Compute and Poverty

Who Collects the Compute Rent?

7 min read

Data as of October 26, 2026. Every figure links to the page it came from.

In Compute Is the Soil I argue that compute is the new land: a base everyone needs, held by a few, and priced by scarcity more than by the work of the people who own it. George called that kind of income rent. An argument like that should be checked against numbers, so here are the ones I lean on, with the counterpoints that cut against me. I am not accusing anyone. The owners are playing the game as it is written; my interest is in the rules.

The landlords’ margins

The clearest sign of rent is a price that stays well above the cost of supplying the thing. In the quarter ended June 30, 2026, Amazon Web Services earned $16.6 billion of operating income on $42.2 billion of sales, a 39.4% operating margin, up from 32.9% a year earlier. Google Cloud’s revenue grew 82% to $24.8 billion, and its operating income rose from $2.8 billion to $8.8 billion, which by my arithmetic lifts its margin from about 21% to about 36%. Microsoft’s Intelligent Cloud segment made $16.0 billion on $39.3 billion in the same quarter, about 41%.

Upstream, the chip itself carries the fattest margin. Nvidia reported $96.2 billion of revenue, $89.0 billion of it from data centers, at a 75.0% gross margin for the quarter ended July 26, 2026.

What a GPU rents for

If chips were an ordinary manufactured good, I would expect the rent on an old one to fall steadily as newer ones arrive. That is not what happened this year. Ornn’s index of transacted H100 rental prices settled at $3.03 per GPU hour on October 25, 2026, up 6.3% in a month, and posted on-demand rates at four clouds ran from $3.49 to $6.16. Andreessen Horowitz, using Ornn’s data, wrote on October 23 that every model of GPU in the series, “including the oldest A100 models, costs more to rent than it did at the beginning of the year,” and that prices sit 9% to 143% above where six year straight-line depreciation would put them.

Two cautions belong right beside that claim. The same note concedes that A100s and H100s spent more time below the line than above it over the longer run, and an independent reading argues the A100 still rents well below its early prices. Also, a16z crypto led Ornn’s $33 million seed round in June 2026. I take the direction seriously, a scarcity premium this spring, and hold the size of it loosely.

The land under the cloud

The physical ground matters too. CBRE counted primary market vacancy of 1.4% in North America, and 0.2% in Northern Virginia, in the first half of 2026, with more than 80% of the 7,481 megawatts under construction already preleased and asking rents up in every size class. When space is that tight, I suspect much of the gain goes to whoever already holds a powered site.

Land prices say the same thing. The home builders’ trade group, which has its own stake in the fight, tallied deals such as 97 acres in Loudoun County sold for $615 million. That works out to about $6.3 million an acre, against a 2025 median of about $125,000 an acre in the county.

Some of this value already flows to the public. Loudoun’s adopted budget for fiscal 2027 expects $879.1 million from its tax on data center computer equipment and $417.7 million from real property tax on data centers, together about 40% of a $3.22 billion general fund by my count. Notice what that tax mostly falls on. George would tax the location and spare the machines, and Loudoun leans the other way: computer equipment pays $4.15 per $100 of value against $0.805 for real property.

Bar chart of Loudoun County's tax on computer equipment in data centers: 581.6 million dollars in FY 2023, 684.8 million in FY 2024 and 699.8 million in FY 2025 as actual collections, then 794.8 million in FY 2026 and 879.1 million in FY 2027 as adopted budget figures.
Loudoun County’s personal property tax on the computers inside data centers. The last two bars are budgets, not collections. Chart by GANDALFO. Data: Loudoun County, FY 2027 Adopted Budget, General Fund Revenue, pp. 106 to 107.

Power is the bottleneck

A data center is worthless without a grid connection, and that is where scarcity bites hardest. Lawrence Berkeley National Laboratory found that data centers used 176 terawatt hours in 2023, 4.4% of US electricity, and projected 6.7% to 12.0% by 2028. Getting new supply connected is slow: the median power project built in 2025 took five years from interconnection request to operation, against under two years for projects built from 2004 to 2008.

In PJM, the grid that serves Northern Virginia, the capacity auction for 2028/29 cleared every zone at the $325 per megawatt-day cap and still fell 6,831 megawatts short of its reliability requirement. PJM’s independent market monitor calls data center load “the primary reason for recent and expected capacity market conditions” and puts the added cost to other customers at $23.1 billion over three auctions. That is the cause I find most telling: when the scarce input is a grid connection, households pay part of the rent.

Bar chart of PJM's regionwide capacity price by delivery year. It ranges from about 29 to 165 dollars per megawatt-day from 2018/19 to 2024/25, then jumps to 269.92 for 2025/26, 329.17 for 2026/27, 333.44 for 2027/28 and 325.00 for 2028/29.
The price PJM pays power plants to be available, by delivery year. The last four are the highest in the series, and the latest cleared at its cap. Chart by GANDALFO. Data: PJM, 2028/2029 Base Residual Auction Report, July 14, 2026, RTO clearing prices.

Concentration

Synergy Research put the cloud market at $143.4 billion in the second quarter of 2026, up 43%, with Amazon at 28%, Microsoft at 20% and Google at 15%. Epoch AI estimates that five hyperscalers held 71% of the world’s AI compute at the end of 2025, up from 63% in early 2024. Rising concentration while the market grows fast is what I would expect if location and scale, not just skill, decide who earns.

The other side, fairly

There is a strong case against my reading. The price of what compute produces is falling fast: Epoch found the cost of reaching a fixed level of model performance fell between 9 and 900 times a year, depending on the benchmark and the level chosen, and the performance per dollar of new AI chips has grown about 49% a year. Users are getting a great deal more for their money.

High margins can also be a return on enormous risk rather than rent. In the same quarter, Alphabet spent $44.9 billion on property and equipment against $39.1 billion of operating cash flow, and Amazon spent $54.2 billion against $45.4 billion. New entrants are arriving as well: Synergy now counts nine neoclouds in the top 40. If the buildout overshoots, today’s scarcity premium could turn into tomorrow’s write-down.

What this does and doesn’t show

It shows high and rising margins at the chip and cloud layers, rental prices for older chips that held up this year, very tight space, land selling far above its old use, and a grid where new data center load has raised prices for everyone. It doesn’t show how much of those margins is pure rent rather than a return on capital and skill; company accounts don’t split them that way. The GPU price series are short and come from a young index. The land figures are deals, not a measured price index.

What would change my view? Several years of falling real GPU rents, cloud margins drifting back toward ordinary levels as neoclouds take share, interconnection times shrinking, or a fall in concentration as the market grows. If those come, I will say so here.

Sources

Chart credits

  1. Loudoun County tax on data center computer equipment. Chart by GANDALFO from Loudoun County data. loudoun.gov. Drawn from published figures.
  2. PJM capacity prices by delivery year. Chart by GANDALFO from PJM data. pjm.com. Drawn from published figures.