AI Future Fridays: AI vs the Warming Sky

Welcome to AI Future Fridays. Every week this column takes one force shaping the near future, tells you the truth about it, and hands you something to build before Monday. This week it's the strangest reversal in climate work: the sky stopped being a place where things disappear.

For two centuries the atmosphere was the perfect dumping ground because nobody could prove what you put there. That's the thing that broke. Machine vision on orbiting spectrometers now attributes a plume to a specific piece of equipment, on a specific afternoon, with a number attached. You're not a bystander in that shift either, because the same class of tooling that catches a leaking pipeline will read your own utility bills and tell you where your money goes up the chimney.

The Methane Signal From a Permian Basin Pipeline

Start with the smallest possible version of this story, because the small version is the one that proves the machine works end to end.

On October 9, 2024, a satellite called Tanager-1 passed over the Permian Basin in West Texas and saw something nobody on the ground had reported. The nonprofit Carbon Mapper, led by CEO Riley Duren, ran the detection through its pipeline and pinned the plume on a gathering line venting roughly 7,000 kilograms of methane an hour. Every hour it ran, that leak did the climate damage of driving 47 gas cars for a year. Carbon Mapper told a state agency, the state told the operator, the operator sent a crew, and a later pass from the same satellite showed clean sky.

Nobody was caught. Something was found. That distinction matters, and it's most of the reason this story is a hopeful one.

Tanager-1 was built by Planet Labs using instrument technology from NASA's Jet Propulsion Laboratory, and the detection itself is an inference problem: a hyperspectral sensor measures how sunlight is absorbed on its way back up through the air, and a model separates the methane signature from dust, cloud, glare and surface chemistry. Doing that at 30 meter resolution across a continent, fast enough to matter, is the part that was impossible ten years ago. In its first year in orbit the satellite published 5,392 methane plumes across dozens of countries, and Carbon Mapper quantified a running total of about 141 million kilograms of carbon dioxide equivalent per hour from the sources it found.

Where AI Is Already Cutting Emissions in 2026

The Permian leak isn't a one-off demo. Three fronts have gone from research to operations in the last two years, and they're worth separating because they fail and succeed for different reasons.

• The grid: forecasting models make wind and solar dispatchable enough to trust, and fault detection cuts the time a line stays down. The IEA puts the outage reduction at 30% to 50% and estimates up to 175 gigawatts of transmission capacity could be unlocked by sensors and smarter management without stringing a single new line.
• Methane: super-emitters are a small number of sources doing a large share of the damage, which makes them the rare climate problem where detection is most of the fix. Carbon Mapper's public data portal and the inventories at Climate TRACE both publish facility-level numbers that regulators, journalists and neighbors can pull up on a phone.
• Wildfire: on July 7, 2026, three FireSat satellites reached orbit from Vandenberg. The constellation is designed to spot a fire as small as five meters square, and the eventual 50-satellite fleet is meant to revisit everywhere on Earth every 20 minutes. Google Research helped design the sensor, Muon Space builds the hardware, and the nonprofit Earth Fire Alliance runs the program.

What these three share is that the AI part is boring and the consequence isn't. No agent negotiates climate policy. A model looks at a very large amount of noisy sensor data and says "that pixel is not like the others," and a human with a truck does the rest. It's the same shape as the local-first tooling we walked through in last week's edition on running models you own: unglamorous inference, pointed at a problem where speed is the whole value.

The Backstory of Climate Models and the 1979 Charney Report

Every future has a lineage, and this one's is older and more flattering to science than the public argument suggests.

In the summer of 1979 a group of scientists gathered at Woods Hole under Jule Charney and produced a report for the National Academy of Sciences called Carbon Dioxide and Climate: A Scientific Assessment. They had two primitive computer models that disagreed with each other, a stack of hand calculations, and no satellites worth mentioning. They concluded that doubling atmospheric carbon dioxide would warm the planet by somewhere between one and a half and four and a half degrees Celsius.

Nearly fifty years and roughly a billion times more compute later, that range has barely moved.

Here's the myth worth killing: the idea that climate science has been a story of models being wrong and getting patched. The opposite happened. The physics was right early, the uncertainty was honest, and what improved wasn't the headline number but the resolution. The 1979 team could tell you what the planet would do. They couldn't tell you what your county would do, which harvest would fail, or which pipeline was leaking on a Tuesday. That's the gap machine learning has been closing, and it's a gap of detail rather than direction.

• 1979: Charney and colleagues fix the warming range using two models and a lot of nerve.
• 2014 to 2020: neural networks start beating physics-only models at nowcasting, and cloud and weather emulators cut simulation costs by orders of magnitude.
• 2024 to 2026: detection goes operational. Satellites and public portals move the argument from "is this happening" to "who left that valve open."

AI's Own Energy Bill and the Data Center Debate

Time to name the darkness properly, because the column's optimism isn't worth anything if it skips this part.

AI runs on electricity and the electricity is not free. The International Energy Agency puts global data center consumption at 415 terawatt hours in 2024, about 1.5% of world electricity, and projects it more than doubling to roughly 945 terawatt hours by 2030. A single AI-focused data center draws as much power as 100,000 households, and the largest ones under construction will draw twenty times that. Emissions from data center electricity go from about 180 million tonnes today toward 300 million tonnes by 2035 in the base case.

The local version is worse than the global one. Nearly half of US data center capacity sits in five regional clusters, so the strain lands on particular grids and particular utility bills rather than being spread thin across a planet. The IEA reckons about 20% of planned projects risk delay from grid constraints alone. Anyone who tells you the power question is a rounding error is looking at the global average and ignoring where people live.

And the tools break. The most sensitive methane instrument ever flown, MethaneSAT, went silent on June 20, 2025 after fifteen months in orbit. The Environmental Defense Fund tried nearly 400 contacts over three weeks before calling it. Years of work and the best eyes we'd ever had, gone to an unexplained power anomaly. The infrastructure of seeing is not yet durable, and pretending otherwise is how you get surprised. It's a reminder that any reporting pipeline you depend on needs a plan for the day its data source stops answering.

The atmosphere doesn't care that the instrument watching it was well funded. It only stops warming when somebody closes the valve.

The Forecast for AI and Climate Through 2027

Here's the honest read on the next 18 months, with the good and the bad in the same paragraph where they belong.

The IEA's assessment is that the panic and the hype are both overdone. Data center emissions stay under 1.5% of total energy sector emissions through 2035, while widespread adoption of AI applications that already exist could cut something like 5% of energy-related emissions over the same window. That's a favorable ratio and an inadequate number. Five percent does not solve this. It's roughly the size of the gap between a climate policy that works and one that misses, which means the technology is worth something precisely because the policy is close.

Three things to watch between now and the end of 2027. First, whether detection data starts carrying legal weight: Oregon has already passed a law requiring a landfill with persistent detected emissions to adopt advanced monitoring, and California is buying satellite data directly. Second, whether the constellations survive contact with reality, since Planet has committed to three more Tanager satellites and FireSat needs dozens more to hit its 20-minute revisit. Both of those trends push the same way for anyone running a business: expect to answer for numbers you used to estimate, which is a good argument for putting a monitoring agent on the data before someone else does. Third, rebound effects, which are the quiet killer: efficiency gains that get spent on more consumption rather than less, a pattern the IEA flags explicitly for autonomous vehicles pulling riders off public transit.

The forecast in one line: the seeing gets much better, the arguing gets much harder to fake, and none of it substitutes for the unglamorous work of closing valves and building transmission.

The Countermove: Build a Home Energy Report Agent

The street finds its own uses for things. The same pattern that catches a super-emitter from orbit works at household scale, and the household version needs no satellite at all.

Buried in the IEA's analysis is a number that doesn't get enough attention: scaling up AI-led optimization in buildings could save around 300 terawatt hours of electricity globally, roughly the combined annual generation of Australia and New Zealand. Buildings are where the boring, enormous, unclaimed savings live. The barrier isn't sensors. It's that nobody reads their own bills closely enough to find the pattern, because reading a year of utility statements by hand is a miserable afternoon that pays off in ways you can't see until you've finished.

Start with the load-bearing basics before automating anything:

• Find your baseline: the draw your home never goes below, even at 3am with everyone asleep. That number is your floor, and it's usually higher than people guess.
• Check your rate plan against your actual pattern: time-of-use plans reward and punish very different households, and most people are on whatever they were defaulted into years ago.
• Download your interval data: most utilities publish hourly or fifteen-minute usage through a customer portal or a Green Button export. That file is the whole game.

Then build the watcher. The agent's job is not to control anything in your house, which is where these projects usually go wrong and get abandoned. Its job is to read: pull your interval data and your bills on a schedule, compute your overnight baseline and your peak hours, compare this month to the same month last year adjusted for how cold it was, and send you one short report with only the things that changed. A spike in baseline draw means something is running that shouldn't be. A shift in peak timing means your rate plan and your life have drifted apart. You can wire the same logic into any of the scheduled agent workflows you already run, and it costs nothing to leave running once it works.

That's the whole power shift in miniature. The utility has always had this data about you. Now you have it about yourself, in a form that talks.

The Horizon for a Planet That Can See Itself

Here's the future worth building, and it's quieter than the usual climate-tech pitch.

Picture the sky above a gas field in 2032. There are dozens of instruments up there now, cheap ones, run by universities and nonprofits and a few governments that got embarrassed into it. They don't look at people. They look at flare stacks and landfill caps and the cracked seal on a compressor station, at tens of meters of resolution, which is precise enough to name a valve and far too coarse to name a person. Every detection lands in a public portal. There's no watchlist, no enforcement drone, no algorithmic sentencing. Just a number on a map, and a phone call to whoever owns the equipment, and a follow-up pass a week later to see if the plume is gone. The accountability runs upward, toward the people with the assets, which is the only direction it has ever needed to run.

Now picture the ordinary version: somebody who fixes leaks for a living outside Odessa, who opens a work order with coordinates already on it instead of walking a line with a sniffer hoping to get lucky. Her job got better. She closes more valves per week than she used to. Multiply her by every basin on Earth and you have a rate of repair that starts to look like a curve bending, not because anyone invented a miracle but because the finding finally got as fast as the breaking.

That's the future this column wants. Not a smarter planet. A planet where the evidence is cheap, public, and pointed at infrastructure instead of at people, and where the argument about what to do next has to start from something true.

Before Monday: Your Home Energy Audit Checklist

You've got a weekend and a utility account you've never really looked at. That's enough.

1. Log into your utility's customer portal and download twelve months of bills plus whatever interval usage data they offer. Green Button export if you have it, a CSV if you don't, PDFs if that's all there is.
2. Find your overnight baseline: the lowest sustained draw in a typical 24 hour period. Write the number down. It's your floor, and everything above it is a choice.
3. Pull up your utility's published rate schedule and check which plan you're on. Compare it against your own peak hours rather than against the marketing copy.
4. Look up your own address on Climate TRACE or the Carbon Mapper portal and see what's emitting near you. Most people have never looked, and the map is public.
5. Build the report agent from the prompt above and schedule it monthly. Let it run for two cycles before you change anything, so you have a baseline to compare against.
6. Pick one thing the first report flags and fix it this month. One. The agent runs whether you're paying attention or not, so the only part that needs your weekend is the deciding.

Frequently Asked Questions

• Does AI help or hurt the climate on balance?+
  On the IEA's numbers, the help is bigger than the harm, and neither number is decisive. Data centers were about 1.5% of world electricity in 2024 and their emissions run under 1.5% of energy sector emissions through 2035. Widespread use of existing AI applications could cut roughly 5% of energy-related emissions by 2035. So the ledger is positive and nowhere near sufficient on its own. AI is a tool in the fight, not the answer to it, and the honest version of this story says both parts out loud.
• What can one household do with an AI energy agent?+
  Plenty, and it starts with reading rather than controlling. Point an agent at your utility bills and your meter data, have it find your baseline overnight draw, rank your biggest loads, and compare your rate plan against your real usage pattern. Most people find one appliance or one billing setting worth real money in the first pass. You don't need a smart home to do this, and a scheduled reporting workflow over a year of PDF bills is enough to start.
• Are emissions-detection satellites a surveillance risk?+
  These instruments are pointed at industrial infrastructure, not at people. Carbon Mapper and Climate TRACE publish plumes attributed to flare stacks, pipelines, landfills and power plants at tens of meters of resolution, which is coarse enough to identify a piece of equipment and far too coarse to identify a person. The accountability runs upward, from a leaking facility to the regulator, which is the opposite direction from the surveillance futures this column refuses to write.
• Do I need new hardware to start?+
  No. The first pass is pure document work: bills, the rate schedule your utility publishes, and whatever interval data your provider already lets you download. Hardware helps later if you want circuit-level detail, but buying a monitor before you've read a year of your own bills is the expensive way around. If you're weighing a gadget against a workflow, the limits are worth knowing first. Start with the paperwork you already own.