Before containers, loading a ship cost about $5.83 a ton and took days; after the world agreed on one standard box, shipping costs fell 90% in two decades and countries that had sat outside global trade suddenly had a door in. The cranes, the goods, the demand all existed — what was missing was the architecture that let the pieces snap together.

That, in a sentence, is market creation — and it’s the subject of a piece by Arushi Sharma Frank, founder of Luminary Strategies, on one idea: you don’t hunt for value that already exists, you build the market and let companies come into it.

Share

Share

“Regulatory rules written well, implemented in technical design, are the foundation of new commercial products, new moats, and new ventures.” — Arushi Sharma Frank (A.S.F.)

In a world where energy demand collides with the power grid: “The battery can respond. The software can control. The load can withdraw. The customer is willing to accept risk. The grid needs the service. Yet the pieces remain scattered because nobody has created the architecture that lets them work together.” Swap the battery for a shipping crate, an airwave, or a retirement account, and the sentence still holds.

Five markets that didn’t exist until a rule invented them

The airwaves. In 1994 the FCC began auctioning invisible spectrum to the highest bidder — turning “sky” into ownable, financeable property. Over 100 auctions later: $233 billion raised and the entire U.S. mobile industry born.

Your retirement. A modest 1978 tax clause — Section 401(k) — meant only to clarify how deferred pay was taxed. Ted Benna saw its potential; today it holds $10.1 trillion and spawned a whole asset-management industry.

Independent power. In 1978, PURPA forced utilities to buy electricity from independent generators (cogenerators and small renewables up to 80 MW), at their own “avoided cost.” It created the non-utility power producer: a market that hadn’t existed.

Rooftop solar. Net-metering rules make your utility buy back the power your panels don’t use, usually at the retail rate — turning a roof into a power plant. The first million U.S. solar installations took 40 years; the country has since blown past 6 million.

The air we breathe. Facing acid rain, the 1990 Clean Air Act created a market for pollution itself — a cap plus tradable permit. Emissions fell 43%, faster than promised and at a quarter of the projected cost.

The Pattern:

In every case the raw capability was lying around in plain sight — ships, radio waves, savings, sunlight on a roof, smokestack scrubbers. What was missing was an agreed-upon architecture: a spec, an auction rule, a tax line, a tariff, a permit, a market rule that makes hardware financeable and performance bankable.

As A.S.F. puts it, “a market rule can give a startup a category, customers and a moat.”

The Moat:

A moat is a competitive edge durable enough to survive competition — switching costs, network effects, cost advantage, efficient scale, and intangible assets — patents, brands, and regulatory licenses. A market rule is a moat by definition: a rare kind that doesn’t just make the category hard to cross, it draws the category’s borders in the first place. It is possible that the deepest moat is the market rule itself — because whoever helped write it understands the terrain before anyone else arrives. A.S.F. calls this “baking the pie”. Most venture chases a growing pie and assumes someone else is doing the baking. Her bet is that the baking is the business.

Final Notes:

Rules make markets, but they also break them. The same lever that turned airwaves into $233 billion can entrench incumbents, pick losing technologies, or freeze a design too early. Spectrum auctions have been faulted for favoring the deepest pockets; net-metering rules have been slashed and restored from Nevada to California; well-meaning standards have locked in inferior formats for decades. A rule is only as good as its architecture — and architecture written by the wrong hands becomes a moat around the wrong castle. Market creation isn’t automatically virtuous. It’s powerful, which is a different and more demanding responsibility.

Which returns us to the most useful idea in A.S.F.’s essay — that the empty space stays empty until someone decides they’re the one to fill it. Her own signature example is a Texas grid rule she sketched in a notebook after a single lunchtime question to ERCOT, which months later existed as market architecture where there had been none. “Power,” she writes, “begins in the belief that the thing can be built and that you may be the person who has to begin building it.”

The container was just a box. The rule that everyone would build the same box is what rebuilt world trade.

SOURCES & NOTES

All quotations from Arushi Sharma Frank are drawn from “Becoming an Angel for Prosperous Infrastructure,” Luminary Strategies (Jul. 2026).

· Frank, “Becoming an Angel for Prosperous Infrastructure” — Luminary Strategies

Luminary Strategies

Becoming an Angel for Prosperous Infrastructure - Interview with Aurora Energy Research

Luminary Strategies LLC (Luminary) is an infrastructure venture platform that builds technical policy moats while backing companies developing the next generation of power infrastructure. We literally build the markets that our advised companies will thrive in, maximizing the prosperity nexus between incentivized private capital and sustained, burgeonin…

Read more

2 months ago · 7 likes · Arushi Sharma Frank

· https://rethinkelectricity.com/library/resources/pclr-builds-the-moat-for-ai-loads

Container loading cost ($5.83 → $0.16/ton; ~90% decline) — Levinson, The Box; ISO https://www.iso.org/news/ref2215.html

· FCC spectrum auctions: 100+ auctions, $233B+ raised since 1994 — FCC https://www.fcc.gov/node/122689

· 401(k) assets ~$10.1T (Q4 2025) — Investment Company Institute https://www.ici.org/statistical-report/ret_25_q4

· PURPA 1978: qualifying facilities, avoided-cost purchase mandate — FERC https://www.ferc.gov/qf

· Net metering (45 states + DC); U.S. passed 6 million solar installations, 97% residential — SEIA https://seia.org/research-resources/6-million-solar-installations/

· Acid Rain Program: 43% SO₂ cut, ~1/4 projected cost — U.S. EPA https://www.epa.gov/acidrain/acid-rain-program

An AI data center runs flat out around the clock — every second, all year. Most of the new power going onto the American grid does not. Of the roughly 86 gigawatts of new generating capacity the U.S. Energy Information Administration (EIA — the federal government’s official energy-statistics agency) expects online in 2026, close to nine in ten megawatts are solar, wind, or batteries: power that flows only when the sun is up, the wind is blowing, or a battery still holds a charge.

The difference between those two kinds of power matters more than the fuel or the price per megawatt. The first — power you can call on any hour, on command — is firm, or dispatchable [a grid operator can turn it up or down at will]. The second is variable [it flows only when the wind blows or the sun shines]. Which kind a region has is what decides whether it can keep an always-on machine fed.

Not all watts are equal

The capacity factor [the share of a full year a power plant produces at its maximum rating]. A 100-megawatt plant that ran flat out every hour of the year would have a capacity factor of 100%. Nothing does. But the spread between power sources is enormous.

According to the EIA, in 2024 a nuclear plant ran at about 92% of its maximum across the year. A natural-gas combined-cycle plant [a gas plant that reuses its own waste heat to squeeze out extra electricity] ran at about 60%. Wind came in near 34%, and solar near 24%.

A solar farm and a nuclear reactor both stamped “100 megawatts” are not the same asset. Over a year, the reactor delivers close to four times as much electricity — and it delivers on a schedule you control, not one the weather controls.

You can see the same truth in the national totals. Nuclear is just 7.7% of America’s installed capacity [the nameplate rating of everything plugged into the grid], yet it produces about 18% of the country’s electricity, because it almost never stops. Wind is the mirror image: about 12% of capacity, but only 11% of generation. Firm power punches above its weight. Variable power punches below.

The grid is adding part-time (variable) power

Break down that 2026 buildout and the imbalance sharpens. Of the 86 gigawatts, solar is 51%, batteries 28%, wind 14%. New natural gas is about 6 gigawatts — and once you subtract the old coal and gas plants retiring, renewables and storage account for roughly 99% of the net additions.Taken at face value, that looks like the grid has simply decided the future is variable. It hasn’t — and the reason it hasn’t is mechanical.

What it means: firmness isn’t gone — it’s back-ordered

2026 looks almost entirely variable not because nobody wants firm power, but because you can’t buy a gas turbine right now even if you want one.

The world’s largest turbine maker, GE Vernova, ended the second quarter of 2026 with a backlog of 116 gigawatts of gas turbines on order — roughly a decade of its entire factory output, already spoken for. About a fifth of the recent growth in that backlog is headed for data centers and wait times for a large turbine have stretched to five to seven years. Solar panels and batteries, by contrast, can be ordered and switched on in a year or two. So the grid fills up with variable power first simply because it’s the only thing that arrives fast — not because it’s the only thing being bought.

And side by side the visible grid, a firm (dispatchable) power buildout is happening that barely registers in the official capacity numbers, because much of it is being built behind the meter [on-site power a data center builds for itself, bypassing the public grid]. The energy research firm Wood Mackenzie finds that in the disclosed pipeline of U.S. data centers, on-site gas now accounts for roughly 48% of total site capacity — developers so pressed for firm power on a timeline that they’re building their own gas plants rather than wait years for a grid connection.

The biggest buyers are going straight for the firmest source there is. Across 2024 and 2025, Microsoft contracted to restart the shut-down Three Mile Island reactor in Pennsylvania; Amazon signed for 1.92 gigawatts of nuclear from the Susquehanna plant; Meta took a 20-year deal for the entire output of Constellation’s Clinton reactor in Illinois. These aren’t side bets. They’re some of the most capital-rich companies on earth paying a premium and waiting until 2027–2031 to lock down power that never blinks.

Won’t batteries fix this?

The strongest objection is storage. If you overbuild solar and pair it with batteries, doesn’t variable power start to act firm? Partly — and that’s exactly why batteries are 28% of this year’s additions.

But look at what today’s batteries do. Most grid batteries are built to discharge for 2 to 4 hours. That’s ideal for one job: soaking up cheap afternoon solar and releasing it into the evening peak when everyone gets home. It smooths out the day. What it does not do is carry a data center through a still, overcast three-day stretch in January. For that you’d need long-duration storage [systems that can discharge for eight hours up to several days]. That technology is real and growing — deployments rose 49% in 2025 — but it isn’t commercially mature. Investment in it fell about 30% last year, and venture funding dropped 72%, as capital chased AI and cheap lithium batteries instead. (Even that decline understates the private pullback, since it counts a large U.S. government loan; strip that out and the commercial retreat is sharper.)

So the fair conclusion: batteries are closing the hours gap fast, and that’s genuinely valuable. The days gap — the one an always-on load runs through — is still open. Storage stretches the day; it does not yet cover the week.

The Signal

· Firmness is repricing: Always-on (”firm”) power is scarce and back-ordered for years; the market likely hasn’t fully priced the premium. (Inference; confidence: Med-High)

· Firm ≠ variable: A firm MW and a variable MW shouldn’t trade at parity on a grid serving always-on load. Value is shifting from nameplate capacity to hours of dependable output.

· Who benefits: owners of existing dispatchable capacity — nuclear fleet operators, flexible gas generators — plus turbine makers with multi-year backlogs.

· What to underwrite: duration and dispatchability, not gigawatts — price the ability to deliver on command.

Sources

1. Capacity factors — nuclear ~92%, gas combined-cycle ~60%, wind ~34%, solar PV ~24% (2024). EIA, Electric Power Monthly, Table 6.07.B (Primary, Tier 1 gov). eia.gov

2. Generation vs. capacity — nuclear 7.7% of capacity → ~18% of generation; wind ~12% capacity → ~11% generation; 2025 generation mix. EIA, Electricity in the U.S. (Primary, Tier 1 gov), 2026 (2025 data). eia.gov

3. 2026 additions — ~86 GW; solar 51%, storage 28%, wind 14%; ~99% of net additions renewables + storage. EIA, Today in Energy / Short-Term Energy Outlook (Primary, Tier 1 gov), 2026. eia.gov

4. Gas-turbine backlog — 116 GW end-Q2 2026, ~20% data-center-driven, 5–7-year lead times. GE Vernova, Q2 2026 financial results (Primary, corporate); Utility Dive (News), Jul 2026. gevernova.com

5. Behind-the-meter gas — on-site gas ≈48% of total site capacity in the disclosed US data-center pipeline. Wood Mackenzie, US Data Center Pipeline (Tier 2 analyst), Q1 2026. woodmac.com

6. Hyperscaler nuclear PPAs — Microsoft–Constellation (Three Mile Island / Crane restart); Amazon–Talen 1.92 GW (Susquehanna); Meta–Constellation 1.1 GW (Clinton). World Nuclear News / NucNet / company announcements (Primary + News), 2024–2025. world-nuclear-news.org

7. Long-duration storage — deployments +49% to >15 GWh in 2025; investment −30% YoY (VC −72%), excluding a $1.76B US DOE commitment. Wood Mackenzie, Long Duration Energy Storage Trends (Tier 2 analyst) via Utility Dive, Mar 2026. utilitydive.com

All sources fall within the 24-month recency window. Facts are labeled where the text moves from established data (EIA, corporate filings, Wood Mackenzie) to inference (the firmness-repricing read in “What it means” and “The Signal”).

20 deals logged. $105.56 billion in disclosed one-time capital commitments (power purchase agreements and the Google–SpaceX monthly lease isn’t counted as one-time capital).

Where it concentrated: Texas and the US grid markets drew the most activity; India saw the single largest multi-year commitment ($30bn, AirTrunk); France had two campuses on former power-plant sites. Two deals bypassed the public grid entirely — built behind-the-meter, next to their own power.

The lead: a platform that finances AI chips like real estate

In June, Broadcom, Apollo, and Blackstone announced a $35 billion platform to finance more than 20 gigawatts of custom AI computing through 2028; the first money funds an expansion for Anthropic (Apollo). It buys Broadcom XPUs — custom AI chips built to one customer’s exact design (unlike a GPU, which anyone can buy off the shelf) and makes that computing power available to frontier labs, named as Anthropic and OpenAI. The notable part is who’s funding it: Apollo and Blackstone through their credit and insurance arms, not a traditional venture fund.

How is works? Long-dated capital, the money behind insurance policies and credit funds pays for the chips up front. The labs commit to pay for the computing power over several years, and those commitments are what the financing rests on. The chip stops being something a lab buys and writes off and becomes an asset a pool of capital owns and gets paid back on over time, closer to how a building or a fleet of aircraft gets funded.

Why it’s worth noticing: for two years the constraint on AI has been physical; power, grid, buildings. This points to a financial constraint: how you fund chips that lose value every year. And the capital being tapped isn’t venture money; it’s insurance and credit.

The open question is whether AI chips make good collateral. A building holds value for decades; a top-end AI chip can be outclassed in a few years. So the structure leans less on the resale value of the hardware and more on the labs’ commitment to keep paying.

Intelligence layer

Verse raised $54M in June 2026, strengthening investment in growth-stage software startups building in this layer — using AI to find grid capacity, forecast demand, or manage on-site power. These are venture-scale rounds ($5–70M). Verse’s $54M in June follows GridCARE’s $64M in May, Emerald AI’s $25M in March, and Amperon’s Series B in January.

Verse’s Dispatch Intelligence manages a data center’s on-site solar and batteries and schedules when the site draws from the grid, allowing it to operate sooner rather than waiting in the interconnection queue (the multi-year line to connect a large load to the grid).

Software-scale startups at the AI/energy intersection are small relative to the physical infrastructure deals — generation and data-center capacity, funded in the hundreds of millions to billions (in June, roughly $1.3bn to $35bn). Both address the same constraint — the multi-year wait to connect a data center to the grid.

The more useful detail is the investor set. Verse’s round was led by Bessemer, with GV and NVIDIA participating. NVIDIA is also an integration partner: it is incorporating Verse’s software into its DSX AI Factory reference design for data-center builders.

The implication is straightforward. NVIDIA is treating the grid connection, not chip supply, as a key constraint, and is addressing it in the software layer as well as through hardware. Tracking which strategic investors recur across these rounds is a way to see where the constraint — and the value — is moving.

Quick explainers

• XPU = a custom AI chip built for one customer’s specific work (a GPU, by contrast, is a general chip anyone can buy).

• Gigawatt (GW) = a unit of power; one gigawatt is roughly enough to power a mid-sized city. “20GW of compute” means data centers drawing about that much electricity.

• Behind-the-meter (⚡) = a data center wired straight to its own power plant, skipping the public grid.

• PPA (power purchase agreement) = a long-term deal to buy electricity from one supplier.

• CMBS = a bond backed by commercial property — here, income from a data center.

• Grid markets (ERCOT = Texas; PJM = Mid-Atlantic; MISO = Midwest; SPP = central plains; TVA = Tennessee Valley; RTE = France) = the regional operators that run the power grid.

Sources

All 20 June deals, in log order:

1. Broadcom + Apollo + Blackstone — Apollo

2. KKR / Helix Digital Infrastructure — Data Center Dynamics

3. Ardian + Verne — Ardian

4. DayOne Data Centers — PR Newswire

5. Related Digital + Blackstone (Stargate “The Barn”) — Oracle

6. Meta + Reliance (India DC) — Meta

7. Meta + CleanMax / Fourth Partner (renewables) — Meta

8. Crusoe + Bergen Engines — Crusoe

9. Hitachi Energy / Canduct — Hitachi Energy

10. Google + SpaceX/xAI (compute lease) — Data Center Dynamics

11. Csquare IPO — The Deal

12. AirTrunk (India) — Forbes

13. Google + Intersect (Meitner, BTM) — POWER Magazine

14. Microsoft + Chevron (Project Kilby, BTM) — TechCrunch

15. Google (Chesterfield, VA) — VPM

16. Digital Realty + Blackstone (NoVa) — Yahoo Finance

17. Google (Bridgeport, AL) — AL Daily News

18. Blackstone / QTS (CMBS) — CoStar

19. EDF + SoftBank + Eclairion (France) — Data Center Dynamics

20. Meta + Crusoe (compute capacity) — Data Center Dynamics

The claim

For forty years, a data center was an IT building that happened to use electricity. That era is ending. A modern AI data center is a power-and-cooling building that happens to contain computers. The software didn’t force that flip. The heat did. Here’s the proof.

The proof

1. The racks got ten times denser. A rack is the cabinet that servers sit in — a metal frame about the size of a large fridge. For decades, a full rack drew 10 to 20 kilowatts. A kilowatt is simply how fast something uses electricity; a typical American home averages a little over one kilowatt across a day. So an old server rack used about as much power as ten to twenty homes.

Nvidia’s current AI rack — the GB200 NVL72 — draws about 120 kilowatts, and low-130s under full load, according to HPE. Same floor footprint as the old cabinet, roughly eight to ten times the power. That is a hundred homes’ worth of electricity flowing into one fridge-sized cabinet. (Nvidia; HPE, 2024–26.)

2. All of that power becomes heat, and air can’t carry it anymore. There is a rule of physics no computer escapes: nearly every watt of electricity that goes in comes back out as heat. For forty years, fans and moving air handled the job, because 10–20 kilowatts per rack is within what air can carry. At 120–140 kilowatts in the same space, it isn’t. Air physically cannot move heat off the chips fast enough. So cold liquid is now piped directly against the chip. Nvidia and the Open Compute Project treat liquid cooling at AI density as a requirement, not an upgrade (Nov 2024).

And once liquid enters, the building changes. A single AI rack weighs about 1.36 tonnes — roughly 3,000 pounds — so floors, doors, and plumbing are re-engineered around it. Berkeley Lab describes a data center as four physical layers: the servers that compute, the storage that remembers, the network that connects, and the infrastructure that powers and cools the other three. In an AI facility, that fourth layer is now the main event. It decides whether the building can host AI at all.

3. The electricity data shows the flip already happened. US data centers used about 176 terawatt-hours in 2023 — 4.4% of all US electricity — up from 58 TWh in 2014 (DOE / Berkeley Lab, Dec 2024). Look at the timing: consumption was flat at roughly 60 TWh through 2016, then turned upward from 2017 — exactly when GPU-accelerated servers entered the fleet. What is known: use tripled in nine years, and the turn matches a change in hardware. What it suggests: this growth is structural. It tracks server architecture, not the economy.

What it means

The limit on AI moved. The chip is no longer the scarce thing — you can order AI hardware and install it within months. The scarce thing is a location that can deliver enormous amounts of power and absorb enormous amounts of heat, at the same spot, at the same time.

In the US, a new power project now waits a median of about five years from grid-connection request to operation (Berkeley Lab, Queued Up: 2025). To be precise: that five years measures power plants joining the grid, not a data center’s own hookup — but if new chips need new power, they are on that clock. That is why “where” now rivals “how much.” The same rack can power up in months in one region and wait years in another. So when you hear “AI needs more compute,” translate it: AI needs more power, in a place that can also take the heat away.

The counterargument

“Efficiency will save us — it did before.” Partly right, and worth taking seriously. From 2010 to about 2016, computing grew enormously while data center electricity stayed nearly flat, because efficiency gains and the shift to hyperscale facilities absorbed the growth. Two honest replies.

First, that flat era ended for a physical reason, not a lack of effort. Efficiency per computation is still improving — but AI raised the power packed into each rack by roughly ten times, and better efficiency now shows up as more computing per building, not less power drawn by it.

Second, keep the scale honest: globally, data centers still use only about 1.5% of the world’s electricity (IEA, Apr 2025). The global total is not the problem. Concentration is. The US carries 45% of that load, clustered in a handful of grid regions, and the strain is local — this substation, this county, this watershed. US data centers directly consumed about 17 billion gallons of water in 2023 (Berkeley Lab). The fact: a small share of world electricity. What it suggests: national averages will keep saying everything is fine while individual grid regions run out of room.

The chip used to decide what a data center computes. Heat now decides what a data center is. So the more interesting map of the next decade isn’t where the best models get trained — it’s which places can hand a building a hundred-plus megawatts and take the heat back. Which places would you put on that map?

Sources

IEA, Energy & AI (Apr 2025) · DOE / Lawrence Berkeley National Lab, 2024 United States Data Center Energy Usage Report (Dec 2024) · LBNL, Queued Up: 2025 Edition (emp.lbl.gov) · Nvidia GB200 NVL72 documentation (developer.nvidia.com) · HPE (2024–26) · Berkeley Lab via Pew Research (water, 2024–25).

The Numbers

Start with one chip. The chips that do AI math are called GPUs graphics processing units. According to the Congressional Research Service, an advanced data-center GPU is rated to draw between 350 and 700 watts. For comparison, that is about what a microwave oven pulls while it is running. One chip, one microwave.

Now stack them. Chips sit inside servers, and servers get bolted into tall metal cabinets called racks. Lawrence Berkeley National Laboratory, the US government lab that tracks data center energy reported in its 2025 update that a rack of ordinary servers draws about 3 to 5 kilowatts. A rack built for AI can draw up to 100 kilowatts. Same size cabinet, twenty to thirty times the power. NVIDIA’s current flagship AI rack packs 72 GPUs into one cabinet and draws roughly 120 kilowatts, according to NVIDIA’s own technical documentation. That is more heat than fans can push out of a metal box, which is why these racks are cooled with liquid pumped past the chips, the way a car engine is.

Now zoom out to the building. The International Energy Agency (IEA) breaks down where a modern data center’s electricity actually goes: about 60% is used by the servers themselves. Cooling takes anywhere from about 7% in the most efficient large facilities to more than 30% in older, smaller ones. This is the part we miss. Every watt that goes into a chip comes back out as heat and removing that heat costs electricity too. You pay once for the computing, then you pay again to un-heat the room.

Stack the buildings, and the numbers start showing up in national statistics. Berkeley Lab found that AI servers in the US used about 2 terawatt-hours of electricity in 2017 and about 40 terawatt-hours by 2023 — twenty times more in six years. (One terawatt-hour is roughly enough electricity to power a mid-sized city for a year.) The IEA projects that electricity used by AI-style servers will keep growing about 30% per year through 2030, compared with about 9% for ordinary servers. And the facilities themselves keep scaling: the IEA estimates a typical AI-focused data center already uses about as much electricity as 100,000 households, while the largest ones under construction will use twenty times that.

What This Means

Here is where we move from fact to inference, what the numbers suggest, not what they prove. If the power draw is set by the hardware itself, then AI’s electricity demand is structural. You cannot fix it with tidier software alone, because the multiplication happens in physical objects: chip times rack, rack times building, building times campus. Each layer inherits the layer below it.

There is a second, quieter implication. The IEA notes that an AI data center is about ten times more capital-intensive than an aluminum smelter — the classic example of a power-hungry factory. When the machines are that expensive, letting them sit idle burns money. So, operators run them as close to around-the-clock as they can. That may make AI load not just large, but unusually steady and inflexible: it does not switch off at night, and it is costly to power down when the grid is tight.

Big, concentrated, and always-on, that specific combination, more than the raw size, is what utilities and grid operators are now working through.

The Honest Pushback

The strongest objection is efficiency. Every chip generation does more math per watt than the one before. Software is improving too, newer models can often do the same work with less computing. The IEA has run this scenario: if efficiency improves faster than expected, global data-center electricity demand in 2035 comes in about 20% below its base projection. And that cooling range: 7% to 30% shows how much room the buildings themselves still have to improve. This is a real, fair objection, and it is why every serious forecast is a range, not a single number.

But notice what has happened at the chip level. Even as each generation got more efficient per calculation, the power rating per chip went up, not down. Designers spent the efficiency gains on more speed instead of less power. And in past computing booms, making computing cheaper mostly made people buy much more of it.

So, the fair conclusion is that efficiency will likely decide how much AI we get per watt, not how many watts get used.

If the draw is set at the chip, the thing worth watching is not the next benchmark score. It’s the power rating on the next generation of chips. If that number climbs again, the electricity question climbs with it — rack by rack, building by building. So far, the draw isn’t an accident of young technology. It’s the design.

Sources

1. GPU power draw: 350–700 W per chip — Congressional Research Service, Data Centers and Their Energy Consumption (Primary, Tier 1 gov), updated May 2026 — congress.gov

2. Rack density: ~3–5 kW conventional vs. up to ~100 kW AI racks — LBNL / DOE, US Data Center Energy Usage Report: 2025 Update (Primary, Tier 1 gov), Jun 2026 — eta.lbl.gov

3. NVIDIA GB200 NVL72: 72 GPUs, ~120 kW per rack, liquid-cooled — NVIDIA technical documentation (Primary, corporate), 2025 — docs.nvidia.com

4. Servers ≈60% of data-center electricity; cooling ~7% (hyperscale) to >30% (enterprise) — IEA, Energy and AI (Primary, Tier 1), Apr 2025 — iea.org

5. US AI (accelerated) server electricity: 2 TWh (2017) → 40 TWh (2023) — LBNL / DOE, 2024 US Data Center Energy Usage Report (Primary, Tier 1 gov), Dec 2024 — report PDF

6. Accelerated servers growing ~30%/yr vs. ~9%/yr conventional, to 2030 — IEA, Energy and AI (Primary, Tier 1), Apr 2025 — iea.org

7. Typical AI data center ≈100,000 households; largest under construction ≈20× that — IEA, Energy and AI (Primary, Tier 1), Apr 2025 — iea.org

8. AI data center ~10× more capital-intensive than an aluminum smelter — IEA, Energy and AI (Primary, Tier 1), Apr 2025 — iea.org

9. High Efficiency Case: 2035 data-center demand ~20% below Base Case — IEA, Energy and AI (Primary, Tier 1), Apr 2025 — iea.org

AI IS MORE THAN SOFTWARE

Here’s the fact: by 2030, the IEA projects US data centers will consume more electricity than the country’s aluminum, steel, cement, and chemical production — combined.

In plain terms: the limit on how much electricity is available may show up before any limit on how smart AI can get. By 2030, the International Energy Agency (IEA) projects that data centers in the United States will use more electricity than the country’s aluminum, steel, cement, and chemical industries combined. That is not a software number. It is a heavy-industry number, attached to a technology that almost nobody thought of as a big power user a few years ago.

Behind every AI response is a data center.
And behind every data center is electricity.

The Numbers

Start with how big this already is. In 2023, data centers used about 4.4% of all the electricity in the US. That is according to a December 2024 report from the US Department of Energy and one of its national labs. Their own estimate says that share could grow to somewhere between 6.7% and 12% by 2028. That is a wide range, and it is wide on purpose: even the people building these models are not sure exactly how fast construction and power hookups will keep pace with demand. In plain numbers, that means data center electricity use could roughly double or triple in five years, from about 176 terawatt-hours a year to somewhere between 325 and 580 terawatt-hours. (One terawatt-hour is roughly enough electricity to power a mid-sized city for a year.)

This is not only a US pattern. The IEA estimates that by 2030, the world’s data centers combined will use roughly as much electricity as Japan’s entire economy uses today, all of it, for data centers alone. In the US, data centers are expected to account for close to half of all new electricity demand between now and 2030. And in 2025 alone, electricity demand from data centers grew about 17% worldwide, while total electricity demand grew only about 3%. Data centers are growing several times faster than the power grid around them.

The spending backs this up. Five of the largest tech companies spent more than $400 billion in 2025 building this infrastructure, and that number is expected to climb roughly 75% higher in 2026, according to IEA reporting. The money is moving fast. The electricity needed to power what that money buys is not moving nearly as fast. It is worth noting that these numbers come from two separate sources, the US government’s own energy lab and the International Energy Agency worldwide, and they point in the same direction independently, which is one reason to take the trend seriously even with the uncertainty built into it.

What This Means

Here is where we move from fact to inference, what the numbers suggest, not what they prove outright. A company can announce and fund a new data center in a matter of weeks. A new power line or power plant typically takes years to get approved and built, no matter how much money is behind it. If that gap holds, the real limit on how fast AI gets built may end up being how fast a region can add electricity, not how good the chips or the models are. That would hand real power to whoever controls land, the process for connecting to the power grid, and approval speed: utility companies, grid operators, and places that already have spare electricity to sell. It also means AI may end up getting built wherever the power already exists, not necessarily wherever the talent or the money wants it to be.

The Honest Pushback

Here is the strongest case against all of this: efficiency. Every year, computer chips do more work per unit of electricity. Cooling systems improve. Companies get better at shifting computing jobs to wherever power is cheapest or most available. If efficiency keeps improving faster than demand grows, this power ceiling could turn out to be smaller than it looks today, the same way some past predictions about computers using runaway amounts of electricity did not fully come true. This is a real, fair objection. It is also exactly why the Department of Energy’s own estimate is a wide range, 6.7% to 12%, instead of one confident number. That range is the industry itself admitting it is not sure how this plays out.

That said, history is not fully on the side of “efficiency will fix it.” In past computing booms, efficiency gains rarely shrank total energy use. At best they kept pace with growth, and in some cases, they made computing so cheap that people simply used far more of it. So, the fair conclusion is that efficiency is more likely to fuel demand than to cap it.

If the real limit is shifting from chips to electricity, the thing worth watching changes too. Not the next model release. The next power line that gets approved or delayed. The next utility that says how much new electricity it can deliver, and by when.