Power and hyperscalers

Behind the AI thesis sits a more physical one: those data centers run on electricity, and after two decades of flat US power demand, AI (alongside electrification and reshoring) turned the grid into a bottleneck. This is the classic "sell picks in a gold rush" framing, applied to the least glamorous, most physical layer of the AI stack, and it is a useful teaching thesis because the constraint is concrete and measurable.

The short version

AI data centers need enormous, reliable, always-on power, and they need it faster than grids can build. That has revived demand for utilities, independent power producers, natural gas, nuclear (including restarts and small modular reactors), grid equipment (transformers, switchgear), and cooling. The bull case is a decade of structural load growth after twenty flat years, benefiting boring regulated assets that rarely get a growth story. The bear case is that the AI capex it depends on may slow, that grid build-outs are slow and politically fraught, and that some "AI power" names ran on enthusiasm ahead of contracted demand. The binding real constraint (interconnection queues, permitting) is also the thesis's clearest gauge.

The second-order idea

When a boom has a physical input everyone needs and few can supply, the input's owners can be a better risk-adjusted bet than the boom itself: less hype, clearer economics, a moat made of permits and physics. AI's physical input is power. A frontier data center can draw as much electricity as a small city, must run 24/7, and cannot wait years for a grid connection, which collides with a US power system built for flat demand and slow to expand. The thesis: whoever supplies, moves, or equips that power participates in the AI build with a different (often more durable) risk profile than the chips and models.

The scale of the demand

Fact	Rough magnitude	Why it matters
US electricity demand growth, 2005-2022	~flat	Utilities had NO growth story for two decades
Data center share of US power, mid-2020s	~4%, projected toward high single digits+	A step-change in a system that plans in decades
A single frontier training campus	hundreds of MW to GW-scale	Equivalent to a mid-size city's peak load, at one site
Grid interconnection queue waits	years	The bottleneck is TIME, not just capacity
Transformer / switchgear lead times	extended into multi-year	Equipment makers gained pricing power they never had

Who benefits, link by link

Regulated utilities in data-center corridors: load growth means rate-base growth, the regulated utility's only real growth engine (see the industry method in Other Intrinsic Methods). Boring assets with a sudden secular tailwind.
Independent power producers selling into tight markets: direct exposure to power PRICES and to premium contracts with hyperscalers, including nuclear operators signing long-term deals.
Natural gas as the practical bridge fuel for firm, dispatchable baseload while renewables and nuclear scale: a commodity-demand thesis inside the power thesis.
Nuclear: restarts of shuttered plants, uprates, and small modular reactors, all suddenly financeable because hyperscalers will sign to pay for carbon-free firm power.
Grid equipment (transformers, switchgear, cabling) and cooling: the shortage economics are the cleanest, with multi-year backlogs and restored pricing power.

The bull case

Structural, not cyclical, load growth after twenty flat years, in assets the market habitually prices for no growth.
Physical moats. Permits, rights-of-way, interconnection rights and existing generation cannot be conjured with capital; scarcity is durable.
Diversified payoff. The power constraint is real across a WIDE range of AI outcomes: even a more modest AI future still needs materially more electricity, so the thesis does not require the maximal AI bull case.
Regulated visibility. Utility cash flows are contracted and rate-based, so the growth arrives with less speculation than the chip layer.

The bear case

It inherits the AI capex risk. If the AI build-out slows (the monetization question in the AI thesis), projected load growth is revised down and the derived demand evaporates: the second-order trade is downstream of the first-order risk.
Grid build-outs are slow and political. Permitting, NIMBY opposition and rate-payer politics (who pays for the upgrades?) can delay or dilute the benefit for years.
Enthusiasm ran ahead of contracts. Some "AI power" names rerated on the story before signing durable demand; the precedent-transactions discipline applies (an event-priced value needs an event probability).
Efficiency risk. If AI compute becomes dramatically more energy-efficient per unit of output, the power demand curve bends below the straight-line projections the bull case extrapolates.

The tell of a good second-order thesis

A second-order trade earns its keep only if it is genuinely more durable OR cheaper than the first-order one, not merely less famous. The power thesis passes on durability (electricity is needed across more AI scenarios than any single chipmaker's margin survives) but fails to escape the FIRST-order risk entirely: if AI capex craters, so does the derived power demand. The honest framing is "a somewhat de-risked, somewhat cheaper way to express a related view," not "a free lunch adjacent to the boom."

What to watch

Hyperscaler power-purchase agreements: actual signed contracts (especially nuclear and long-term deals) convert story into contracted demand.
Interconnection queue movement and permitting reform: the bottleneck is time, so anything that speeds or stalls connection is the real catalyst.
The same capex-to-revenue gauge as the AI thesis: the upstream risk that governs everything downstream.
Equipment backlogs and lead times: the cleanest real-time read on whether the shortage economics are intensifying or normalizing.
Regulatory decisions on cost allocation: who pays for the grid upgrades determines how much of the growth reaches shareholders versus rate-payers.

Glossary for this guide

Total addressable market: The revenue available if a company served every possible customer. Useful as a ceiling and abused as a pitch; the discipline is asking what fraction is realistically serviceable and at what cost.
Capex: Cash spent on long-lived assets: plants, equipment, software. Maintenance capex keeps the current business running; growth capex expands it. Both reduce free cash flow now in exchange for cash flows later.
Moat: A structural barrier that stops a company's high returns from being competed away: network effects, switching costs, brands and patents, cost advantage, or efficient scale. The durable question is never whether a business is profitable but what protects the profits.
Supply-demand balance: The commodity analyst's model: production plus inventory change must equal consumption, quarter by quarter. Price is the negotiator that keeps the identity true when the balance tightens.
Return on equity: Net income divided by shareholders' equity: what the owners earn on the capital they have in the business. High ROE sustained for years is the signature of a strong business or heavy leverage; the analysis is telling those apart.
Catalyst: The event expected to force the market to reprice toward your view: earnings, a spin-off, a regulatory decision, a contract. Cheap without a catalyst can stay cheap for years; the catalyst is the thesis's clock.
Falsifier: The observable fact, chosen in advance, that would prove the thesis wrong and trigger exit. Deciding it while calm is cheap; deciding it mid-drawdown is expensive, which is why professionals write it down first.
Book value: Shareholders' equity as the balance sheet states it: assets minus liabilities. Meaningful where assets are financial or tangible (banks, insurers), nearly meaningless where the real assets are brands and code.
Dividend discount model: Valuing a share as the present value of its future dividends, usually with the growing-perpetuity formula. The natural method for payout-defined businesses: utilities, telecoms, and banks.

Power and hyperscalers

The short version

The second-order idea

The scale of the demand

Fact	Rough magnitude	Why it matters
US electricity demand growth, 2005-2022	~flat	Utilities had NO growth story for two decades
Data center share of US power, mid-2020s	~4%, projected toward high single digits+	A step-change in a system that plans in decades
A single frontier training campus	hundreds of MW to GW-scale	Equivalent to a mid-size city's peak load, at one site
Grid interconnection queue waits	years	The bottleneck is TIME, not just capacity
Transformer / switchgear lead times	extended into multi-year	Equipment makers gained pricing power they never had

Who benefits, link by link

Regulated utilities in data-center corridors: load growth means rate-base growth, the regulated utility's only real growth engine (see the industry method in Other Intrinsic Methods). Boring assets with a sudden secular tailwind.

Independent power producers selling into tight markets: direct exposure to power PRICES and to premium contracts with hyperscalers, including nuclear operators signing long-term deals.

Natural gas as the practical bridge fuel for firm, dispatchable baseload while renewables and nuclear scale: a commodity-demand thesis inside the power thesis.

Nuclear: restarts of shuttered plants, uprates, and small modular reactors, all suddenly financeable because hyperscalers will sign to pay for carbon-free firm power.

Grid equipment (transformers, switchgear, cabling) and cooling: the shortage economics are the cleanest, with multi-year backlogs and restored pricing power.

The bull case

Structural, not cyclical, load growth after twenty flat years, in assets the market habitually prices for no growth.

Physical moats. Permits, rights-of-way, interconnection rights and existing generation cannot be conjured with capital; scarcity is durable.

Diversified payoff. The power constraint is real across a WIDE range of AI outcomes: even a more modest AI future still needs materially more electricity, so the thesis does not require the maximal AI bull case.

Regulated visibility. Utility cash flows are contracted and rate-based, so the growth arrives with less speculation than the chip layer.

The bear case

It inherits the AI capex risk. If the AI build-out slows (the monetization question in the AI thesis), projected load growth is revised down and the derived demand evaporates: the second-order trade is downstream of the first-order risk.

Grid build-outs are slow and political. Permitting, NIMBY opposition and rate-payer politics (who pays for the upgrades?) can delay or dilute the benefit for years.

Enthusiasm ran ahead of contracts. Some "AI power" names rerated on the story before signing durable demand; the precedent-transactions discipline applies (an event-priced value needs an event probability).

Efficiency risk. If AI compute becomes dramatically more energy-efficient per unit of output, the power demand curve bends below the straight-line projections the bull case extrapolates.

The tell of a good second-order thesis

What to watch

Hyperscaler power-purchase agreements: actual signed contracts (especially nuclear and long-term deals) convert story into contracted demand.

Interconnection queue movement and permitting reform: the bottleneck is time, so anything that speeds or stalls connection is the real catalyst.

The same capex-to-revenue gauge as the AI thesis: the upstream risk that governs everything downstream.

Equipment backlogs and lead times: the cleanest real-time read on whether the shortage economics are intensifying or normalizing.

Regulatory decisions on cost allocation: who pays for the grid upgrades determines how much of the growth reaches shareholders versus rate-payers.

Glossary for this guide

Total addressable market: The revenue available if a company served every possible customer. Useful as a ceiling and abused as a pitch; the discipline is asking what fraction is realistically serviceable and at what cost.
Capex: Cash spent on long-lived assets: plants, equipment, software. Maintenance capex keeps the current business running; growth capex expands it. Both reduce free cash flow now in exchange for cash flows later.
Moat: A structural barrier that stops a company's high returns from being competed away: network effects, switching costs, brands and patents, cost advantage, or efficient scale. The durable question is never whether a business is profitable but what protects the profits.
Supply-demand balance: The commodity analyst's model: production plus inventory change must equal consumption, quarter by quarter. Price is the negotiator that keeps the identity true when the balance tightens.
Return on equity: Net income divided by shareholders' equity: what the owners earn on the capital they have in the business. High ROE sustained for years is the signature of a strong business or heavy leverage; the analysis is telling those apart.
Catalyst: The event expected to force the market to reprice toward your view: earnings, a spin-off, a regulatory decision, a contract. Cheap without a catalyst can stay cheap for years; the catalyst is the thesis's clock.
Falsifier: The observable fact, chosen in advance, that would prove the thesis wrong and trigger exit. Deciding it while calm is cheap; deciding it mid-drawdown is expensive, which is why professionals write it down first.
Book value: Shareholders' equity as the balance sheet states it: assets minus liabilities. Meaningful where assets are financial or tangible (banks, insurers), nearly meaningless where the real assets are brands and code.
Dividend discount model: Valuing a share as the present value of its future dividends, usually with the growing-perpetuity formula. The natural method for payout-defined businesses: utilities, telecoms, and banks.