AI & Climate: The Connection Brief 2026-06-09 · 21 min · Human Pace, The Audio Edition ---------------------------------------------------------------- Hi, I'm T. I translate risk for people making capital decisions in clean energy and artificial intelligence infrastructure. This is Human Pace, the audio edition. Today's piece is called Artificial Intelligence and Climate — The Connection Brief. Artificial intelligence may be one of the most useful tools to help solve climate change challenge. It is also becoming one of its fastest-growing energy demand. How we generate energy has climate implications. Holding the two opposing forces together is the discipline that I hope to explore through this brief with you. Artificial intelligence now forecasts weather faster and farther than many physics-based models. It can spot a wildfire the size of a classroom from orbit, help balance a grid carrying more wind and solar, and search millions of candidate compounds for a better battery. The same machine runs on electricity and water. The technology companies building it are reporting rising emissions as data centers expand. So imagine a ledger with two columns: climate value on one side, climate cost on the other. In this brief, there are eight connections between AI and energy. The first connection is AI giving us sharper eyes on the planet. AI improves weather forecast and helps keeping an eye on what is already happening. For example, Google’s GraphCast makes a ten-day weather global forecast in under a minute. Google's WeatherNext 2 runs eight times faster. NVIDIA’s CorrDiff sharpens forecasts to kilometer scale at five hundred times the speed. None of that is a lab demo. The European Centre for Medium-Range Weather Forecasts has taken its own AI forecasting system into operations, where it issues tropical-cyclone track forecasts that beat the physics-based model. The same intelligence that reads the atmosphere reads the surface, making climate harm visible as it happens. Let's start with methane. Methane traps heat more than eighty times as effectively as carbon dioxide over a twenty-year span according to the UN Environment Program. And according to Google, Google can apply AI to satellite imagery, mapping the world’s oil-and-gas infrastructure and tracing methane leaks back to their source. Find the few, and you have found most of the problem. The same eyes watch for fire. Google’s FireSat compares any five-by-five-meter patch of the planet against earlier imagery — close enough to flag a wildfire the size of a classroom, with global coverage refreshed every twenty minutes. What does that mean? That means this is where AI's climate value is least disputed. Weather forecast is faster, cheaper, and further-reaching. It's a great planning tool for everyone exposed to weather such as utilities scheduling power, ports timing operations, and insurers pricing a storm. Additionally, AI turns a vast and blurry planet into a map with high fidelity of methane emission and active wild fire. With all that knowledge, what is the risk and who carries it? Emergency managers, grid operators, and risk managers are the people who must decide whether to trust the picture when capital and lives ride on the call. Sharper eyes move the decision earlier and make the problem visible sooner; they do not make the decision. The judgment, and the responsibility for acting on it, stays human. That is for the first connection. The second connection is a tension between load and lifeline. A grid carrying more wind and solar and more data centers is harder to balance. And AI sits on both sides of that problem. Let's start with the helpful side. The International Energy Agency, or IEA, finds artificial-intelligence-based fault detection can cut outage durations by 30 to 50 percent. It also finds that remote sensors managed by artificial intelligence could unlock as much as 175 gigawatts of transmission without building a single new line. That is capacity you get from software instead of steel. Then, in January 2026, the other edge showed up in the real world. Winter Storm Fern strained the grid, and Northern Virginia wholesale power jumped from $200 to $1,800 per megawatt-hour in a single day. The U.S. Department of Energy invoked emergency powers, authorizing grid operators in Texas, the Mid-Atlantic, and the Carolinas to call on data centers to deploy their backup generation. So for one weekend, the data centers that strain the grid became a shock absorber for it. The IEA now urges regulators to reward data centers for using their backup power and storage flexibly; doing so, turning a grid liability into grid support. Meanwhile, the hyperscalers have become the grid's biggest clean energy buyers: 49% of all global corporate clean energy contracting in 2025, new nuclear included. So how to think about this? Here the two columns of the ledger touch most visibly. On one side, AI adds load and on the other side, it supplies the tools to run the grid it strains. Tools like fault detection, forecasting, and curtailable demand. Whether the steadying outpaces the straining is a coordination question. A data center that can throttle itself is a shock absorber; one that can't is just load. The third connection is the materials research. Better batteries and clean energy materials have always been gated by one thing: search. The chemical design space runs to an estimated ten to the sixtieth power of possible compounds. With that many combinations, we cannot brute-force our way into discovery, we can only guess well. So AI was pointed at the guessing. DeepMind’s GNoME predicted 2.2 million new crystal structures, of which 380,000 are stable enough to pursue — including 528 candidate lithium-ion conductors, twenty-five times the number a prior study had found. At Argonne National Laboratory, AI foundation models are now being trained to navigate that same space for electrolytes and electrodes directly. But the catch sits on the other side of the screen. A predicted structure is a candidate, not a product. Synthesis, testing, and commercial scale-up still run on the timescale of chemistry and factories — years, not minutes. AI shortened the search. It did not shorten the build. This is the promise of AI. The slowest step in materials science which is the search has collapsed from decades of trial and error into a database query, handing the clean energy transition a catalogue of candidates it would have taken lifetimes to find. The conviction it asks for is to act on that head start: to back the synthesis, testing, and manufacturing that carry the most promising candidates into real batteries and panels. So where is the exposure? Arguablely, it’s the investors and manufacturers who back the build-out. AI has done the hard part, the finding, and what remains is the conviction to fund scale-up before the market crowns a winner. Our climate can gain from this effort. Every candidate that reaches production is a faster path to cheaper energy storage and more cleaner power. In this bet, whoever carries the risk also carries the reward. To me, it sounds promising. Now the other column of the ledger. The cost column. The fourth connection is the technology companies' rising emissions. That is the cost. The figures are audited and disclosed by the technology companies themselves. Google's total emissions reached 14.3 million metric tons of carbon dioxide equivalent in 2023, a 48 percent increase from its 2019 base year. The company attributes the increase primarily to data-center energy use and supply-chain emissions. Scope 3, meaning emissions across the broader value chain, accounts for 75 percent of the total. Microsoft tells the same story from a different baseline. The emissions up 29% since 2020, driven by the construction of more data centers and the embodied carbon that comes with them. And the effort to offset it is real. These same companies are among the world’s largest corporate buyers of clean energy, together contracting roughly half of all corporate clean-power deals in 2025. Google is candid about the bind: reaching its 2030 climate goals now faces significant uncertainty, because AI’s non-linear growth in energy demand, shifts in energy policy, and the slow scale-up of carbon-free power make its own trajectory harder to predict. When a company that measures everything says it cannot forecast its own emissions, that is worth hearing. These emissions are the cost of meeting demand — and that demand reaches across the whole economy, from every business and household now leaning on AI and the cloud. That appetite lands as electricity and carbon, and it shows up on the operators' books because that is where the infrastructure sits. So, in this case, who carries the risk? The risk is shared, because the demand is. The companies' 2030 climate targets were set before AI reshaped the demand curve. And since the appetite for these services belongs to the whole economy, so does the exposure: emissions counted on a few balance sheets are generated on behalf of everyone who uses them. The fifth connection is the load behind the model. As a starting point, the U.S. Data centers consumed about 4.4% of the country’s electricity in 2023 — 176 terawatt-hours — and they are projected to reach somewhere between 6.7% and 12% by 2028. That is 325 to 580 terawatt-hours according to Lawerence Berkely National Laboratory. Globally, the IEA estimates data-center electricity will roughly double by 2030, from about 415 terawatt-hours in 2024 to around 945 terawatt-hours. Just under 3% of world demand and roughly what all of Japan consumes. Here is how I read it. The large language model is the brain; the data center is the body that runs it. The intelligence feels weightless, but the body is all megawatts drawn from a grid already strained. The more capable the brain, the larger the body it needs and its appetite is outpacing how fast the grid can green up, so the climate absorbs whatever gets burned to meet it. Where the marginal megawatt is fossil, the cost of the model's convenience is paid in emissions no single user sees on the bill. The sixth connection is water: the thirsty giant. The index and analytics firm MSCI puts global data-center water use at roughly 560 billion liters today, on track for 1.2 trillion liters by 2030, the draw of more than four million U.S. households. But the volume is not the risk. There are many ways we use water way more. The location is. The World Resources Institute finds two-thirds of data centers built or in development since 2022 sit in water-stressed areas — which means the draw is heaviest exactly where there is least to draw on. So, a data center can buy clean power from anywhere on the grid; but it drinks from the watershed it sits on top. That makes water a coordination problem among operators and communities over a shared and shrinking resource. So we have a genuine tool and a genuine cost, both growing. Which leaves the question everyone wants settled in one direction or the other. Does it net out? The most authoritative read on the net comes from the IEA. It estimates that the broad use of AI could cut emissions equal to around 5% of energy-related emissions in 2035 — far larger than data centers' own footprint, yet far smaller than what the climate needs. Data-center emissions themselves stay under 1.5% of the energy sector’s total through that period — even while ranking among its fastest-growing sources. And AI’s own footprint can shrink: one 2025 study finds smarter model selection alone could cut AI energy use by 27.8% in a single year. So the arithmetic looks favorable. Here is why it may not hold. The IEA warns that AI is no silver bullet, that rebound effects can undercut its benefits. Efficiency gains can spur more consumption, so better technology alone need not deliver net reductions. Cheaper, faster AI invites more of it. This is the heart of the net question and the most authoritative read on it is deliberately modest. The IEA puts AI's climate help at a few percent of emissions: real, but far short of what's needed, and easily eroded by rebound. No one has yet netted AI's benefit against its full climate cost to a single figure. Who carries the risk here? We want the net to come out positive — but we still need the data to prove it. The honest position is that the answer is not in yet, and the deployment choices being made this decade are what will write it. That is the definition of a conviction risk: acting before the outcome is settled. That question stays open, and it may stay open for years. But capital does not wait for it to close. The money is already committed which means the question I can actually answer is a different one. The eighth connection asks who carries the bet. For example, Gen Re documents generative-AI weather models letting reinsurers refresh forecasts intra-day, run larger scenario ensembles, and sharpen the tail quantification that sets capital. And the losses they price keep climbing. Global insured catastrophe losses ran about $137 billion in 2024, against a $181 billion protection gap — then about $107 billion in 2025, the sixth consecutive year above $100 billion. Meanwhile the thing AI is building has itself become a fast-growing risk to price. U.S. data-center construction spending grew from $1.8 billion in 2014 to $28.3 billion in 2024. That is not a trend line. That is a step change. And it is concentrating geographically. The IEA finds half of all U.S. data centers under development sit inside pre-existing large clusters which raises the risk of local bottlenecks in power, water, and land. That clustering is a property-insurance problem in the making. Concentrate sites in a few regions, and a single hurricane, wildfire, or windstorm can strike many at once turning what looks like a set of separate buildings into one correlated loss. So what would getting ahead of it look like? Four things. 1st, Price a Probable Maximum Loss both site by site and across the whole portfolio — not one or the other. 2nd, Buy a single portfolio program rather than a patchwork of standalone policies, because every standalone program adds cost. 3rd, Set deductibles to your real risk appetite, resisting the reflex to buy the lowest one, where the premium just trades dollars with the insurer. 4th and final, because these assets run for decades, build in risk control from the start and keep widening the roster of carriers, so the capacity is there as you grow. Artificial intelligence's climate cost eventually appears as a property-insurance bill: fast-growing and geographically concentrated. The timing sharpens it. As capital expenditure floods into artificial intelligence infrastructure, the same inflation that makes these assets cost more to build also makes them cost more to insure. Insurance becomes part of the cost of capital, structured at the portfolio level and planned well before the next renewal. So who is holding the risk? The asset owners, first. How much of this exposure they keep versus transfer is theirs to decide: a portfolio priced to its real risk and a program built to a clear appetite turn it into a managed cost; left to default, the exposure simply lands — on the owner's balance sheet, and on the financiers and ratepayers behind it. Risk is managed, not eliminated. To recap, here is what this brief reveals. Eight connections, two columns. On one side, AI as climate tool — faster forecasts, sharper eyes on fire and carbon, a grid it both strains and helps steady, new materials. On the other, AI's own climate cost — emissions climbing at the very firms that build the tools, electricity headed for 945 terawatt-hours by 2030, water by the billions of liters in the driest regions. The net does not yet resolve to a number. No published figure balances AI's climate benefit against its climate cost at scale. And the system tends to feed itself: a more volatile climate raises the demand for AI's forecasting and resilience, which raises the resource load, which strains the grids and watersheds that volatility already threatens. So, let’s not wait for the net to settle; it may not for years. Let’s translate the unsettled ledger into the next decision. Fund clean power and resilient siting that shrink AI's costs over time. Read the insurance market as the early signal of where the physical risk has already landed. Let’s act on that signal before a loss forces the issue. This brief is one of a 3 connection briefs — AI and Energy, Energy and Climate, and AI and Climate. The series is complete and the written version, with every source linked, is at cleanpowerwhisperer.ai If you are making a capital decision inside this system and you’d like to think it through with someone, you know where to find me. Stay curious. Be safe. Be well.