Big Tech Climate Pledges Skip the Barrels AI Helps Pump

A peer-reviewed study finds AI that helps oil companies raise output could add 3.3 to 13.3 times the carbon of AI data centers. Microsoft and Google would not discuss the paper. Amazon did, and talked about carbon per barrel.

The findings landed in August in npj Climate Action. Questions went to the three cloud vendors on August 19. The gap they were asked about is still not on their climate ledgers.

The Study Puts Oilfield AI Above Data Centers

Holly Alpine and Will Alpine, former Microsoft staff who now run the Enabled Emissions Campaign, modeled AI as a productivity shock on both fossil fuels and renewables. The paper, written with two other researchers, treats AI as a tool that can cheapen oil recovery and also tighten wind and solar operations, then asks which effect wins in an economy still running mostly on fossil fuels.

Under parallel adoption, the model produces a 0.47 to 1.8 gigatonnes a year net rise in CO2, or 1.2 to 4.8 percent of 2024 energy-related carbon. Enabled emissions beat avoided emissions whenever fossil-sector gains are above zero. Net cuts show up only if renewables get a 4 to 5 times larger productivity lift than oil and gas.

AI CARBON, TWO LEDGERS

Fossil fuels still supply about 80 percent of primary energy, the authors note, and oil and gas output would drop about 8 percent a year without constant reinvestment. AI that lowers finding costs and lifting costs is, in that setting, a way to keep fields economic. The paper calls this a directional result across 64 scenarios, not a single-year forecast.

Silence From Microsoft and Google

The companies were asked whether they accept the findings, whether they count carbon their tools help unlock, and whether they would limit contracts meant to expand extraction. Google confirmed the questions, then stopped answering. Microsoft’s media team said the company has nothing to share.

Alpine, who used to manage sustainability work at Microsoft, said the quiet is the tell. In earlier rounds, she said, the firms pointed to responsible-AI write-ups, low operating emissions, or greenhouse-gas reporting rules that do not require enabled carbon to be counted.

Their silence tells you more than their talking points ever could.

Holly Alpine, co-author, Enabled Emissions Campaign

Microsoft still aims to be carbon negative by 2030. Melanie Nakagawa, its chief sustainability officer, wrote that the systems it builds “must also support the long-term health of the planet.” Google aims for net zero across operations and its value chain by 2030. Its AI principles say it will mitigate unintended or harmful outcomes. None of those texts names custom oilfield software as an out-of-bounds use.

The public argument about AI and climate still orbits power plants beside server halls, water for cooling, and tax breaks. That fight is loud because you can photograph a substation. Extra barrels from a seismic model do not show up on a campus tour, and they do not hit Scope 2.

Amazon Talks Intensity, Not Total Pollution

Amazon was the only firm that sent a written reply. It did not contest the paper’s ranges. It also did not say whether AWS tools are raising oil and gas output. A spokesperson said energy firms should get the same cloud kit as everyone else, and that AWS will keep serving “legacy businesses” to make them less carbon intensive while also speeding renewable work.

The energy industry should have access to the same technologies as other industries. We will continue to provide cloud services to companies in the energy industry to make their legacy businesses less carbon intensive and help them accelerate development of renewable energy businesses.

Amazon spokesperson, emailed statement

Alpine said no one had asked for a ban on energy customers. The live issue is application, whether the tools are used to find and produce more oil and gas, cheaper than the same wells would have been without them. Amazon’s oil and gas pages still sell that job. Explorers are told they can pull insights faster, identify reservoirs faster and cheaper, and reduce the cost per barrel.

WHAT AWS STILL PROMISES UPSTREAM

• Time to first oil: Machine learning on seismic data is pitched to speed exploration, drilling, and production.
• Reservoir work: High-performance computing is sold to run models faster and optimize extraction.
• Lifting cost: Automation and global scale are framed as a way to protect profit when budgets shrink.
• Drilling risk: Sample uses include fewer drilling hazards and more productive wells.

Those lines sit on a current industry page, not a leak. Amazon also said it supports its own sustainability programs and works with partners to cut demand for carbon fuels. Alpine called those programs real and beside the point, because they do not touch supply growth that AWS helps finance through faster, cheaper barrels.

What Facilitated Emissions Would Capture

Ask a climate staffer how a cloud company is doing and you will hear matching, carbon-free energy percentages, and net-zero years. Microsoft says it matched 100 percent of its electricity with renewables in its latest reported year. Google reports about 65 percent carbon-free energy on an hourly basis across data centers and offices in 2025. Amazon says it matched 100 percent of electricity with renewables in 2025, the third year in a row. Those are operating stats. They do not score a model that shaves months off a seismic workflow for an oil company.

CARBON INTENSITY

• The metric: Carbon intensity is pollution per barrel, per dollar, or per unit of energy, not the total tons released.
• The dodge: Intensity can fall while total output rises, so a firm can claim progress as more oil reaches the market.
• The pledge hole: Net-zero programs built on a company’s own stacks can stay on track while client production, and client carbon, grow.

The Greenhouse Gas Protocol, which writes the corporate carbon rulebook most of these pledges lean on, has been drafting a new Scope 3 bucket for this class of problem. Working-group slides from July 24, 2025, describe an optional Category 16 for facilitated emissions. A facilitated activity, in the draft, is a third-party emitting source that a company’s services enable or influence, that the company does not own, and that still pays the company.

The full technical working group backed writing that category down. The draft still says a company may report it. The worked examples are insurance, underwriting, and other financial services. Oilfield AI is a clean fit for the definition and is not named. Until that line is required, a cloud vendor can tell the truth about warehouse power and say nothing about barrels.

Oil Advocates Treat Faster Drilling as the Win

The American Petroleum Institute, the U.S. oil trade group, disputed the idea that more energy and lower emissions are in conflict. That is the industry’s core reply: denser fuels, more output, and AI as a general accelerator, not a climate choice. On energy forums the same week the paper spread, the attack was blunter. The study, in that telling, wants to punish software for doing its job, because cheaper oil is treated as a harm.

The paper already grants the mechanism those critics like. AI does not pick sides. It raises productivity where it is installed. The result in the model is that fossil incumbency holds unless clean-power gains run several times hotter, or policy steers the tools. Authors also note that without that steering, AI raises the carbon intensity of the whole economy. Efficiency at the well is not the same fact as less carbon in the air.

WHERE THE ARGUMENT SPLITS

• The paper: Extra oil and gas from AI outweighs AI’s help to renewables in every case where fossil gains are not zero, and that extra carbon dwarfs data-center power.
• Oil groups: More energy and lower emissions need not clash, and a tool that cuts cost per barrel is a production win, not a climate case against the vendor.
• The vendors: Microsoft and Google will not engage the finding. Amazon talks intensity, equal access, and its own net-zero plan, and does not give a barrel count.

The two loops now run together. Chevron said it would build a 2.67 gigawatt gas plant under a 20-year deal to feed a Microsoft data-center campus in Reeves County, Texas, with first power expected in 2028. Cloud firms sell AI into oilfields and buy gas to keep GPUs on. Accounting still treats those as separate stories.

Dates, Stacks, and the Barrels Left Out

Amazon was asked whether the paper should unsettle its Climate Pledge, the 2019 vow of net-zero carbon by 2040. A spokesperson said the pledge has not changed, that the company is “being transparent about the challenges,” and that “growth and sustainability don’t have to be in opposition.” Alpine said that sentence only works because the books were never drawn to see this part of the business. The pledge tracks Amazon’s operations. It does not add the oil a customer produces with AWS.

She does not want the next move to wait on a press shop. The first levers she names do not need Microsoft or Google to agree.

THREE PLACES THE LEDGER COULD CHANGE

• Investors: Ratings files can ask whether a tech firm’s AI or cloud work is expanding a client’s fossil output, not only what the tech firm emits itself.
• Accounting bodies: The live Scope 3 rewrite can put facilitated emissions on a required line instead of an optional one, and can say whether oilfield software counts.
• Policy: Disclosure rules can force a vendor to say if a system it builds or licenses is meant to expand or extend fossil production.

That is not automatic. It starts with treating enabled barrels as a number someone has to own. Microsoft, Google, and Amazon can decline a journalist’s list. They have a harder time declining the same question from staff, customers, and the people who hold the stock, if the question is specific: how much oil and gas production is the software enabling, and what, if anything, will change.

Frequently Asked Questions

Which journal published the AI oil emissions study?

It ran in npj Climate Action, a Nature Portfolio journal, as volume 5, article 71, on August 4, 2026, and is open access. The team used a global computable general equilibrium model, a whole-economy simulation that traces how a productivity shock in one sector moves prices, output, and fuel use.

Who wrote the enabled emissions paper besides Holly and Will Alpine?

Independent researcher Nathan Geldner and Maksym Chepeliev of Purdue University are co-authors. The Alpines founded the Enabled Emissions Campaign in 2024 after leaving Microsoft, and the campaign is listed as their affiliation on the paper.

Do current climate rules make cloud companies report oil their AI helps expand?

Not as a required line item. The GHG Protocol draft would let firms report facilitated emissions under Category 16, and the draft treats that category as optional except where another industry standard already demands it. Worked examples on the table are insurance, underwriting, and other financial services, not oilfield software.

What extra climate dates sit behind Microsoft, Google, and Amazon’s headlines?

Microsoft also says that by 2050 it will remove the carbon it has emitted, directly or from electricity, since 1975. Google’s 2030 net-zero aim includes a planned 50 percent cut in combined Scope 1, market-based Scope 2, and Scope 3 from a 2019 base. Amazon describes 2040 as 10 years ahead of the Paris Agreement’s mid-century mark, and the pledge is limited to its own global operations.

How do 2035 data-center forecasts change the study’s comparison?

The authors also stack enabled emissions against the IEA’s projected 2035 data-center emissions of 0.3 to 0.5 gigatonnes a year. On that later, higher server baseline, enabled oil-and-gas carbon is still 1.2 to 8 times larger in the paper’s figure.