AI Stack Concentration Dashboard

AI capability rests on a stack of inputs. The models people use sit on top. Under them are the chips and the compute, under those the power and the data centers, and under everything the physical materials. Each layer depends on the ones below it, so concentration low in the stack becomes leverage over everything above. Other projects already map the upper layers: Epoch AI on compute and training runs, the Stanford AI Index on capability and investment, and CSET at Georgetown on chips. The question we start from is one Andy Hall posed: how many chokepoints are there in the stack, and can we measure each one? The first version, called v1 now spans the whole stack: the power and materials layers at the bottom, where the build-out binds first and the least mapping had been done, up through the compute-and-chips layer, where the sharpest chokepoints sit, to the models layer at the very top, where the frontier is trained. The measure is the same at every layer, so the layers can be compared.

Every chart carries a confidence badge, defined in the methodology below: measured (reported as published), derived (a mechanical transform with no added assumptions), modeled (depends on assumptions), or proxy (a stand-in for data we cannot measure directly, with the gap stated). Hover over any badge to see its source.

Power: demand is climbing again

For about fifteen years, U.S. electricity demand barely moved. Efficiency gains kept pace with economic growth, and monthly generation sat on a flat plateau from the mid-2000s into the late 2010s. That has changed. Output has pushed above the old ceiling, and EIA's short-term forecast has it rising further into 2027. New data centers are one of the reasons analysts give for the turn, alongside electrified transport and re-shored manufacturing. We show the measured turn here. The data-center share of it is a separate, modeled question, and it is the one we take up next.

Data centers: how much of the demand is AI

This is the one number on the page that is explicitly about AI, and it is modeled, not measured. LBNL's 2024 report puts data centers at 4.4% of U.S. electricity in 2023, with their total consumption having tripled from 58 to 176 TWh between 2014 and 2023. For 2028 the report gives a wide range rather than a single figure: between 6.7% and 12% of national electricity, depending on how fast AI accelerators are deployed. We show that range as a band, because the range is the honest answer.

Materials: two supply chains beneath the stack

The materials layer is the least mapped part of the stack, and it is not one thing. It splits into two supply chains that feed different parts of AI. The compute chain is what the chips are made from: silicon refined into wafers, gallium for the wider power and radio-frequency electronics, and tungsten in the interconnects and the tools that machine the hardware. The energy chain is what powers, builds, cools, and backs up the data centers: copper for the wiring and the grid, rare earths for the magnets in motors and generators, and cobalt and graphite for the batteries that carry a site through a grid dip. We measure each one the same way, with the share of world production held by the largest producer and a concentration index across all of them, reported as concentration facts with no AI-demand attribution. The figures are annual USGS estimates for 2020 through , and each year shows USGS's latest revision of that year. The chart below tracks the whole run.

Pick a year and every number below follows it, from the cards to the country charts. It starts on the newest year. The chart above always shows the whole run.

The energy chain: what powers, builds, and cools the data centers

The energy chain is the larger of the two and the one where the buildout binds first. Copper is the backbone, and its chokepoint is refining rather than mining. The battery and magnet minerals beneath it are shown at the mine stage only. Refining, magnet making, and battery cells are more concentrated still, and a later version will add those stages.

Copper: the leverage is in refining, not mining

Copper ore is spread across many countries (Chile, Peru, the DRC). Refined copper is not, and the gap is widening: China's share of world refining has climbed every year of the series, from in to in .

The two Herfindahl-Hirschman indices show where the leverage builds up: mining is unconcentrated, while refining sits in the moderately concentrated band. The U.S. DOJ thresholds at 1,500 and 2,500 are marked for reference.

The copper price is market context. It reflects global supply and demand, and we make no claim that AI causes it.

Rare earths: the magnets that move the heat

Rare earths are the permanent magnets in the motors, generators, and the fans that push heat out of a data center. China mined about of the world's supply in , up from in 2020, for a production HHI of out of 10,000. This is the mine stage only. Separation and magnet making sit even more tightly in China, and a later version will measure them.

Cobalt: the batteries that hold a site up

Cobalt goes into the lithium-ion batteries that keep a data center running through a grid dip. Here the concentration is geographic rather than Chinese: the Democratic Republic of the Congo mined about of the world's cobalt in , up from in 2020, for a production HHI of . As with the other battery minerals, this is the mine stage; refining is more concentrated and is not shown here.

Graphite: the other half of the battery

Graphite is the anode in those same batteries, the counterpart to cobalt. China's share of world natural graphite mining has moved between and over the series, at in , for a production HHI of . China's hold on anode-grade processing is larger still, which these mine-stage figures do not capture.

The compute chain: what the chips are made from

The compute chain is shorter but no less concentrated. Gallium is the extreme case, close to a single-country monopoly for the whole series. Silicon and tungsten sit behind it, each with most of the world's supply held in one place.

Gallium: a single-country chokepoint

For gallium we make no AI-attribution claim. We simply document the concentration, and it is the most extreme case in the materials layer. China's share of primary production has stayed near total across the series while world output roughly tripled. USGS's estimate rounds China's output to the entire world total, so the share reads ; small producers elsewhere disappear into the rounding. The production HHI is out of 10,000.

Silicon: the base of every wafer

Silicon metal is the feedstock refined into the polysilicon and wafers every chip is built on. China made about of the world's silicon metal in , for a production HHI of . Its series starts in 2022: earlier USGS editions report silicon metal only combined with ferrosilicon, so there is no clean earlier number to show.

Tungsten: the interconnects and the tools

Tungsten runs through the chip itself, in the interconnects, and through the hard tooling that machines semiconductor and data-center hardware. China mined about of it in , down from in 2020 and the one materials series that has drifted lower, for a production HHI of .

Compute and chips: a chokepoint at every step

Above power and materials sits the layer that turns them into working AI: the chips themselves. Read this layer differently from the ones below it. Minerals and power are metered production, counted and reported. Here the numbers are market-share estimates from industry analysts (TrendForce, Counterpoint, Yole, CSET), so most of them are modeled rather than measured, they are softer, and they shift from quarter to quarter. Every card shows its confidence badge and its source, so you can see exactly how firm each number is. One of them, ASML's hold on EUV machines, is not an estimate at all but a structural fact: no other company on earth builds them. The chain runs from the software that designs a chip to the accelerator that ships, and there is a chokepoint at every step.

One scale across the whole stack

Where a market's firm-level shares are public, the same concentration index built for the minerals applies to the chips. Two compute stages, EDA software and HBM memory, land on the identical zero-to-ten-thousand HHI scale, right next to copper, gallium, and the rest. HBM's three-firm memory oligopoly turns out to be about as concentrated as the rare-earth mines. This is the comparison the dashboard was built to make: one measure, one scale, from Congo cobalt to Nvidia silicon. The materials bars are measured production and the compute bars are modeled market shares, so read them as neighbours, not identical twins.

Models and capability: the stack opens, then closes again

At the very top of the stack sits the AI people actually use, and here the pattern of the whole dashboard finally breaks. Every layer below has a chokepoint held by a few firms or a single country. But by developer, the frontier is the least concentrated layer in the stack: dozens of labs now train frontier-scale models, those above 10²³ FLOP of training compute, and no one of them dominates. Its developer concentration index is only about , well under the level economists call unconcentrated. In the whole-stack chart above it is the shortest bar.

And yet, step back from firms to nations and the openness vanishes. Almost every one of those labs sits in one of two countries. The United States and China together make about of all frontier-scale models, national concentration is climbing, and in China passed the United States for the first time, to , up from a few percent in 2022. The number of frontier-scale models has itself gone from about ten a year in 2021 to in .

The demand side: what the market actually runs

Who trains the frontier is one question; which models people actually run is another. There is no clean public measure of total AI usage, so this is a proxy: OpenRouter, the largest open model-routing marketplace, publishes the token volume each model handles. Read it for exactly what it is. This is the developer and open-model market; it does not capture the traffic that flows straight to ChatGPT, Claude, or the Gemini app, where U.S. closed models dominate, so it overstates open and Chinese models. With that caveat stated plainly, the finding still lands: Chinese-developer models take about of the tokens routed through OpenRouter in the week shown, against about for U.S. models. China's rise is on the demand side too, not only in who builds the frontier.

Methodology

Every observation is checked against metrics.yaml, our methodology spine, before it reaches this page, and every download is logged with a SHA-256 hash in data/provenance.csv. The materials figures are annual series, 2020 through , assembled from successive editions of the USGS Mineral Commodity Summaries; for each year we use USGS's latest revision of that year's figure, so final numbers supersede first estimates. Concentration measures use the published world total as the denominator and pool unlisted producers into a single "Other" bucket. That makes every figure a conservative lower bound on the true concentration. v1 now spans all four layers of the stack: power and materials are metered production (measured and derived); the compute-and-chips layer is analyst market-share estimates (modeled), badged as such; and the models layer is a measured model census from Epoch AI. The confidence badge on every chart says which is which. Later versions will add grid-interconnection queues, a bounded copper-demand range for AI data centers, the refining and processing stages of the battery and magnet minerals, and firm-level HHIs for the compute stages that are top-share-only today.