Early this month Anthropic posted a job listing for a “Gas Supply and Transport Lead”, the first dedicated owner of natural gas for its data centres in the United States and Canada. The listing, reported by Business Insider on 13 November, offers $320,000 to $405,000 and contains a plain description of a dependency. “Some of our campuses will run on gas-fired, behind-the-meter generation,” it says. “At those sites fuel is the power supply: if gas does not arrive, in the volume and at the pressure the plant needs, the compute does not run.”

The sentence is a small piece of engineering candour, and it sits in a larger argument about how AI companies should get electricity. A preprint posted on 12 November by Alex Newkirk and colleagues at Carnegie Mellon University and the University of California, Berkeley, takes up that argument with a model. The paper has not been peer reviewed. Its claim is that the price of a kilowatt-hour is the wrong number to compare, because the number that decides the outcome is how long the chips wait for power.

Why waiting is the main cost

The authors start from the bottleneck. Grid interconnection queues, they write, exceed five years in some regions, and the transformers, breakers and electrical steel that a connection needs often have multi-year backlogs. New gas turbines have waiting lists the paper puts at four to seven years. Firms have responded with exclusive power purchase agreements and with generation built on site, sometimes fully islanded from the grid.

The paper compares four ways of powering a facility of 100,000 GPUs, about 131 megawatts at design load: solar panels with batteries wired together as alternating current, the same with direct current, islanded natural gas, and a grid connection. Direct-current coupling joins panels and batteries before the power is inverted, which removes inverters and some conversion losses. The chips themselves already run on direct current. Across the continental US the model finds this saves 18 percent of lifetime energy cost on average, about 9 cents per kilowatt-hour. For a facility in the Permian Basin of West Texas it gives $0.27 per kilowatt-hour against $0.33 for the alternating-current design, a difference of $570 million in present value.

The central step is a new measure the authors call speed-adjusted cost of energy. Each month a finished building waits for power is a month in which its GPUs earn nothing, and the model charges for that idle time at $2.40 per GPU-hour, an H100 rental index price that the authors take as the floor value of a chip-hour. They give solar and batteries a 24-month timeline, taken from an earlier industry report. Gas plants get 40 to 42 months of construction, drawn from a federal cost study, with turbine delivery assumed to take 36 months. Grid connection takes the median of each region’s past queue, between two and a half and five years depending on the state.

On energy cost alone, direct-current solar is rarely the cheapest option. Once the idle months are priced, it is the lowest-cost option in 84.9 percent of the locations modelled. Grid power wins where cheap industrial electricity meets a fast queue: the authors name Maine, where the grid is the fastest in the country to connect, and East Texas and Georgia, where industrial rates are below 6 cents and interconnection takes three years. Gas wins in few places in the baseline, because places where gas is cheap tend to have cheap grid power as well.

How far the result stretches

The conclusion depends on two assumptions, and the authors test both. If a chip-hour is worth less than $2.40 to its owner, waiting costs less. They find the value at which a firm would be indifferent between solar and the grid is a median of $1.61 an hour, and that most of the country sits below the $2.40 baseline. If gas plants were running in three years, which the authors say would need idealised parallel construction or turbines delivered well ahead of schedule, solar would still lead in about 55 percent of locations, gas would lead in about 35 percent, mostly in solar-poor regions of the upper Midwest, the Appalachian Basin and the Pacific Northwest, and the grid would lead in about 9 percent. Gas also gets cheaper with size. Small plants of 10 to 20 megawatts cost 25 to 30 cents a kilowatt-hour, and in the Colorado example the paper shows, gas matches direct-current solar at around 30 megawatts.

Reliability is where the two bets differ most. The model sizes every system to serve the full load in 99 percent of hours. For solar, the authors find that cost follows the worst five-day stretch of low sunlight far more closely than the annual average: a dry Colorado site and a fog-prone northern California site with similar yearly insolation come out at $0.21 and $0.61 a kilowatt-hour. A gas plant faces a different weather risk, in the pipeline, which is what the Anthropic listing is about. It asks the new hire to design “second pipeline feeds, storage, backup fuel, cold-weather performance, and curtailment priority”.

The limitations the authors list are substantial. They model grid connection as buying power at industrial rates after the wait and leave out upgrade costs that utilities may pass on to other customers, which makes their grid figures a lower bound. The direct-current equipment they assume is not yet mature: design standards are not codified, some fault-protection gear is not yet on the market and the skilled workforce is built around alternating current. The authors say early projects may capture only part of the savings. They assume land can be acquired without delay for solar fields of hundreds to thousands of acres, and they note that a longer approval timeline for solar than for gas could change the result. Their gas options are combustion turbines, so an order of engines with different delivery times falls outside what they test.

What the buyers are doing

Anthropic’s own filings and hiring show the gas bet. Business Insider reports that its IPO prospectus says it expects to compete for natural gas in some areas where it is helping to build data centres, and that it has committed about $900 billion to infrastructure, about half of it due by 2030. The Financial Times reported this week, in Business Insider’s account, that the company is spending $3.4 billion on a 1.7-gigawatt batch of gas-fired engines from a Texas supplier. The prospectus warns that dependence on gas exposes it to cost swings and delays, and that environmental standards could change amid “shifting public and political views regarding climate change.”

Two other recent events frame the paper’s numbers. On 10 November CoreWeave’s chief executive told investors it was selling short-term capacity at more than $40 million per megawatt a year, against $20 million to $25 million for four- and five-year contracts. And on 3 November voters in more than 30 localities across nine states backed moratoriums, bans and other limits on data centres, according to Semafor. The authors treat land cost as negligible, but they concede that siting fights could lengthen timelines unevenly, and solar fields need far more land than gas plants do.

The authors end by calling for measurements. A direct-current campus of 100 megawatts or more has not been built and run, and until one is, the paper’s savings are expected values. The job listing keeps a narrower question. Anthropic asks its first gas lead to find out, early in site selection, “which sites have a real fuel path and which do not.”

References

  • Anthropic (2026). Gas Supply and Transport Lead, US and Canada (job listing). job-boards.greenhouse.io
  • Baranko, K., Campbell, D., Hausfather, Z., McWalter, J., & Ransohoff, N. (2024). Fast, scalable, clean, and cheap enough: How off-grid solar microgrids can power the AI race. offgridai.us
  • Bellan, R. (2026). CoreWeave turns to customer prepayments, shorter contracts to fund AI growth. TechCrunch, 11 November. techcrunch.com
  • Capelouto, J. D. (2026). US voters roundly reject data centers in ballot measures. Semafor, 4 November. semafor.com
  • Council, S. (2026). New Anthropic job listing offers $405,000 to land natural gas for its post-IPO data center buildout. Business Insider, 13 November. businessinsider.com
  • LaRose, A., Diefenderfer, J., Bowers, R., & Vincent, N. (2024). Capital Cost and Performance Characteristics for Utility-Scale Electric Power Generating Technologies. U.S. Energy Information Administration. eia.gov
  • Newkirk, A. C., Gerber, D. L., Fuchs, E. R. H., Hubbard, A., Morgan, M. G., Shehabi, A., Smith, S. J., & Samaras, C. (2026). Speed-to-Power Dominates Energy Costs for AI Data Centers: Technoeconomic Analysis of Powering AI in the Continental United States. arxiv.org
  • Rand, J., Manderlink, N., Gorman, W., Wiser, R. H., Seel, J., Kemp, J. M., Jeong, S., & Kahrl, F. (2024). Queued Up: 2024 Edition, Characteristics of Power Plants Seeking Transmission Interconnection As of the End of 2023. escholarship.org