Compute Supply and the Cost Line: Can Supply Keep Up, and When Does It Pay Off

The solid line is actual: from $1,320 in the fourth quarter of 2025 to $2,335 in the second quarter of 2026, coinciding with public-cloud GPU price increases and the B300 markup; a surging gauge means supply short of demand. Dashed lines are projections: scenario revenue divided by Epoch's data-center completion supply. The green shading is the zone above breakeven: its floor at $4.5k is the breakeven of a new site with depreciation, and the $6.5k line is the full annualized cost of ownership; the thick orange dashed line at $5.6k is the key threshold, beyond which new compute earns a positive return at a 10% cost of capital; the diamond marks when the base case crosses. The numerator is model companies' ARR, not the rent cloud providers collect.

section 1 run backwards: each row is a date, “compute at that date” is market-wide H100e under Epoch’s known-site schedule, and each threshold column is compute × threshold, the combined ARR of the five companies needed to touch that line; the “base-case ARR” column is the base case used for the dashed line in section 1. How to read it: by mid-2027 compute is about 60 million chips, so ARR needs $268B to reach $4.5k and $333B to reach the $5.6k key threshold, while the base case at $239B is still short; by the end of 2027 the base case of $372B against $375B just touches $4.5k; by the end of 2028, $711B against $726B is a hair below $5.6k. The table uses the lower edge of the supply band (known sites), so the more supply comes online, the more ARR it takes to touch a line.

The average yardstick = total ARR ÷ cumulative compute; the marginal yardstick = new ARR in the quarter ÷ new compute in the quarter, increment against increment with no stock inertia, so it turns before the average does. On history: in the fourth quarter of 2024 the marginal figure was only $52, as new compute went into training and monetization lagged; by the fourth quarter of 2025 it was back to $1,836 and above the average, so while the market read percentages as stagnation the marginal curve was already announcing 2026; in the second quarter of 2026 it was $4,488, 1.9× the average, and higher again in the third. Marginal above average means the squeeze continues; falling below average is the early warning of a top. The latest quarter includes a run-rate estimate and is drawn translucent.

A second check on section 1 in dollars: part of the rise in revenue per chip comes from chips getting dearer, so only if the dollar-based ratio is still rising has demand really caught up. Left: market-wide cumulative chip spend against the combined run rate of five model companies. Right: the monetization rate = annualized revenue ÷ cumulative chip spend, how much annualized revenue each dollar of chip capital supports. Reading: with the price per unit of compute no longer falling, a monetization rate still sliding would be evidence of oversupply, revenue failing to keep up after price increases; a rising rate means demand is keeping up. Today the price has stopped falling and the monetization rate is still rising.

This is the key chart for the whole supply side. IT power coming online each half year, stacked by owner; past dates are verified from satellite imagery, future dates are Epoch’s estimated completions. Every half year from 2026 to 2028 brings 3 to 7 GW online, and in the second half of 2028 Oracle alone is most of it. The drop after 2029 is because Epoch only records sites that have broken ground, with two to three years of visibility, not a real supply cliff.

Left: the sum of 71 tracked sites, where a steeper slope means faster build-out; the dashed line is cumulative capital cost. Right: the ten largest sites by peak power, each line one site’s IT-power build-out, with future dates as planned values. Use the left chart for the pace of supply growth and the right one for individual projects.

Left: new H100e compute each quarter stacked by chip model, with the B300 now the largest block. Right: Epoch’s cost estimates summed, which is the implied chip purchase spend. Coverage follows Epoch’s own description: only Nvidia, Google TPUs, Amazon Trainium, AMD Instinct and Huawei Ascend, which Epoch considers the large majority of global AI compute; custom chips from Microsoft, Meta and Tesla and consumer GPUs are not included. Nvidia data starts in 2022, the other four only in 2024; Nvidia estimates are the most confident, Amazon the least. Some quarters are incomplete and drawn translucent, so the rightmost quarters are understated rather than falling demand.

Left: by user attribution of Epoch's data centers, the trajectory of capacity online plus planned completions, where the slope is how hard each player is grabbing compute; "unlabeled" is mostly sites with undisclosed customers, and coverage thins after 2028. Right: cumulative H100e stock by owner through the first quarter of 2026, in-house chips included (Google TPU and Amazon Trainium under their own names), showing how capex turns into compute assets and who is gaining: Oracle and xAI surging, China held back by export controls.

Left: owners' share of cumulative stock as of the first quarter of 2026, in-house chips included. Right: users' share, counting only capacity online, not future plans. The two differ in definition: owners are who bought the chips, users are who runs the data centers.

Left: compute cost as a share of revenue; above 100% means compute alone exceeds all revenue, and dropping below 100% is the turning point of the business model. Anthropic is around twenty percent, OpenAI still above seventy in 2025. Right: daily tokens processed, on a log axis by default: a straight line means a constant growth multiple, and the Google line flattening means the yearly multiple fell from 50× to 7× while volume itself keeps compounding, which the linear view shows; the gap between the explosion in volume and revenue growth is the revenue deflation per token. Google discloses all-products volume once a year (3.2 quadrillion a month in May 2026, about 107 trillion a day) and first-party API volume every earnings call (22 billion a minute in the second quarter of 2026, about 32 trillion a day); OpenAI’s API ran at 15 billion a minute in April 2026. Scopes are drawn separately and absolute levels are not comparable across lines; new US company disclosures will be added as they come.

Left: the blended purchase price = market-wide chip spend in the quarter ÷ new H100e in the quarter, down from $38k in 2022 to $13.5k in the third quarter of 2025, then the decline stopped and the price rose to $14.1k in the first quarter of 2026, consistent with B300 volume and HBM price increases; the dashed line is the coverage-adjusted series. Right: price per unit of compute by chip generation; B200 to B300 is the first time in the data that a new generation costs more per unit of compute; H20 and Ascend look expensive because their compute is constrained, not because they are premium.

Based on Epoch AI's total-cost-of-ownership model for a 1GW data center: $38B of upfront capex, $0.9B a year of operating cost, $8.5B a year of annualized total cost, with servers at 60%; five-year IT equipment life, fourteen-year buildings, 71% utilization, a 10% discount rate. Converted per H100e: 1GW is about 1.24 million chips, existing sites cost $30.6k per chip, new-generation sites only $13k to $19k. The table lists the formula, key assumptions and range for every threshold line.