Vector 1
Forecasting & SimulationAI Strategy & EconomicsEnergy & Grid

There Is Plenty of Power for AI. The Shortage Story Is Misinformed or Self-Serving.

Brian J. Curry
Brian J. Curry
October 18, 2026 · 39 min read

Inflated forecasts, outdated per-query myths and a market that pays incumbents for scarcity have turned a pricing problem into a supply panic. A model built on flexible compute, smarter wires, nuclear, geothermal and storage shows AI can pay for a bigger grid that lowers household bills.

Summary

There is enough energy for AI. What has been missing are rules that make AI pay for it.

  • AI can easily afford its own power. Electricity is 3% to 4% of what its compute rents for, and only 7% to 11% even at the price of new nuclear.
  • Forecasts are inflated. About 25 of the 90 GW in utility forecasts may never show up, and queues may overstate real demand three to fivefold. When Ohio made developers pay for capacity they reserved, its queue shrank from 30 GW to 13 GW.
  • Per-query fears are out of date. A median chatbot prompt uses roughly nine seconds of TV’s worth of power. The real growth is in reasoning, video and agent workloads, which use far more.
  • The shortage story has beneficiaries. These include owners of existing plants collecting scarcity prices, utilities with bigger capital plans, turbine makers, speculative developers, and AI companies seeking subsidies.
  • Running AI at home or on your phone helps only when the model is small and the hardware sleeps when idle. A home AI box left on all day can use about 90 times the energy of sending the same questions to the cloud.
  • The real problem is price, not supply. Rules that let AI bid up the existing grid and spread the bill across everyone are what raise household costs.
  • The fix is to make AI the anchor buyer: it brings its own supply, flexes in peak hours, pays for its grid upgrades, funds early nuclear and geothermal, and pays a small dividend to households. In the model, that cuts household rates 31% by 2040. Letting costs spread across everyone raises them 24% by 2030.

In July, PJM, the grid operator that serves about 65 million people from Illinois to North Carolina, cleared its capacity auction for 2028/2029 at $325 per megawatt-day. That is more than 11 times the $28.92 it cleared for 2024/2025, the third auction in a row to hit the price cap, and the region still came up roughly 6,800 MW short of its reliability target. PJM’s independent market monitor puts most of the increase on forecast data center load. Nationally, the Energy Information Administration expects the average residential price to reach 18.2 cents per kWh this year, up from 16.5 cents in 2024.

Two stories have grown up around numbers like these. One says AI is about to swallow the grid, that every chatbot answer drinks water by the bottle, and that the country will run short of power. The other says AI is quietly raising your electric bill. The first story is mostly overstated. The second is real, but the cause is not the physics of electricity. It is a set of rules that lets a new, price-insensitive buyer bid up the grid households already paid for and then spreads the bill across everyone.

My argument is that there is plenty of power for AI, and that AI is the first electricity customer in a century that can pay several times the cost of power and barely notice. Treated as an ordinary customer, AI bids up existing capacity and households pay a scarcity rent. Treated as an anchor buyer, required to bring its own supply, flex in the few tight hours of the year, pay first-of-a-kind prices for nuclear and geothermal, and send a dividend back to the people who built the grid, AI finances a larger and cleaner system that households inherit at learning-curve prices.

This piece tests that argument with a three-part model you can run: a national speed-to-power stack that sets every available lever against AI load, a learning-curve model of what AI pays as the first buyer of new nuclear, a household bill model that compares four ways of connecting AI to a utility, and a module that tests moving inference to homes and devices. It also connects to my work on the Autonomy Economy, because if AI really does drive down the cost of goods and labor, electricity becomes the input that decides whether those gains reach people.

Who benefits from the shortage story

Not everyone warning about shortages is wrong. Grid operators are right that reserve margins in some regions are thin, and engineers who say a specific substation cannot take another 500 MW next year are usually correct. But the national story, that AI is running the country out of power, has an audience of people who gain from believing it, and it helps to name them.

Owners of existing power plants are the clearest winners. A capacity auction pays the clearing price to every megawatt that clears, old or new. When forecast data center load pushes PJM’s price from $28.92 to the $325 cap, a plant built decades ago earns the same scarcity price as one built last year. PJM’s market monitor attributed about $6.2 billion of the 2027/2028 auction’s cost to data centers that had not been built. In the model later in this piece, the socialized case sends households about $225 each in a single year to generators that already exist, for no new output.

Utilities have a structural interest too. Under cost-of-service regulation, an investor-owned utility earns a regulated return on the capital it builds, so a larger load forecast supports a larger capital plan. LBNL and Brattle found investor-owned utility spending on generation rose 22% in 2025. That does not make the spending wrong, but it means the party producing the forecast also profits from a bigger one.

Speculative developers benefit from inflated queues. Filing the same project with several utilities costs little and holds a place in line. When AEP Ohio required new data centers to pay for 85% of contracted capacity whether they used it or not, its queue fell from more than 30 GW to about 13 GW within months.

Equipment makers are selling into a seller’s market. GE Vernova is offering gas turbine slots as late as 2031 on a 116 GW backlog, Wood Mackenzie puts turbine costs up more than 195% since 2019, and manufacturers have started charging reservation fees to hold a place in the queue.

AI companies also use the story. A grid in crisis is a strong argument for federal incentives, fast-tracked permits and exemptions from local review. OpenAI has pushed for federal support for large-scale AI infrastructure, while Microsoft has said it supports help on permitting but not electricity subsidies.

And the story travels because it is vivid. A dramatic number about chatbots and water spreads faster than a correction measured in fractions of a milliliter.

None of this requires bad faith. It only requires that each party’s incentive points the same way: toward forecasts that are too high and prices that are paid by someone else. The rest of this piece shows what the numbers look like when you strip that out.

How much of the power panic is real

Start with scale. The International Energy Agency estimates that data centers used 485 TWh of electricity worldwide in 2025, up 17% in a year, and projects about 950 TWh by 2030, roughly 3% of global demand. That is fast growth from a small base. The IEA’s Energy and AI report put data centers at about one tenth of global electricity demand growth to 2030, less than industrial motors, air conditioning or electric vehicles. The picture changes by geography: in advanced economies data centers account for more than 20% of growth, and in the United States the IEA expects them to drive almost half of it.

Then look at the forecasts. Lawrence Berkeley National Laboratory measured US data center use at 176 TWh in 2023, 4.4% of the national total, and projected 325 to 580 TWh by 2028. A range that wide tells you how little anyone knows. Grid Strategies has estimated that roughly 25 GW of the 90 GW of data center load in utility forecasts may not materialize, because developers file for the same project in several places. Jonathan Koomey, who has studied data center energy use for more than two decades, told an Ohio manufacturers’ conference in August that interconnection queues may overstate the demand that actually shows up by three to five times. The evidence is already visible. In January, PJM cut its 2028 peak forecast by about 4.4 GW, citing better vetting of data center requests.

The per-query numbers are where the coverage drifted furthest. In August 2025 Google reported that its median Gemini text prompt used 0.24 watt-hours, about as much as running a television for nine seconds, and 0.26 milliliters of water, and that energy per prompt had fallen 33-fold in a year. Researchers criticized the paper for leaving out training, water used at power plants and location-based carbon accounting, and those criticisms are fair. Even so, the IEA estimates that converting every conventional web search to a simple AI query would use under 4 TWh a year, less than 1% of current data center consumption. The real energy story is somewhere else: the IEA puts video generation, reasoning and agentic tasks at hundreds to thousands of times the energy of a simple text query, and that is where demand is heading.

So the national panic gets the scale wrong. What it gets right is local and financial. PJM’s market monitor estimated that forecast data center load above existing levels accounted for $21.3 billion, or 45%, of the $47.2 billion cost of capacity cleared across three auctions. Gas turbines are back-ordered for years. And the households least able to absorb a price increase are already stretched: LBNL and Brattle found that for the roughly one third of households earning under $50,000, electricity takes more than 5% of income, and about one household in six is behind on energy bills.

Both halves point to the same fix. If the load is smaller and more flexible than forecast, and the damage comes from how costs are allocated, the answer is rules that make AI carry its own risk and fund new supply.

Who pays for the marginal megawatt

A residential rate covers three things: energy, capacity (the promise that enough generation exists to meet the annual peak) and wires (transmission and distribution, mostly fixed costs recovered per kWh). In cents per kWh:

r_res = e + (P_cap × 365 × M_res) / Q_res + W / (Q_base + w × Q_dc)

Here e is the energy component, P_cap is the capacity price in dollars per MW-day, M_res is the residential capacity obligation in MW (its share of the peak plus the reserve margin), Q_res is residential energy, W is the annual wires revenue requirement, Q_base is energy sold to everyone except the data center, Q_dc is the data center’s billed energy and w is the share of a pro-rata wires allocation the data center actually pays.

The equation shows three channels. Scarcity pricing: if AI adds peak demand faster than supply arrives, P_cap rises on every megawatt households already buy, not just the new ones. Socialized upgrades: if the substations built for a data center go into W, households pay for them. And the channel that runs the other way: most of W is fixed, so if the data center pays a fair share (w above zero), the same cost is spread over more kWh and everyone’s wires charge falls.

A worked example on a stylized utility: 10,000 MW of peak, 50 TWh of annual sales (17.5 TWh residential), $4.5 billion of annual wires cost, a 7-cent energy component, residential load at 45% of the peak, a 15% reserve requirement and a capacity price of $118.50 per MW-day.

e = 7.00   capacity = 5,175 MW × $118.50 × 365 / 17.5 TWh = 1.28   wires = $4.5B / 50 TWh = 9.00
r_res = 17.28 cents per kWh

That is close to the national average. Now add 3,000 MW of AI load at an 85% load factor, 22.3 TWh a year. If it is socialized, supply arrives late and short, the reserve margin falls to about 8%, capacity clears at the $325 cap, energy picks up a scarcity premium, $120 million a year of upgrades lands in the shared pot, and the data center pays only 30% of a pro-rata share of existing wires. The residential rate rises to about 20.2 cents. If instead the data center brings capacity for its own peak, pays its own upgrades and half of a pro-rata share of existing wires, capacity barely moves and the residential rate falls to about 15.7 cents.

Same data center, same households. A 4.5-cent gap comes from the rules.

AI can afford to pay full freight

The asymmetry that makes the anchor-buyer idea work is willingness to pay.

An Nvidia H100 SXM is rated at 700 watts, about 0.84 kW at the wall with a power usage effectiveness of 1.2. At an industrial price of $80 per MWh, its electricity costs $0.067 per GPU-hour. Published on-demand rental prices for the same GPU ran from about $1.99 to over $12 an hour in October 2026, so even at the cheapest listing, power is 3.4% of what the compute sells for.

Now price that power at what new nuclear costs. Lazard’s 2026 analysis puts nuclear at $175 to $255 per MWh. At those prices, power costs $0.147 to $0.214 per GPU-hour, 7.4% to 10.8% of the lowest rental price. AI can pay first-of-a-kind prices for the most expensive firm clean power on the menu and power would still be about a tenth of what its compute sells for. A household cannot. Electricity is a necessity with few short-run substitutes, and the lowest-income homes spend the largest share of income on it.

The design rule follows directly: put price risk and early-technology cost on the buyer with high willingness to pay and the ability to flex, and protect the buyer that has neither.

The supply toolbox

There is no single answer to AI power, and the debate goes wrong when it treats one technology as the answer. The better question is what each lever can deliver, how fast and at what cost, and which of them AI’s money should pay for first. Here is the toolbox, roughly in order of speed.

Demand hygiene is the fastest supply there is. Every gigawatt of speculative load removed from a forecast is a gigawatt that capacity markets stop pricing. Minimum-take tariffs like AEP Ohio’s do this at no cost to anyone except speculators.

Flexible compute comes next. Researchers at Duke University’s Nicholas Institute estimated in February 2025 that the US system could absorb about 76 GW of new load if that load accepted curtailment for 0.25% of its annual energy, about 98 GW at 0.5% and about 126 GW at 1%, with average events lasting 1.7 to 2.5 hours. In a 2025 demonstration in Phoenix under EPRI’s DCFlex program, Emerald AI cut the draw of a 256-GPU cluster in an Oracle data center by 25% for three hours during Salt River Project’s evening peak while meeting its compute targets. Google has signed agreements with Indiana Michigan Power and the Tennessee Valley Authority to curtail machine learning workloads during grid stress.

Smarter wires make that headroom usable where data centers actually are. A UC Berkeley and GridLab study found that reconductoring existing lines with advanced conductors can up to double their capacity within existing rights of way, usually for less than half the cost of new lines, and could enable 764 GW of transmission-connected solar by 2035 in indicative corridors.

Surplus interconnection reuses grid connections that already exist. Researchers at UC Berkeley found about 500 GW of economically viable solar could be sited within six miles of existing fossil plants and share their interconnection, over 800 GW with wind, at more than 75% of the fossil generators they studied. Paired with storage, that capacity can be firmed for the evening peak. BloombergNEF put stationary storage packs at $70 per kWh in 2025, down 45% in a year, and the IEA expects 20 to 25 GW of batteries inside data centers worldwide by 2030.

Nuclear restarts are the fastest firm clean capacity available. Constellation expects to restart the 835 MW Crane Clean Energy Center, the former Three Mile Island Unit 1, in 2027 under a 20-year Microsoft contract and with a $1 billion federal loan, although a PJM analysis suggested the grid connection might not happen until 2031. NextEra plans to restart the 615 MW Duane Arnold plant in Iowa by 2029 with Google as primary customer and a $1.9 billion federal loan closed in September. Palisades in Michigan began loading fuel in August but has slipped well past its original target, a reminder that restarts are fast for nuclear and slow for everything else.

Uprates squeeze more out of reactors that already run. The US fleet is about 98 GW, and estimates of uprate potential range from 3 to 8 GW. The NRC expects applications adding 2.4 GW through 2032. Constellation plans more than $4.3 billion of uprates across 11 reactors to add nearly 1 GW, and Vistra has a $4.2 billion federal loan for 433 MW at three plants, with Meta buying the output. Per kilowatt, uprates cost less than new reactors but more than new solar with storage or combined-cycle gas.

Enhanced geothermal has crossed from pilot to power plant. Fervo Energy declared commercial operation of the first 33 MW block of its Cape Station in Utah this month, with about 100 MW expected by January 2027 and another 400 MW under construction for 2028. Fervo expects future blocks to take about 18 months to build, which would make geothermal the fastest new-build firm clean option on this list.

New reactors are where AI’s willingness to pay matters most. Ontario Power Generation is building four 300 MW BWRX-300 small modular reactors at Darlington for CAD 20.9 billion, with the first unit at CAD 6.1 billion, the fourth at CAD 4.1 billion and first power in 2030. Meta’s January agreements could finance up to 6.6 GW by 2035, including uprates at Vistra plants, up to 1.2 GW from an Oklo campus in Ohio as early as 2030, and two TerraPower Natrium units with rights to six more. In large reactors, the federal government’s October 2025 agreement with Westinghouse covers at least $80 billion of AP1000s, each about 1,110 MW. Vogtle’s two AP1000s cost roughly $35 billion and ran about seven years late, but Unit 4 cost about 30% less than Unit 3. Executive orders signed in May 2025 set a goal of 400 GW of US nuclear by 2050, 10 large reactors under construction by 2030 and 5 GW of uprates.

Gas is the bridge, and it is getting more expensive to build. Lazard’s 2026 ranges put combined-cycle gas at $51 to $129 per MWh against $40 to $98 for utility solar and $37 to $99 for onshore wind, and IRENA found the installed cost of utility solar fell 87% from 2010 to 2024. With turbine slots stretching to 2031, gas cannot arrive faster than most of the levers above. The IEA expects 15 to 27 GW of onsite gas at data centers by 2030 and notes that reliable onsite systems need 30% to 70% more generation than peak demand, an overbuild that a grid-connected, flexible data center avoids.

Efficiency sits under all of it. Stanford’s AI Index found the price of GPT-3.5-level inference fell 280-fold between November 2022 and October 2024, and the IEA estimates energy per AI task has recently fallen by at least an order of magnitude a year. Rebound eats much of that gain, but it is the reason straight-line demand forecasts keep missing.

The data center is not the only place AI runs

Everything above assumes AI runs in data centers. A growing share will not. McKinsey expects inference to make up more than half of AI compute by 2030, and AMD’s chief technology officer, Mark Papermaster, has said that over the next three to six years most inference will move out of data centers and onto phones, laptops and other devices. Some people are going further and running open models at home on purpose-built machines such as Nvidia’s DGX Spark or a Mac Studio.

Per answer, a small local model can be very cheap. A 2026 benchmark by Zähl and colleagues on a consumer RTX 4060 Ti measured 0.27 joules per token for a 0.5 billion parameter model, with 7 billion parameter models using up to 8.6 times more. For a 300-token answer, that is about 0.023 watt-hours for the small model, a tenth of Google’s 0.24 Wh median Gemini prompt, and about 0.197 Wh for the 7B model, roughly 0.82 times the Gemini figure. A published benchmark of a DGX Spark running a quantized 4B model flat out implies about 0.005 Wh per answer. These comparisons flatter the local machines. The GPU readings leave out the rest of the computer, Google’s figure includes fleet overhead, and a half-billion parameter model is not doing the same work as a frontier model.

Idle power is where home AI loses. The same DGX Spark benchmark measured 45 watts at idle, which is 1.08 kWh a day if the machine stays on, the energy of about 4,500 median Gemini prompts. A Mac Studio idling at 9 watts uses about 900 prompts’ worth a day. Someone who asks 50 questions a day uses about 12 Wh in the cloud. Leave a Spark on all day to answer the same 50 and the total is about 90 times higher. Data centers win because they keep hardware busy. A Hugging Face study on H100s found that even small batches cut energy per generated token by about a third, and production servers batch far more heavily than that.

The grid question is different from the efficiency question. Moving inference to homes takes load off the part of the grid that is causing the price problem, the transmission lines and capacity auctions, and puts it on household circuits. But it lands on the least flexible part of the system, the residential evening peak, and outside the large-load tariffs that make AI pay its way. I added a small module to the model to test this. It takes 65 GW of real AI load, assumes 55% of it is inference, moves 20% of that inference to homes and devices, and assumes household AI use runs 1.5 times its average at the evening peak.

Moving that work avoids about 5.4 GW of firm capacity at data centers. Whether the grid comes out ahead depends on how much energy the edge uses per task relative to the cloud. The break-even ratio is about 0.59. At a ratio of 0.25, typical of small on-device models handling simple tasks, the grid gains about 3.1 GW of relief. At 0.5 it gains 0.8 GW. At 1.0, it loses 3.8 GW, because the household peak grows faster than data center demand shrinks. At 2.0, typical of a mid-size model on a home rig that idles all day, it loses 12.9 GW.

So edge inference is a real lever, with conditions. It helps when it means small, efficient models running on devices people already own, sleeping when idle, ideally charged from rooftop solar or a home battery. A home rig running a mid-size model every evening moves load onto the grid’s hardest-to-flex customers and out of reach of the rules this piece argues for. Idle-power standards for AI hardware and time-of-use rates for households would push the edge toward the good version.

Model part one: speed to power

The first model asks a national question: how many gigawatts of firm capacity does AI actually need by 2032, and how much of it can the levers above cover before anyone orders a new gas turbine?

It starts from the roughly 90 GW of data center load in utility forecasts, optionally removes the 25 GW Grid Strategies flags as duplicate or speculative, treats 25% of AI peak as flexible (the share Emerald AI demonstrated), and then subtracts each lever at an accreditation factor that reflects how much it counts toward firm need. Two assumptions drive the result, and they are mine: that 30% of the Duke headroom is deliverable where data centers are, and that reconductoring raises that by another 15 points. Solar plus four-hour storage on surplus interconnection is counted at 50% accreditation, and only 5% of the Berkeley potential is assumed built by 2032.

Taking the forecast at face value, the levers bring firm need from 67.5 GW down to about 15.9 GW by 2032, which is what gas has to bridge. With the de-duplicated forecast, the levers more than cover firm need by 2032 with about 2.8 GW to spare. By 2035, adding new nuclear (Meta’s 6.6 GW portfolio and about five AP1000s), more geothermal and more solar with storage covers both cases with no gas bridge left.

The sensitivity matters more than the central case. If only 10% of the headroom is deliverable, nothing flexes and the full 90 GW is real, the 2032 gas bridge is 65 GW, more than half of GE Vernova’s entire global backlog. With a de-duplicated forecast, 25% flexibility and 20% deliverable headroom plus 10 points of reconductoring, it falls to 8.6 GW by 2032 and zero by 2035. The difference between those two worlds is almost entirely forecast discipline, flexibility and wires, which are the cheapest levers on the list.

Model part two: AI as the first buyer of new nuclear

The second model asks what happens if AI buys the expensive early units of a new reactor design and households buy the later ones.

Darlington gives a real starting point. Converting at the exchange rate implied by its budget, unit 1 costs about $14,691 per kW and unit 4 about $9,874. Two doublings of cumulative units took a third off the cost, which implies a learning rate of about 18% per doubling. Wright’s law then gives the cost of unit n:

C(n) = C(1) × n^log₂(1 − LR)        with C(1) = $14,691/kW and LR = 18.0%

Converting capital cost to a levelized cost with a 7% cost of capital, a 30-year life, a 93% capacity factor and $30 per MWh for operations and fuel, unit 1 comes out at $175 per MWh, which matches the low end of Lazard’s nuclear range. Unit 10 falls to $7,594 per kW and $105 per MWh, and unit 40 to $5,104 per kW and $80 per MWh, competitive with new combined-cycle gas.

If AI buys the first 10 units and utilities serving households buy units 11 to 40, those 30 units cost 20% less than if households had bought units 1 to 30 themselves, about $13.3 billion saved on 9 GW. That is the anchor-buyer logic in one number. AI pays the first-of-a-kind premium it can easily afford, and households inherit the cheaper reactors.

The caveat is large and should be stated plainly. US nuclear construction has mostly shown costs rising from one plant to the next, not falling, and Darlington’s learning is a budget projection, not a result. Vogtle’s Unit 4 coming in 30% cheaper than Unit 3 is the best recent evidence that learning is possible when the same team builds the same design back to back. That is exactly what a large, patient buyer with a fleet order can provide.

Model part three: household bills

The third model returns to the stylized utility and runs four regimes from 2026 to 2040, with AI load growing from zero to 2,000 MW by 2031 and 3,000 MW by 2035. Everything is in real 2026 dollars.

Socialize: AI connects as an ordinary large customer. Matching supply arrives three years late and covers 90% of its firm need, upgrades go into the shared rate base, and the data center pays 30% of a pro-rata share of existing wires.

Cost causation: AI brings 100% of its firm capacity at connection, pays its own upgrades, signs an 85% minimum-take contract and pays half of a pro-rata share of existing wires.

Flex + dividend: cost causation, with 25% of AI peak flexible and a Kilowatt Dividend of $10 per MWh on billed energy returned to households as a flat credit.

Anchor buyer: flex + dividend, and AI contracts new firm supply equal to 120% of its firm need, so the extra 20% (about 518 MW on this system) is sold into the market. The energy component declines 4% a year in real terms from 2027 as low-marginal-cost nuclear, geothermal, solar and storage displace higher-cost generation. Solar’s installed cost fell about 13.6% a year from 2010 to 2024, so 4% is a cautious assumption.

The full code for the model:

Anchor Buyer model: how AI load can be served, who pays, and what households see.

Three parts, all stylized and in real 2026 dollars:
  1. household_rates()  five ways of connecting AI load to a 10 GW utility
  2. speed_to_power()   a national stack of supply and demand levers against AI load
  3. learning_curve()   what AI pays as the first buyer of new nuclear, what follows

Run:  python3 anchor_model.py
"""
from dataclasses import dataclass, replace
import math

HOURS = 8760

# ---------------------------------------------------------------- 1. household rates
@dataclass
class System:
    base_peak_mw: float = 10_000
    base_supply_mw: float = 11_700      # 17% reserve margin
    base_energy_mwh: float = 50e6
    res_energy_mwh: float = 17.5e6
    res_peak_share: float = 0.45
    households: float = 1.667e6         # about 10,500 kWh per home per year
    base_growth: float = 0.005
    wires_rev_req: float = 4.5e9        # $/yr
    energy_c_kwh: float = 7.0
    reserve_req: float = 0.15
    cap_price_max: float = 325.0        # $/MW-day
    cap_price_min: float = 30.0
    rm_tight: float = 0.10
    rm_loose: float = 0.20
    dc_upgrade_per_mw_yr: float = 40_000
    dc_load_factor: float = 0.85

@dataclass
class Policy:
    name: str
    firm_share: float        # share of AI peak served firm (1 minus flexibility)
    supply_match: float      # new supply as a multiple of AI's firm need
    supply_lag: int          # years between load and matching supply
    dc_pays_upgrades: bool
    dc_wires_weight: float   # share of a pro-rata allocation of existing wires the DC pays
    min_take: float          # share of contracted energy billed regardless of use
    dividend_per_mwh: float  # Kilowatt Dividend on billed DC energy, $/MWh
    energy_decline: float    # real annual decline in the energy component from 2027
    scarcity_beta: float = 0.5

def dc_peak(year: int) -> float:
    """AI peak MW: 0 in 2026, 2,000 MW by 2031, 3,000 MW by 2035, flat after."""
    if year <= 2026:
        return 0.0
    if year <= 2031:
        return 400.0 * (year - 2026)
    if year <= 2035:
        return 2000.0 + 250.0 * (year - 2031)
    return 3000.0

def cap_price(rm: float, s: System) -> float:
    if rm <= s.rm_tight:
        return s.cap_price_max
    if rm >= s.rm_loose:
        return s.cap_price_min
    frac = (rm - s.rm_tight) / (s.rm_loose - s.rm_tight)
    return s.cap_price_max - frac * (s.cap_price_max - s.cap_price_min)

def household_rates(p: Policy, s: System = System(), start: int = 2026, end: int = 2040):
    rows = []
    for y in range(start, end + 1):
        g = (1 + s.base_growth) ** (y - start)
        base_peak = s.base_peak_mw * g
        dcp = dc_peak(y)

        dc_need = dc_peak(y - p.supply_lag) * p.firm_share * (1 + s.reserve_req)
        supply = s.base_supply_mw * g + p.supply_match * dc_need
        rm = supply / (base_peak + dcp * p.firm_share) - 1
        price = cap_price(rm, s)
        res_cap = s.res_peak_share * base_peak * (1 + s.reserve_req) * price * 365

        wires = s.wires_rev_req * g + (0 if p.dc_pays_upgrades else dcp * s.dc_upgrade_per_mw_yr)
        dc_energy = dcp * HOURS * s.dc_load_factor
        billed_dc = max(dc_energy, p.min_take * dcp * HOURS * s.dc_load_factor)
        res_energy = s.res_energy_mwh * g
        res_wires = wires * res_energy / (s.base_energy_mwh * g + p.dc_wires_weight * billed_dc)

        shortfall = max(0.0, s.reserve_req - rm) / s.reserve_req
        decline = (1 - p.energy_decline) ** max(0, y - 2027)
        energy_c = s.energy_c_kwh * decline * (1 + p.scarcity_beta * shortfall)

        res_kwh = res_energy * 1000
        rate_c = (res_cap + res_wires) / res_kwh * 100 + energy_c
        hh = s.households * g
        bill = rate_c / 100 * res_kwh / hh
        dividend = p.dividend_per_mwh * billed_dc / hh
        rows.append(dict(year=y, rm=rm, cap_price=price, rate_c=rate_c,
                         bill=bill, dividend=dividend, net_bill=bill - dividend))
    return rows

SOCIALIZE = Policy("Socialize", 1.0, 0.9, 3, False, 0.3, 0.0, 0, 0.0)
CAUSATION = Policy("Cost causation", 1.0, 1.0, 0, True, 0.5, 0.85, 0, 0.0)
FLEX_DIV = Policy("Flex + dividend", 0.75, 1.0, 0, True, 0.5, 0.85, 10, 0.0)
ANCHOR = Policy("Anchor buyer", 0.75, 1.2, 0, True, 0.5, 0.85, 10, 0.04)
POLICIES = (SOCIALIZE, CAUSATION, FLEX_DIV, ANCHOR)

# ---------------------------------------------------------------- 2. speed to power
# Each lever: (name, GW of nameplate or headroom, accreditation toward firm need, source note)
LEVERS_2032 = [
    ("Flexible headroom, deliverable share", 76 * 0.30, 1.00, "Duke: 76 GW at 0.25% curtailment; 30% assumed deliverable"),
    ("Reconductoring adds deliverable headroom", 76 * 0.15, 1.00, "assumed +15 points of the Duke headroom"),
    ("Solar + 4h storage on surplus interconnection", 25.0, 0.50, "UC Berkeley: ~500 GW solar potential; 5% built"),
    ("Nuclear restarts", 0.835 + 0.80 + 0.615, 0.95, "Crane, Palisades, Duane Arnold"),
    ("Nuclear uprates", 2.4, 0.95, "NRC expected applications through 2032"),
    ("Enhanced geothermal", 0.5, 0.90, "Fervo Cape Station phases I and II"),
]

LEVERS_2035_EXTRA = [
    ("New nuclear: SMRs and AP1000s", 6.6 + 5.5, 0.95, "Meta portfolio 6.6 GW; ~5 AP1000s"),
    ("More geothermal", 2.5, 0.90, "assumed industry scale-up"),
    ("More solar + storage on surplus interconnection", 25.0, 0.50, "another 5% of potential"),
]

def speed_to_power(gross_gw: float, overstated_gw: float, flexible_share: float, levers):
    """Firm GW still needed after demand hygiene, flexibility and each supply lever."""
    real = gross_gw - overstated_gw
    firm_need = real * (1 - flexible_share)
    steps = [("Utility-reported AI load", gross_gw),
             ("Less duplicate and speculative requests", real),
             (f"Firm need with {flexible_share:.0%} of peak flexible", firm_need)]
    remaining = firm_need
    for name, gw, accreditation, _ in levers:
        remaining -= gw * accreditation
        steps.append((name, remaining))
    return steps, max(0.0, remaining)

# ---------------------------------------------------------------- 3. learning curve
CAD_TO_USD = 15.1 / 20.9          # from the Darlington budget in both currencies
UNIT1_USD_KW = 6.1e9 / 300e3 * CAD_TO_USD   # Darlington unit 1
UNIT4_USD_KW = 4.1e9 / 300e3 * CAD_TO_USD   # Darlington unit 4
LEARNING_RATE = 1 - (UNIT4_USD_KW / UNIT1_USD_KW) ** 0.5   # two doublings, unit 1 to unit 4

def unit_cost(n: int, c1: float = UNIT1_USD_KW, lr: float = LEARNING_RATE) -> float:
    """Wright's law: cost falls by lr with each doubling of cumulative units."""
    return c1 * n ** math.log2(1 - lr)

def lcoe(capex_kw: float, wacc: float = 0.07, years: int = 30, cf: float = 0.93, om_fuel: float = 30.0) -> float:
    crf = wacc * (1 + wacc) ** years / ((1 + wacc) ** years - 1)
    return capex_kw * crf / (HOURS / 1000 * cf) + om_fuel   # $/MWh

def learning_curve(ai_units: int = 10, later_units: int = 30):
    ai_first = sum(unit_cost(n) for n in range(1, ai_units + 1))
    households_after = sum(unit_cost(n) for n in range(ai_units + 1, ai_units + later_units + 1))
    households_first = sum(unit_cost(n) for n in range(1, later_units + 1))
    return dict(lr=LEARNING_RATE, unit1=unit_cost(1), unit4=unit_cost(4),
                unit10=unit_cost(10), unit40=unit_cost(ai_units + later_units),
                saving_share=1 - households_after / households_first,
                saving_usd=(households_first - households_after) * 300e3)

def power_cost_per_gpu_hour(price_mwh: float, gpu_kw: float = 0.700, pue: float = 1.2) -> float:
    return gpu_kw * pue * price_mwh / 1000

# ---------------------------------------------------------------- 4. edge inference
GEMINI_WH_PER_PROMPT = 0.24      # Google, median text prompt, fleet-level accounting

def local_wh_per_prompt(j_per_token: float, tokens: int = 300) -> float:
    return j_per_token * tokens / 3600

def edge_shift(real_gw: float, shift: float, energy_ratio: float, inference_share: float = 0.55,
               flexible_share: float = 0.25, load_factor: float = 0.85, peak_factor: float = 1.5):
    """Move a share of AI inference from data centers to homes and devices.

    energy_ratio: energy per task at the edge relative to the cloud.
    peak_factor: how much higher household AI use is at the evening peak than on average.
    Returns firm GW avoided at data centers, GW added to the household peak, and the net.
    """
    moved = real_gw * inference_share * shift
    avoided = moved * (1 - flexible_share)
    added = moved * load_factor * energy_ratio * peak_factor
    return avoided, added, avoided - added

def edge_break_even_ratio(flexible_share: float = 0.25, load_factor: float = 0.85, peak_factor: float = 1.5) -> float:
    return (1 - flexible_share) / (load_factor * peak_factor)

if __name__ == "__main__":
    print("1. Household rates")
    for pol in POLICIES:
        rows = household_rates(pol)
        pick = {r["year"]: r for r in rows}
        print(f"  {pol.name:16s} rate 2030 {pick[2030]['rate_c']:.2f}  2040 {pick[2040]['rate_c']:.2f} c/kWh | "
              f"net bill 2030 ${pick[2030]['net_bill']:,.0f}  2040 ${pick[2040]['net_bill']:,.0f} | "
              f"15-yr ${sum(r['net_bill'] for r in rows):,.0f} | RM 2040 {pick[2040]['rm']:.1%} "
              f"cap ${pick[2040]['cap_price']:.0f}")

    print("\n2. Speed to power (firm GW still needed)")
    for label, gross, over in (("Forecast at face value", 90, 0), ("De-duplicated forecast", 90, 25)):
        steps, gap32 = speed_to_power(gross, over, 0.25, LEVERS_2032)
        _, gap35 = speed_to_power(gross, over, 0.25, LEVERS_2032 + LEVERS_2035_EXTRA)
        print(f"  {label}:")
        for name, v in steps:
            shown = f"{v:6.1f}" if v >= 0 else f"{-v:6.1f} surplus"
            print(f"    {name:48s} {shown}")
        print(f"    gas bridge needed by 2032: {gap32:.1f} GW, by 2035 with new firm clean: {gap35:.1f} GW")

    print("\n   Sensitivity: gas bridge GW in 2032 / 2035")
    print("    real load  flexible  deliverable headroom + reconductoring")
    for gross, over in ((90, 0), (90, 25)):
        for flex in (0.0, 0.25):
            cells = []
            for deliv, recond in ((0.10, 0.0), (0.20, 0.10), (0.30, 0.15)):
                lv = list(LEVERS_2032)
                lv[0] = (lv[0][0], 76 * deliv, 1.0, "")
                lv[1] = (lv[1][0], 76 * recond, 1.0, "")
                g32 = speed_to_power(gross, over, flex, lv)[1]
                g35 = speed_to_power(gross, over, flex, lv + LEVERS_2035_EXTRA)[1]
                cells.append(f"{deliv:.0%}+{recond:.0%}: {g32:4.1f}/{g35:4.1f}")
            print(f"    {gross - over:5d} GW   {flex:4.0%}     " + "   ".join(cells))

    print("\n3. Learning curve and affordability")
    lc = learning_curve()
    print(f"  learning rate per doubling {lc['lr']:.1%}; $/kW unit 1 {lc['unit1']:,.0f}, unit 4 {lc['unit4']:,.0f}, "
          f"unit 10 {lc['unit10']:,.0f}, unit 40 {lc['unit40']:,.0f}")
    print(f"  LCOE unit 1 ${lcoe(lc['unit1']):.0f}/MWh, unit 10 ${lcoe(lc['unit10']):.0f}, unit 40 ${lcoe(lc['unit40']):.0f}")
    print(f"  households buying units 11-40 instead of 1-30 save {lc['saving_share']:.0%} "
          f"(${lc['saving_usd']/1e9:.1f}B on 9 GW)")
    for label, price in (("industrial power $80/MWh", 80), ("FOAK nuclear $175/MWh", 175),
                         ("FOAK nuclear $255/MWh", 255), ("Kilowatt Dividend $10/MWh", 10)):
        c = power_cost_per_gpu_hour(price)
        print(f"  {label:28s} ${c:.3f} per GPU-hour = {c/1.99:.1%} of a $1.99 rental")

    print("\n4. Edge inference")
    for label, jpt in (("0.5B model on a consumer GPU", 0.2747), ("7B model on a consumer GPU", 0.2747 * 8.6),
                       ("4B model on a DGX Spark, batched", 0.058)):
        wh = local_wh_per_prompt(jpt)
        print(f"  {label:34s} {wh:.3f} Wh per 300-token answer = {wh / GEMINI_WH_PER_PROMPT:.2f}x a median Gemini prompt")
    for label, watts in (("DGX Spark idle", 45), ("Mac Studio idle", 9)):
        kwh_day = watts * 24 / 1000
        print(f"  {label:34s} {kwh_day:.2f} kWh/day = {kwh_day * 1000 / GEMINI_WH_PER_PROMPT:,.0f} Gemini prompts")
    print(f"  break-even energy ratio: {edge_break_even_ratio():.2f}")
    for ratio in (0.25, 0.5, 1.0, 2.0):
        a, b, n = edge_shift(65, 0.20, ratio)
        print(f"  20% of inference to the edge, ratio {ratio:4.2f}: data center firm -{a:.1f} GW, "
              f"household peak +{b:.1f} GW, net grid relief {n:+.1f} GW")

A few choices deserve explanation. The capacity price follows a straight-line demand curve between $30 at a 20% reserve margin and the $325 cap at 10%, a simplified version of PJM’s curve. The scarcity premium on energy rises with the shortfall below the reserve requirement. Minimum-take billing means AI pays wires and the dividend on 85% of contracted energy even if it uses less, which is what protects households when a project underperforms.

The household results, rounded:

Under socialization, the residential rate climbs 24% by 2030. Under the anchor regime, it falls 31% by 2040 and the net bill falls 38%. In 2030 the gap between the two is $765 a year per household, and over 15 years it is $10,392. Even with no decline in energy costs at all, the anchor regime takes the 2040 rate to 14.77 cents, 15% below today, because AI-funded surplus capacity pushes the capacity price to its floor.

The most sensitive parameter is w, AI’s contribution to existing wires costs. Under cost causation, at w = 0 the 2040 rate is 17.42 cents, essentially flat. At w = 0.5 it is 15.87, and at w = 1.0 it is 14.77. Paying your own way keeps households where they are. Helping carry the system you join is what makes them better off. That matches LBNL’s finding that from 2019 to 2025 the states with the fastest load growth generally saw real all-sector prices fall, by more than 1 cent per kWh.

For a household earning $40,000, electricity burden in 2040 ranges from 4.63% under socialization to 2.81% under the anchor regime if real income stays flat, and from 4.03% to 2.44% with 1% annual real income growth. For a median household at $83,730 with the same growth, it ranges from 1.92% to 1.17%.

Where this meets the Autonomy Economy

In The Autonomy Economy I proposed a framework for a world where AI and robotics make most labor optional, funded by a Public Trust that owns a share of automated capital and pays a universal Autonomy Dividend. Version 2 added a deflation channel and assumed energy prices would fall about 75% by 2040.

The model here makes me revise that, and I would push back on anyone, including me, who treats generation costs and retail bills as the same thing. Wires and capacity are about 60% of the residential rate. Even with the energy component falling 41% by 2040 and AI funding surplus firm capacity, the best regime lowers the delivered rate by 31% and the net bill by 38%. That is a large gain for households. It is not 75%.

The more important result runs the other way. The task-level evidence for AI productivity is strong: Brynjolfsson, Li and Raymond found that a generative AI assistant raised customer support productivity by 14% on average and 34% for novice workers. The macro range is wide, from Goldman Sachs’s estimate of a 7% lift to global GDP and 1.5 percentage points of annual productivity growth over a decade, to Daron Acemoglu’s estimate of no more than 0.66% total factor productivity growth over ten years. If the optimists are right and non-energy prices fall by half while electricity does not, electricity’s share of household spending rises from 2.4% to about 4.7%. In a deflationary AI economy, energy becomes the binding input for real purchasing power, and the gains from cheaper goods can be clawed back through the one bill nobody can skip.

That also clarifies what the money under socialization is. In the model’s 2030 socialized case, households pay about $382 million more for capacity than under cost causation, about $225 per household, and almost all of it goes to generators that already exist. It is a scarcity rent, a transfer from households to incumbent capital created by a rule rather than by any new cost of production. It is the same pattern I have been working on in a separate piece on rent extraction, where income flows to whoever owns a bottleneck rather than to whoever adds to output.

The anchor regime reverses that flow twice. AI’s money builds new capacity instead of bidding up old capacity, and the Kilowatt Dividend returns part of AI’s payment to the people who financed the grid through decades of rates. At $10 per MWh, the dividend raises about $223 million a year on this system, about $125 per household by 2040, and adds $0.008 per GPU-hour, 0.4% of the lowest rental price. It is small next to the Autonomy Dividend, but it is a working version of the same mechanism built from tariff tools that exist today.

Six rules that make it work

The first rule is cost causation with a minimum take. In July 2025 Ohio regulators approved an AEP Ohio tariff for new loads above 25 MW that bills at least 85% of contracted capacity for up to 12 years after a four-year ramp, with exit fees and financial guarantees. The queue collapse that followed is the rule working.

The second rule is to bring your own capacity, firm or flexible. Texas’s SB 6, signed in June 2025, requires new loads of 75 MW or more to install equipment that lets utilities disconnect them during firm load shed, and creates a program for large loads to ramp down or switch to backup generation. On June 18, 2026, FERC issued show cause orders to six regional grid operators to address large loads, covering cost shifts, co-located and behind-the-meter generation, and new transmission services for flexible loads.

The third rule is to make flexibility the fast lane. A data center that accepts curtailment for a fraction of a percent of its energy should connect years sooner than one that demands firm service on day one. Pair that with reconductoring and surplus interconnection so the headroom is deliverable where the load is.

The fourth rule is that large loads contribute to embedded costs, not just their own upgrades. The sensitivity shows that this is the difference between holding households flat and lowering their bills.

The fifth rule is anchor procurement. Regulators can give the fastest interconnection to large loads that contract new firm clean supply above their own firm need, in particular restarts, uprates, geothermal and early units of new reactor designs, and require the surplus to be offered into the market. That turns AI’s willingness to pay into the first-of-a-kind premium that new technologies need.

The sixth rule is the Kilowatt Dividend: a per-MWh charge on billed large-load energy, held in a state fund and returned as an equal credit per residential account or weighted toward low-income households.

Voluntary commitments help but do not replace these rules. Microsoft’s January 2026 Community-First plan pledged that its power costs would not be passed to residential customers. On March 4, seven large AI and technology companies signed a White House ratepayer protection pledge to secure their own power, pay for grid upgrades and accept rates regardless of usage, and by July the White House reported more than 200 signers. The pledge is non-binding. As one Georgia commissioner pointed out, it does nothing about a contract that is terminated and leaves other customers holding the generation built for it. Only tariffs approved by state commissions and grid operators can enforce these terms.

Where it breaks

All three models are stylized and the results are directional. The speed-to-power stack depends heavily on how much flexible headroom is deliverable where data centers are, a number nobody has measured well. The household model depends on the capacity demand curve, the scarcity premium and AI’s wires contribution, all of which are my assumptions. The code is short so you can change them.

Nuclear timelines slip. Palisades missed its restart date, PJM has suggested Crane’s grid connection could come years after the plant is ready, and no new reactor design has yet been built in the US on schedule. The learning curve is a projection, and US nuclear history mostly runs the wrong way. An anchor buyer improves the odds without guaranteeing them.

Flexibility has been tested mainly on training and batch work. As inference grows as a share of AI compute, the flexible share may shrink below 25%.

Gas bridges have a way of becoming destinations. Plants built for a 2030 shortfall will run for decades unless contracts say otherwise. The IEA projects data center emissions roughly doubling to about 350 million tonnes by 2035 under current trends, and local water use is a separate fight in every dry region.

Data centers are not the only reason bills rise. LBNL’s work points to wildfire mitigation, storm recovery, gas prices and distribution spending as major drivers. And in a multi-state market like PJM, a data center in Virginia raises capacity prices in Illinois, whose regulators have no say over the Virginia tariff. State rules will leak until the grid operator’s rules close the gap.

The path forward

None of this needs new technology or new federal law. State regulators can require minimum-take contracts, own-upgrade payment and embedded-cost contributions for every new load above 25 to 75 MW, and price flexible service as the cheaper and faster option. Utilities can publish their large-load queues with the share backed by financial commitments, so that capacity auctions price load that is actually coming. Grid operators answering FERC’s orders can build flexible interconnection that uses existing headroom before procuring capacity at the cap, and can fast-track reconductoring and surplus interconnection. AI companies can treat restarts, uprates, geothermal and early reactor units as the price of fast interconnection, because the arithmetic says they can afford it. And anyone who pays an electric bill can find their utility’s large-load tariff docket and file a comment, because these cases are decided in proceedings most households never see.

There is plenty of power for AI. What has been missing are rules that make it pay its way. AI does not need cheap power. It needs fast, reliable power and can afford to pay for it. Households need affordable power and cannot easily cut back. A grid designed around that difference lets AI pay for the expensive early part of the next power system, and lets everyone else inherit it.

About the Author

Brian James Curry is a Kansas City AI researcher and data scientist, the founder of Vector1 Research, an independent applied AI lab, and of KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group and grew to more than 1,500 members. He has more than twenty years of experience leading analytics, data science and machine learning at enterprises including Hallmark, Garmin, Koch Industries, Tractor Supply Company and Vail Resorts. His open-source work, including Papilon, PyCausalSim, MeaningFlow, Daedalus and tincast, is on GitHub. He grew up in a small Kansas farming town and attended Kansas State University. He writes about agentic AI, causal inference and the economics of AI on Medium and Substack. Contact: brian@vector1.ai.

Sources

Related research