AI & Energy: The Connection Brief
2026-05-28  ·  24 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 Energy — The Connection Brief.

Artificial intelligence feels weightless. You type a question, and an answer appears. Behind that answer is a physical system: land, chips, cooling equipment, transmission lines, water, and a great deal of electricity.

Artificial intelligence and energy are now structurally coupled. Each one's growth shapes the other's risk profile, capital requirements, and infrastructure timeline. Artificial intelligence is the largest new source of electricity demand in a generation. Energy availability is becoming a binding constraint on its expansion.

There are nine connections in this brief.

Let's start with the first connection: scale of AI.

According to the International Energy Agency, or IEA, global data center electricity consumption reached approximately 415 terawatt-hours in 2024 — about 1.5% of total world electricity use — growing at 12% per year over the last five years. The IEA estimates worldwide data center demand will more than double by 2030, to around 945 terawatt-hours — slightly more than the total electricity consumption of Japan. Gartner’s estimate runs higher still: worldwide data center electricity demand rising from 448 terawatt-hours in 2025 to 980 by 2030, with a 16% increase in 2025 alone.

And the share attributable to AI keeps climbing. According to Gartner, AI-optimized servers are projected to account for 44% of data center power consumption by 2030, up from roughly 21% in 2025. The IEA frames it more starkly: electricity demand from AI-optimized data centers is projected to more than quadruple by 2030. Quadruple. In five years.

Here is the translation.

Two separate modeling efforts — IEA and Gartner — converge on the same conclusion: data center electricity demand will roughly double by 2030. The difference between 945 and 980 terawatt-hours is a rounding error. The structural message is identical. AI infrastructure is building a country-sized electricity demand inside an electric grid designed for a different load profile.

So who carries the risk? Utilities and grid operators who must plan capacity for a demand curve they cannot predict, and ratepayers who absorb the cost of infrastructure built for demand that may or may not materialize at the projected pace.

Scale leads directly to the second connection: pressure on the grid.

Rising data center electricity demand is raising the risk of blackouts across a wide swath of the U.S. during extreme weather events. The supply growth that data centers need has not kept pace with the demand already added.

NERC — the North American Electric Reliability Corporation — issued a Level 3 Essential Action Alert. That is its highest category. What triggered it: computational loads connecting to the bulk power system produce sudden large load reductions, and oscillations that occur in seconds, leaving little or no room for real-time response. The alert outlines seven corrective actions that registered entities must implement by August 2026.

And the shortfall is already showing up in the auctions. PJM Interconnection, the largest grid operator in the United States, procured 145,777 megawatts in its 2027/2028 capacity auction — 6,517 megawatts below its own reliability requirement. Planning itself has become difficult. Grid operators face deep uncertainty about which announced data center projects will actually materialize, which makes long-range capacity planning something closer to guesswork.

So what does that actually mean?

The grid is not failing from neglect. It is failing from speed. Demand from computational loads arrived faster than the infrastructure, the planning processes, and the regulatory frameworks that govern them. NERC's alert is not a forecast — it is a response to oscillations already observed.

There are more than 67 million people served by PJM and similar grid operators. When reliability margins shrink, the consequence falls on households and businesses during extreme weather — the same moments when electricity is most critical.

That is the demand side. Hold those numbers, because the next question is the harder one. If AI wants that much power, where is it supposed to come from?

The third connection is nuclear fission.

Every major hyperscaler — the companies operating the world's largest cloud platforms — has signed nuclear power commitments. Microsoft, Amazon, Google, and Meta now account for more than 9 gigawatts of committed capacity. The deals span reactor restarts, existing plant offtake, and first-of-a-kind small modular reactors. The first nuclear power dedicated to artificial intelligence is expected by 2027.

Here is the order of how the deals landed.

September 2024. Microsoft signs a 20-year power purchase agreement, or PPA, with Constellation to restart Three Mile Island Unit 1 — 835 megawatts, and an estimated 16 billion dollars of lifetime economic impact to Pennsylvania.

October 2024. Google signs the first corporate small-modular-reactor power purchase agreement, with Kairos Power — 500 megawatts by 2035, first deployment by 2030.

June 2025. Amazon expands nuclear offtake with Talen Energy to 1,920 megawatts from Susquehanna through 2042. Separately invests 700 million dollars in X-energy for up to 12 Xe-100 reactors.

2025. Meta announces landmark agreements with Vistra, TerraPower, and Oklo for up to 6.6 gigawatts by 2035.

2027 (projected). First nuclear-to-AI electrons: TMI Unit 1 restart.

Essentially, AI wrote the check that revived nuclear energy. AI companies need carbon-free, 24/7 baseload power at a scale that renewables alone cannot deliver on the required timeline. Nuclear fits that profile. The deals happened in months, not decades.

So where does that exposure land? With the hyperscalers that signed 20-year PPAs on reactors that haven't been built or restarted yet. If Three Mile Island, or TMI, restarts on schedule in 2027, it validates the model. If timelines slip, those companies carry the gap between committed demand and unavailable supply. Fossil generation may fill the gap.

The fourth connection moves from fission to fusion.

Fusion energy has moved from theory to construction. Two companies are building real hardware on real timelines, backed by hyperscaler capital. Helion Energy, backed by Sam Altman, hit 150 million degrees plasma temperature — a milestone that validates the physics.

Microsoft has already signed a PPA to purchase electricity from Helion's first fusion plant, scheduled for 2028. Commonwealth Fusion Systems is constructing a 400 megawatt ARC reactor in Virginia, with power purchase agreements from Google and Eni targeting first power in the early 2030s.

Here are the sequence of the milestones.

July 2025. Helion breaks ground on first fusion plant — Orion — in Chelan County, Washington.

September 2025. Chesterfield County, Virginia approves zoning for Commonwealth Fusion Systems ARC reactor, at 400 megawatts.

February 2026. Helion reaches 150 million degrees Celsius plasma — the first deuterium-tritium fuel operation in the private sector.

2026 (target). Commonwealth Fusion Systems SPARC achieves first plasma.

2028 (target). Helion delivers first fusion power to Microsoft.

In practice, what that means?

Fusion was "always 30 years away" for decades. AI's appetite for clean baseload power turned fusion from a science experiment into a capital allocation decision. Now, at least, two companies have construction timelines, signed PPAs, and hyperscaler backing.

I asked myself: Who carries the risk?

Microsoft's Helion PPA and Google's Commonwealth Fusion Systems commitment are bets that first-of-a-kind fusion plants will deliver commercial power on schedule. If the physics works but the engineering slips, these companies hold contracts for energy that doesn't exist yet — while their data centers still need power.

The fifth connection sounds like science fiction: data centers in space.

AI's energy appetite is outgrowing the planet's infrastructure. The response: move compute off-world. Six companies — Aetherflux, Axiom Space, Kepler Communications, Planet, Sophia Space, and Starcloud — are building AI infrastructure for orbit on NVIDIA platforms.

Google is exploring the same frontier with Project Suncatcher — a moonshot equipping solar-powered satellite constellations with tensor processing units, or TPU, and free-space optical links.

The scale ambition is staggering. Terafab — a joint venture between Tesla, xAI, and SpaceX — targets one terawatt per year of output — double the current annual electricity consumption of the United States.

And the physics is genuinely attractive. In orbit, data centers receive more solar energy than on Earth, with no land constraints and no grid to overload. The math that does not work on the ground starts to work above it.

This sounds like science fiction until you follow the logic: space offers unlimited solar energy, no land-use conflict, no water constraints, no not-in-my-backyard opposition, and no grid interconnection queue. The physics works. The economics are the question — and launch costs are falling faster than most energy forecasts assume.

Who pays if this one goes wrong? The investors and ventures betting on orbital compute at scale. The physics of space-based solar is favorable. The engineering of launching, cooling, maintaining, and connecting thousands of orbital compute nodes remains unproven at commercial scale. SpaceX itself acknowledges that orbital data centers are part of its growth plans. Investors anticipating a possible initial public offering, or IPO, are therefore also betting on that technology.

But if launch costs stop falling, or if latency and reliability cannot match earth-based alternatives, these bets strand capital in orbit. Literally.

So the supply is being arranged. Contracts signed, ground broken, reactors restarted. But every one of those projects has to land somewhere — and somewhere has neighbors. Every earthbound project still has to land somewhere.

The sixth connection is community opposition.

According to Data Center Watch, $18 billion in data center projects have been blocked and another $46 billion delayed over the last two years, driven by opposition from residents and activist groups. The water footprint is scaling with demand. The International Energy Agency estimates global data center water consumption at roughly 560 billion liters per year, potentially rising to 1.2 trillion liters by 2030 — equivalent to the annual consumption of more than four million U.S. households.

Communities that host data centers absorb the infrastructure strain. The ones that block them absorb the economic loss.

The earth-based compute supply side cannot scale without community consent, and community consent requires addressing real resource impacts — water, noise, land use, grid strain. The $64 billion in blocked and delayed projects is a market signal: the social license to build AI infrastructure is not guaranteed.

Let's follow the exposure on both sides. Developers carry delay risk and stranded site-selection costs when projects stall. Communities carry the resource burden when projects proceed without adequate engagement. The mismatch between AI's deployment speed and local planning processes creates friction that neither side can resolve alone.

So how is the industry answering? Two ways, and they point in opposite directions.

Enter our seventh connection: Efficiency counter narrative.

The demand narrative is real. The efficiency counter-narrative is also real. Research is demonstrating that proper model selection alone can meaningfully reduce AI's energy footprint — choosing the right-sized model for a given task rather than defaulting to the largest available. I included the links to the studies on the brief.

For example, in one study, inference optimizations can reduce energy use by up to 73% against unoptimized baselines. In another study, model compression and knowledge distillation deliver roughly 60% faster inference with about 40% fewer parameters, while retaining some 97% of baseline performance.

But usage is scaling faster than efficiency can offset. Google processed 9.7 trillion tokens per month two years ago. Last year, 480 trillion. Today, 3.2 quadrillion per month — a 330x increase in two years.

The cost pressure is real enough to change behavior. Uber burned through its entire 2026 artificial intelligence budget in four months, prompting its chief operating officer to question whether the spend was worth it.

But let's hold the conclusion. Efficiency buys time. It does not eliminate the demand curve.

AI's energy problem has two sides: how much power the infrastructure demands, and how efficiently the software uses it. A 73% reduction in inference energy changes the math on grid capacity, water consumption, and capital requirements. The question is whether efficiency gains will outpace demand growth — or merely slow the curve.

Who carries the risk here? Everyone who assumes cheaper means less. Efficiency and demand are both expanding — but demand is expanding faster. When the cost per token falls, consumption doesn't hold steady. It explodes. Google went from 9.7 trillion to 3.2 quadrillion tokens per month in two years. Cheaper tokens didn't reduce energy demand — they unlocked more of it. That is the risk: every efficiency gain lowers the barrier to use, which drives more usage, which drives more energy consumption. The curve bends down per unit and up in aggregate at the same time.

The eighth connection is the hyperscaler energy arms race.

Accordingly to Bloomberg New Energy Finance, four companies — Meta, Amazon, Google, and Microsoft — were responsible for 49% of all global corporate clean energy buying in 2025. In the U.S. the concentration is sharper still: those four signed 16,777 megawatts of corporate renewables contracts, roughly 80% of the national total according to S&P Global.

Meta and Amazon led globally, contracting a combined 20.4 gigawatts — including 4.7 gigawatts of nuclear power accordingly to Bloomberg. Yet overall corporate clean energy buying fell in 2025, after nearly a decade of growth. So the market is splitting. Hyperscalers are accelerating while the broader corporate universe pulls back.

So what does that actually mean?

The hyperscalers are not just buying clean energy. They are becoming the clean energy market. When four companies control half of global corporate procurement, their investment decisions shape grid economics, renewable project finance, and energy policy. The risk is concentration: what happens to the clean energy pipeline if even one of these buyers changes strategy?

If AI capital expenditure tightens, or if hyperscalers shift to virtual PPAs over physical offtake, the pipeline of new renewable capacity loses its largest source of demand — and the gap between clean energy goals and actual clean energy delivered widens.

Which brings me to the question I always end up asking. However this resolves — whether the demand is overstated or understated, whether the power arrives on time or years late — somebody is carrying that risk.

Here we go to the ninth connection, the insurance architecture challenge.

Capital expenditure is spiking. The data center construction market is growing. Project values are increasing, and concentration risk becomes more apparent as data centers cluster within twenty-mile radiuses. A single regional catastrophe can hit a high concentration of insured value at once.

The market is adapting. For example, Zurich launched Data Center Project Guard — builders risk covering climate control failure, off-site storage, post-construction business interruption, and weather parametric triggers.

Another example is AIG. AIG built a multi-line lifecycle program spanning environmental, professional, directors and officers liability, political risk, marine cargo, builder’s risk, operational all risks, cyber, and general liability.

Legal advisors such as Covington mapped the full insurance coverage stack a data center demands across its lifecycle.

Here is the big picture. Individual insurance policies do well covering individual types of risk. They do less well at surfacing how those risks interconnect. A cooling failure could trigger property damage, business interruption, contractual penalties, environmental liability, and cyber exposure — simultaneously. The visibility across those lines often only emerges at the moment of loss. The insurance architecture for data centers must be as integrated as the assets themselves.

And brokers and risk advisors who default to placing siloed coverage miss the opportunity to demonstrate a proactive, holistic approach to risk management — the kind of approach these assets demand.

So, here we have it. Nine connections. One pattern: AI and energy are in a feedback loop. AI drives unprecedented energy demand. Energy constraints shape where, when, and whether AI infrastructure gets built. Communities are blocking $64 billion in projects. Grids are issuing their highest alerts. Insurance markets are writing policies for facilities whose loss profiles are still emerging.

The capital commitments keep accelerating — more than 700 billion dollars of hyperscaler capital expenditure in 2026, nuclear deals measured in decades, fusion plants breaking ground, and artificial intelligence models trained in orbit.

Whether this produces a better energy system depends on whether the speed of AI buildout is matched by the infrastructure, societal diffusion, and financial markets that have to absorb it. That is a risk management question — and it is one side of a larger system.

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.

That's the piece. 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.