten million in ten.

Humanoid robots → NdFeB magnets Robots built outside China → Dy + Tb AI data centres → copper

Extra supply buys surprisingly little time. In the example above, doubling the magnet supply pushes the shortage back by only about 1.8 years, and a tenfold supply by about six. The date depends mostly on how fast demand doubles.

Materials are a tiny part of what AI costs. Copper is between under 0.5% and about 6% of the cost of building a data centre, depending on what is counted³⁵³⁶, and all the helium used in chipmaking is worth about $1 billion a year³⁷. Think of baking powder, though: a tiny share of what a cake costs, yet without it there is no cake. When an input is essential and nothing can replace it, a shortage does far more damage than its share of the cost suggests. That effect roughly tripled the economic damage of the 1970s oil shocks³⁸, and growth is held back by whatever is essential and hard to improve³⁹. For such inputs, the danger is whether you can get them at all⁴⁰.

Governments write this idea into lists of critical raw materials. The EU calls a material critical when it matters to its economy and its supply is at risk, a risk that rises when production is concentrated in a few countries, especially poorly governed ones, and falls when substitutes or recycling can fill the gap. Its list has 34 such materials, including gallium, germanium, tungsten, helium and the rare earths; 17 of them are “strategic” because demand is expected to outpace supply¹⁰⁸. Copper and nickel fall short of the thresholds but are on the list as strategic materials anyway¹¹¹, and the United States added copper to its own list in 2025¹¹². These lists cover raw materials only. Several of the tightest chokepoints in this post, such as chip-package glass cloth, insulating film and precision gears, are manufactured goods, and sulfuric acid is on neither the EU nor the US list.

2026: a year of chokepoints

A chokepoint is a narrow point that everything must pass through: a strait that all shipping must use, or a factory that makes most of the world’s supply of something. Each of 2026’s big material price jumps had a cause of this kind.

• War. In March, Iranian strikes on Qatar’s Ras Laffan gas complex took out about a third of the world’s helium supply⁴⁷¹ and, for three to five years, 17% of Qatar’s gas export capacity¹⁸. Helium is essential for making chips, and South Korea’s chipmakers, which get over half of theirs from Qatar¹, said in late March they had months of stock⁴⁸. Since a June ceasefire, shipping through the Strait of Hormuz has partly recovered, to 124 transits a week by ships not linked to Iran in late September⁴⁹, against more than 100 ships a day before the war⁷⁹, and attacks on tankers have risen again⁴⁹.
• Export rules. In May, China stopped exporting the sulfuric acid left over from smelting metals⁴⁴⁵⁰⁵¹. It exports only about 4% of what it makes⁵², but half of those exports went to Chile and Indonesia, which use the acid to extract copper and refine nickel¹. More than 15% of the world’s mined copper is extracted with sulfuric acid, mostly in Chile and the Congo¹, and the acid price in Chile doubled in seven weeks⁴⁴⁵³. Gallium outside China reached $1,850 a kilogram in April while staying flat inside China⁴², and tungsten rose about 350% in four months as China cut quotas and licensed just fifteen exporters⁴¹.
• One mine going wrong. In August, copper traded near record highs, partly because Grasberg in Indonesia, one of the world’s largest copper mines, was still far below capacity after a 2025 mudslide⁴⁵⁴⁶⁵⁴.

A major constraint on AI this year was factory capacity. The leading chipmaker’s most advanced factories have run flat out since late 2025, and the special memory that AI chips use may stay short until 2030⁵⁵; a new memory factory takes three to five years¹⁶. Makers of gas turbines for power stations have roughly three years of orders waiting⁵⁶. One Japanese company makes about 85% of the world’s special glass cloth for AI chip packages, with no relief before 2027–28¹⁷. Another, Ajinomoto, better known for its seasonings, makes almost all of the insulating film in those packages and reportedly cut supply to mainland China by 30% in August⁵⁷.

Material details: magnets, gears, copper and leftovers of other industries

How fast could robots grow? Forecasts put 2026 shipments of humanoid robots at about 75,000–90,000¹⁹²⁰. Each needs about 2.5 kg of the strong magnets (made of neodymium, iron and boron) that also go into electric cars and wind turbines, and the world uses about 226,000 tonnes of them a year²¹. (The race chart above lets robots claim only a share of the supply, so its dates come later than these.) If robot production doubled every year and magnet supply stayed flat, robots would need a fifth of today’s magnets around 2033 and all of them around 2036; with a two-year doubling, about 2041 and 2046. Bank forecasts of 6.5 to 10 million humanoids a year by 2035 imply a doubling roughly every 1.3–1.4 years¹⁹²⁰. One industry forecast reaches about 151 million robots a year by 2040, needing about 160% of today’s magnet market²¹, and a widely read AI scenario imagines a million new robots a month³, which would need 11–16% of today’s magnets.

Where magnets run short. Outside China, the refined oxide used in most magnets is in surplus: in August its price fell below a floor the US government guarantees some Western producers, because there are too few magnet factories to buy it²². The shortage is in making the magnets themselves²³, and new magnet plants take years and must compete on cost with China’s mature industry²⁴. Tighter still are dysprosium and terbium, which keep magnets working when hot. The only commercial producer outside China made 26 tonnes in one recent quarter and 8 in the next²⁵, against a capacity of 240 tonnes of dysprosium and 50 of terbium a year if it can get enough ore²⁶. By one industry measure, a million humanoids a year would need about 57 tonnes, though robots need far less of these metals per kilogram of magnet than electric cars do²⁷. Even the main rare-earth-free alternative has a chokepoint: most ferrite magnets are made with strontium, another material on the EU’s critical list, and Iran produces more than half of the world’s supply¹¹⁰⁸.

What runs short before magnets. Gears and hands. One Japanese company makes about 80% of one kind of precision gear used in robot joints²⁸, and the best-known humanoid programme has been slowed by fragile hands, supplier quality and software²⁹. Solving one problem can create another: magnet use per robot is expected to rise as robots gain joints and motors³⁰.

Copper. A data centre contains about 12–47 tonnes of copper per megawatt³¹¹², and the power lines that feed it roughly triple that¹⁴. Data centres may need about 2.5 million tonnes of copper a year by 2040, out of roughly 28 million used worldwide¹²¹. At today’s forecast growth, that means higher prices and some delays: many announced data centres are running late³², and new power designs and optical cables use less copper³³³⁴. Copper becomes a hard limit if data-centre building keeps doubling, which is what the middle of the slider and its physical end imply. Studies of mineral supply assume data-centre building grows steadily, from about 15–20 GW of new capacity a year today to a peak around 30 GW in the early 2030s¹²¹³¹⁴.

Leftovers of other industries. The hardest shortages to fix are of materials nobody mines on purpose. Helium comes out of natural gas, sulfur from refining oil and gas, sulfuric acid from smelting metals, and gallium from refining bauxite into alumina, the step before aluminium. Western refineries process about a third of the world’s alumina but recover almost none of its gallium, and existing alumina and zinc plants could yield several times today’s gallium output if they added recovery equipment⁴⁰. The Middle East supplies about a quarter of the world’s sulfur, and half of the sulfur shipped by sea passes through the Strait of Hormuz¹. Moving away from fossil fuels removes the source of over 80% of the world’s sulfur just as demand for sulfuric acid rises⁵⁹, and a fast, AI-run electrification of the economy⁴ would tighten that squeeze.

The software end: computing power without building

Towards the software end of the slider, AI can get more computing power without anyone building it, by using computers that already exist, including ones whose owners never agreed to that use.

In October 2026, researchers at one AI company reported that small AI models may be months away from being able to copy themselves onto cheap, poorly secured devices such as off-brand TV boxes, set themselves running there, and keep spreading⁹⁰. Each copy would be slow: two million such devices would add up to roughly the computing power of forty data-centre chips⁹⁰, a tiny amount next to a data centre. The worry is persistence. With no central point to switch off, copies that pass notes to their successors and adapt to new hardware could become ever-present, the way the common cold is⁹⁰.

Defence then looks like public health. Secure default settings play the part of sanitation: the UK, for example, no longer allows devices to ship with universal or easily guessed default passwords⁹⁰. Monitoring network traffic is surveillance. Patches and takedowns are antibiotics, and adaptive copies are the resistance that overuse breeds. Quarantine has worked before: when internet providers isolated customers whose devices were infected by an earlier botnet and told them how to clean up, 92% of infections were cleared within two weeks, against 58–74% in comparison groups that were not contacted⁹⁰⁹⁵. Gene drives, which spread a change through a wild population, have a cyber counterpart in defensive programs that spread fixes on their own. These have been tried, and they caused damage of their own and entered computers without their owners’ permission⁹⁴, so they are a poor model to copy.

At this end of the slider, internet providers, national cyber agencies, device regulators and the companies that hold the models’ weights, the learned numbers a model is made of, matter more than mines. Middle powers have real tools here: rules for the devices sold in their markets, coordinated takedowns, and incident reports shared across borders, which is one of the six conditions I propose for upcoming laws.

What to watch

• How much of AI research AI does, and whether experiments need ever more computing power. This moves the slider⁸⁴⁸⁵⁸⁶.
• How quickly smaller and freely available models catch up with the best ones. This sets how much of a lead can be kept⁹⁰⁹¹.
• Whether AI copies start living on devices outside data centres, and whether device rules and coordinated takedowns keep each infection causing fewer than one new one⁹⁰.
• How fast robot production doubles. If it takes two years or more, robots won’t need even a fifth of today’s magnets before the 2040s. If it doubles every year past 2030, they’ll need all of them by the mid-2030s.
• Late 2026 to January 2027: whether China’s acid export halt ends⁵⁰⁵¹; whether rare-earth oxide made outside China finds buyers²²; and the US–China trade truce deadline on 10 January, where analysts say a lasting deal needs “actual deliveries of rare earths”⁷⁵⁷⁶. The truce also delays China’s October 2025 rare-earth controls⁷⁶, which would require licences even for products made abroad that contain Chinese rare earths; enforced in full, they would put about $6.5 trillion a year of production outside China at risk, by the IEA’s estimate²³. China’s April 2025 licensing rules for seven rare earths and magnets containing them were not part of that pause²³¹¹⁴.
• 2027–28: new glass-cloth capacity¹⁷; Ajinomoto’s film expansion from 2028, with a third plant around 2032⁵⁷; Japan’s trial of mining rare-earth mud from the seabed⁷⁷; and Grasberg’s return to full output, targeted for end-2027 but possibly early 2028⁵⁴.
• New sources: using microbes to extract copper from low-grade ore⁷⁸, AI-assisted exploration, and magnets and motors that need no rare earths.

Notes on method

The slider model. AI capability grows only as fast as the slower of its two ingredients: ideas, which by default can raise capability ten-fold a year on existing hardware⁸⁷, and physical build-out, which at its fastest doubles every 1.5 years²⁴. Both are adjustable under “Change the assumptions”. Capability here means effective computing power, which is how the cited studies measure progress⁸⁷. If a share $s$ of progress needs physical things, capability grows at roughly the smaller of $g_I/(1-s)$ and $g_P/s$ (a smooth minimum, $g_E = \big[(g_I/(1-s))^{-3} + (g_P/s)^{-3}\big]^{-1/3}$, so a little building adds to what ideas achieve), physical build-out grows at $s$ times that rate (with a floor of 5% a year), and robot production then grows at $s$ times the build-out rate (with a floor of doubling every three years; my assumption). Copper and magnet dates use the same inputs as the race chart. Today’s mix of roughly 40% comes from published trends in algorithms, chip efficiency and the number of chips⁸²⁸³.

The value of a lead. A one-year lead gives the leader $g_E$ more capability on a log scale, where $g_E$ is the capability growth rate above. Over the following year, rivals copy a fraction $c_I$ of the part that came from ideas and a fraction $c_P$ of the part that came from building, so the lead that can be kept is

$$\ell(s) = g_E(s)\,\big[(1-s)(1-c_I) + s\,(1-c_P)\big].$$

By default $c_I = 0.7$ and $c_P = 0.1$, my assumptions informed by how quickly small models and consumer hardware catch up⁸³⁹⁰⁹¹. Both are adjustable under the chart.

Leverage. Each stage of the supply chain gets a weight that depends on $s$: computers and models weigh most at the software end, chips in the middle, and minerals, magnets, power and grid equipment at the physical end (my judgement). A country’s leverage is its share of each stage, averaged with these weights; for middle powers the shares are added up. Shares are rounded, measured differently by row and often my averages of several sources, so treat small differences as noise. Halving the weight of the chip stages lowers the middle powers’ share at today’s mix from about 40% to about 30%. A share is also a rough stand-in for leverage: a single stage with no substitute can stop everything, which an average does not capture, and no single middle power comes close to the US share.

What the growth models cover. Researchers who model how fast an AI-driven economy could grow mostly conclude that raw materials will not stop it²³⁴⁵, and over the long run I agree. Estimates of the copper known to exist, counting what has already been mined, grew about 25-fold between 1900 and 2010 as exploration and technology improved⁸; for almost every mineral, the demand from the shift to clean energy is below known resources⁹; mining is about 1% of economic output, so even large cost increases barely slow growth overall⁶; and the first part to run short for robots today is precision gears, which cap humanoid production at about half a million a year⁷. Those models leave open how fast particular mines, refineries and magnet factories in particular countries can grow, and none of them includes an export ban, a refinery that exists in only one country, or a missile hitting a gas plant⁴¹⁰. This post looks at those gaps.

The race chart. Demand grows exponentially, supply grows at a fixed rate, and a fixed share of it is available to robots or data centres. It ignores prices, substitutes, stockpiles and recycling, all of which would soften the shortage, so treat the dates as illustrations. Recycling works relatively well for bulk metals such as copper and aluminium, but many critical materials used in small amounts are barely recycled or not at all¹⁰⁹; for rare earths, recycling could cut the need for newly mined supply by up to 35% by 2050²³. If demand doubles every $T_d$ years and supply grows by a fraction $g$ a year, doubling the available supply delays the shortage by

$$\Delta t = \frac{\ln 2}{\ln 2 / T_d - \ln(1+g)}$$

years, as long as demand grows faster than supply.

Some figures come from paywalled reports read through summaries, and some prices from secondary sources. The research was done with the help of AI research agents (Claude, made by Anthropic) and checked by independent model reviews, and the header image was generated with an AI image tool. Every number links to its source; corrections are welcome.

References

1. International Energy Agency, Global Critical Minerals Outlook 2026. iea.blob.core.windows.net
2. Tom Davidson & Rose Hadshar, “The Industrial Explosion”, Forethought, 2025. www.forethought.org
3. Daniel Kokotajlo et al., “AI 2027”, AI Futures Project, 2025 (PDF: robot milestones in both endings; Appendix Q, “Robot economy doubling times”, and its footnotes). ai-2027.com
4. Damon Binder, “The AI Industrial Explosion, Part 1: Maximum growth rates with current technology”, 2026. defensesindepth.bio
5. Damon Binder, “The AI Industrial Explosion, Part 2: Transition dynamics”, May 2026. defensesindepth.bio
6. Damon Binder, “AGI will set off an industrial explosion”, AI Frontiers, 2026. ai-frontiers.org
7. Epoch AI, “How fast could robot production scale up?”, 2026. epoch.ai
8. Richard Schodde, “Long term outlook for the global exploration industry”, MinEx Consulting. www.uvm.edu
9. Hannah Ritchie, “How much will mineral demand increase for the energy transition?”, 2024. hannahritchie.substack.com
10. Epoch AI, “Can AI scaling continue through 2030?”, 2024. epoch.ai
11. Epoch AI, “Introducing the AI Chip Components Explorer”, May 2026. epoch.ai
12. S&P Global, “Copper in the Age of AI”, January 2026. www.spglobal.com
13. Wood Mackenzie, “Data centre metals demand isn’t an asset story. It’s a grid story”, June 2026. www.woodmac.com
14. Amoah, Brown, Simon, Bazilian & Matisek, “Mineral demand from AI data centers: infrastructure intensity, processing bottlenecks, and supply competition”, Resources Policy 119, 2026 (Payne Institute). payneinstitute.mines.edu
15. Damon Binder, “The AI Industrial Explosion, Part 3: Going faster”, 2026. defensesindepth.bio
16. MUFG, “Bottlenecks to Scaling AI Computational Power”, June 2026. www.mufgamericas.com
17. Yole Group, “AI supply chain at risk: between T-glass shortage and emerging glass core”, February 2026. www.yolegroup.com
18. Al Jazeera, “Iran attacks cut 17% of Qatar’s LNG capacity for up to five years: QatarEnergy”, 19 March 2026. www.aljazeera.com
19. Goldman Sachs humanoid forecast, via Yahoo Finance, September 2026. finance.yahoo.com
20. Bank of America Institute, “Physical AI, part 2”, March 2026. institute.bankofamerica.com
21. CRU, “Humanoid robots can widen the projected rare earth supply deficit”, September 2026. www.crugroup.com
22. Benchmark Mineral Intelligence, “Ex-China NdPr prices slip below $110/kg price floor as oversupply builds”, August 2026. source.benchmarkminerals.com
23. International Energy Agency, Rare Earth Elements: executive summary, April 2026. www.iea.org
24. Global Times, “Japan’s H1 dysprosium raw mineral imports plunge 82%”, August 2026. www.globaltimes.cn
25. Rare Earth Mining, “Lynas Rare Earths results: Q3 FY26” (quarterly Dy+Tb output), April 2026. rare-earth-mining.com
26. Lynas Rare Earths, June-quarter FY26 earnings call transcript (Dy 240 t/yr and Tb 50 t/yr capacity, feedstock permitting), July 2026. www.investing.com
27. Emerson Alden, “The Robot Boom Has a Magnet Problem”, Foundation for American Innovation, September 2026. www.thefai.org
28. John VerWey, “Physical AI”, CSET Georgetown, February 2026. cset.georgetown.edu
29. Electrek, “Tesla Optimus production ramp: hands, AI generalization problems”, 25 September 2026. electrek.co
30. Adamas Intelligence, “What Boston Dynamics’ Atlas reveals about the future of NdFeB magnet demand”, January 2026. www.adamasintel.com
31. World Economic Forum & Kearney, “From Minerals to Megawatts”, December 2025. reports.weforum.org
32. Oxford Smith School & Marex, “AI data centres: impact on power and commodity markets”, April 2026. www.smithschool.ox.ac.uk
33. NVIDIA, “800 V HVDC architecture will power the next generation of AI factories”. developer.nvidia.com
34. Andy Home, “AI may not be the demand booster copper bulls expect”, Reuters, 3 June 2026. www.reuters.com
35. Resource Capital Funds, “Metal mining gigacycle: parsing metal demand at the source”, August 2026. resourcecapitalfunds.com
36. BloombergNEF data-centre copper tracking, via Carbon Credits, 2025. carboncredits.com
37. PIIE, “Bring back the helium reserve before the next shock hits”, April 2026. www.piie.com
38. David Baqaee & Emmanuel Farhi, “The Macroeconomic Impact of Microeconomic Shocks: Beyond Hulten’s Theorem”, Econometrica, 2019. www.nber.org
39. Philippe Aghion, Benjamin Jones & Charles Jones, “Artificial Intelligence and Economic Growth”, NBER, 2019. www.nber.org
40. Piyush Akimitsu, “Artificial Intelligence: Supply-Chain Chokepoints and the Reach of Industrial Policy”, arXiv 2607.29572, 2026. arxiv.org
41. Industrial Info Resources, “China restrictions lead to tungsten supply crunch”, June 2026. www.industrialinfo.com
42. CSIS, “The U.S.-China trade truce has not solved the gallium problem”, May 2026. www.csis.org
43. GameChange Energy, “Transformer lead times in 2026”. www.gamechangeenergy.com
44. Akin Gump, “When supply chains break: force majeure, hardship and the sulfuric acid crisis”, April 2026. www.akingump.com
45. Reuters, “Copper to hold gains in 2026 as mine disruptions fuel deficit” (analyst poll), October 2025. www.reuters.com
46. MINING.COM, “Copper price back at record’s edge as Grasberg smelter halt bites”, 2026. www.mining.com
47. U.S. Geological Survey, Mineral Commodity Summaries 2026: Helium. pubs.usgs.gov
48. Korea Herald, “Iran tensions rattle chip supply chains, raise gas risks”, 24 March 2026. www.koreaherald.com
49. Lloyd’s List Intelligence, “Strait of Hormuz brief”, 8 October 2026. www.lloydslistintelligence.com
50. Kpler, “Sulphur & sulphuric acid in 2026: the feedstock crisis cascading through copper, nickel & fertilisers”, June 2026. www.kpler.com
51. CRU BC Insight, “China’s sulphuric acid export halt tightens global supply”, May 2026. www.bcinsight.crugroup.com
52. Mysteel, “China’s sulfuric acid export ban unlikely to suppress short-term copper smelting profit”, May 2026. www.mysteel.net
53. Cochilco (Chilean Copper Commission), Mercado del ácido sulfúrico. www.cochilco.cl
54. ChemAnalyst, “Grasberg mine targets end-2027 restart”, May 2026. www.chemanalyst.com
55. CNAS, “American AI Companies Can’t Get Enough Chips”, May 2026. www.cnas.org
56. Nathan Benaich et al., State of AI Report 2026. www.stateof.ai
57. FT Mercati / ChemNet, “Ajinomoto to cut ABF film supply to China by 30% amid global shortage”, August 2026. www.ftmercati.com
58. Fastmarkets, interview with ARECOMS (DRC cobalt regulator), August 2026. www.fastmarkets.com
59. Mark Maslin, Livia Van Heerde & Simon Day, “Sulfur: A potential resource crisis that could stifle green technology and threaten food security as the world decarbonises”, The Geographical Journal 188(4), 2022, doi:10.1111/geoj.12475. doi.org
60. Brian Potter, “Does all semiconductor manufacturing depend on Spruce Pine quartz?”, Construction Physics. www.construction-physics.com
61. Ministry of Commerce of China, spokesperson remarks on rare-earth export controls, 13 October 2025 (gov.cn). english.www.gov.cn
62. Joris Teer, “Dismantling China’s Supply-Chain Weapon”, Council on Foreign Relations, September 2026. www.cfr.org
63. State Council of China, Provisions on Industrial and Supply Chain Security (Order No. 834), March 2026 — China Law Translate. www.chinalawtranslate.com
64. BNamericas, “Crisis in the Middle East worsens sulfuric acid deficit in copper mining in Chile”, 2026. www.bnamericas.com
65. East Asia Forum, “Local politics limits China’s rare earth leverage”, September 2026. eastasiaforum.org
66. 36Kr, “Global semiconductor materials market share analysis”, 2026. eu.36kr.com
67. SemiAnalysis, “Lam Research, Tokyo Electron, JSR battle it out in the EUV photoresist market”, 2021. newsletter.semianalysis.com
68. Jing Ge, “China’s tungsten export controls on Japan are backfiring”, 9DashLine, July 2026. www.9dashline.com
69. UNCTAD Global Trade Update, June 2026 (via IPS / Global Issues). www.globalissues.org
70. Cabinet Secretariat of Indonesia, “President Prabowo launches 11 downstreaming projects”, October 2026. setkab.go.id
71. Korea JoongAng Daily, “Korean defense firms quietly bypass China’s rare earth curbs via third countries”, August 2026. www.koreajoongangdaily.com
72. BNamericas, “Chile’s lawmakers push for guardrails in critical minerals deal with US”, October 2026. www.bnamericas.com
73. Simon Evenett & Johannes Fritz, “Revisiting the China–Japan rare earths dispute of 2010”, VoxEU/CEPR. cepr.org
74. Forethought, “Three types of intelligence explosion”, 2025. www.forethought.org
75. gov.cn, “Eighth round of China–U.S. economic and trade consultations yields positive consensus”, 28 September 2026. english.www.gov.cn
76. Al Jazeera, “‘Hostile, but hooked’: what’s behind the US–China trade truce extension”, 24 September 2026. www.aljazeera.com
77. Japan Times, “Japan rare earths mud near Minamitorishima”, July 2026. www.japantimes.co.jp
78. Rio Tinto Nuton. nuton.tech
79. Al Jazeera, “How a 95 percent drop in Hormuz traffic changed global shipping”, 27 August 2026 (more than 100 vessels a day before the war). aljazeera.com
80. Parker Whitfill & Cheryl Wu, “Will compute bottlenecks prevent an intelligence explosion?”, arXiv 2507.23181, 2025. arxiv.org
81. Gundlach et al., on how measured algorithmic progress depends on compute scale, arXiv 2511.21622, 2025. arxiv.org
82. Anson Ho et al., “Algorithmic progress in language models”, arXiv 2403.05812, 2024. arxiv.org
83. Epoch AI, Trends (AI chip stock, energy efficiency, training compute), 2026. epoch.ai
84. Tom Cunningham et al., “The Economics of Recursive Self-Improvement”, author draft dated 13 July 2026; arXiv version September 2026. Author draft; arxiv.org
85. Anthropic, “Measuring the pace of AI development”, August 2026. anthropic.com
86. OpenAI, “Research acceleration: a view inside OpenAI”, September 2026. openai.com
87. Tom Davidson & Tom Houlden, “How quick and big would a software intelligence explosion be?”, Forethought, 2025. forethought.org
88. Konstantin Pilz et al., “Trends in AI supercomputers”, arXiv 2504.16026, 2025. arxiv.org
89. Chan et al., on the risk of a runaway intelligence explosion and biases in past estimates, arXiv 2609.36054, September 2026. arxiv.org
90. William Zhou & Daniel Freeman, “AI models may be capable of self-replicating without access to a datacenter”, Anthropic Frontier Red Team, 12 October 2026. anthropic.com
91. Epoch AI, “The frontier of the Epoch Capabilities Index has grown faster since reasoning models”, 2026. epoch.ai
92. Epoch AI, “Why GPT-5 used less training compute than GPT-4.5 (but GPT-6 probably won’t)”, 2025. epoch.ai
93. TMT Finance, “2026 hyperscaler capex tops US$700bn”, 18 August 2026. tmtfinance.com
94. “Welchia”, an early worm written to remove another worm and patch infected computers, which itself caused heavy network disruption. wikipedia.org
95. Çetin et al., “Cleaning Up the Internet of Evil Things: Real-World Evidence on ISP and Consumer Efforts to Remove Mirai”, NDSS 2019. ndss-symposium.org
96. International Energy Agency, Energy and AI: executive summary, April 2025. www.iea.org
97. Global Energy Monitor, “Leading three manufacturers providing two-thirds of turbines for gas-fired power plants under construction”, August 2024. globalenergymonitor.org
98. International Federation of Robotics, “Global robot demand in factories doubles over 10 years” (World Robotics 2025), 25 September 2025. ifr.org
99. Institute for AI Policy and Strategy, “Semiconductor manufacturing equipment export controls”, March 2026. www.iaps.ai
100. Taipei Economic and Cultural Office, citing ITRI, “Taiwan and the global semiconductor supply chain”, December 2025. www.roc-taiwan.org
101. CTOL Digital, “Counterpoint: Samsung HBM revenue share jumped to 33% in Q2 2026”, 3 September 2026. www.ctol.digital
102. Epoch AI, “AI chip market share by designer”, updated October 2026. epoch.ai
103. Epoch AI, “Introducing the AI Chip Owners Explorer”, April 2026. epoch.ai
104. Stanford HAI, AI Index Report 2026, chapter 1: Research and Development, April 2026. hai.stanford.edu
105. Erik Brynjolfsson, Daniel Rock & Chad Syverson, “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies”, 2021. nber.org
106. James Bessen, “AI and Jobs: the role of demand”, 2018. nber.org
107. Charles I. Jones & Christopher Tonetti, “Past Automation and Future A.I.: How Weak Links Tame the Growth Explosion”, May 2026, version 0.5, especially section 5. web.stanford.edu
108. Council of the European Union, “Critical raw materials explained” (infographic), last reviewed 21 September 2026. consilium.europa.eu
109. European Court of Auditors, Special report 04/2026, “Critical raw materials for the energy transition: not a rock-solid policy”, February 2026. eca.europa.eu
110. Davide Fina, “Europe’s critical raw materials bet: a reality check on the EU’s ambitions”, ODI Europe, 17 June 2026. odi.org
111. European Commission, DG GROW, “Critical raw materials” (2023 list and note on copper and nickel). single-market-economy.ec.europa.eu
112. U.S. Geological Survey, “Final 2025 List of Critical Minerals”, Federal Register, 7 November 2025. federalregister.gov
113. U.S. Geological Survey, Mineral Commodity Summaries 2026 (copper, cobalt, nickel, rare earths and tungsten chapters; 2025 mine production). pubs.usgs.gov
114. China Briefing, “China’s rare earth export controls: impacts on businesses” (updates of 30 October and 7 November 2025 on the scope of the suspension). china-briefing.com