AI Capital Cycle ⑥ — When Watts Beat Chips

September 4, 1882. Thomas Edison powered up the world's first commercial generating station on Pearl Street in Manhattan. In its first days, it lit 400 light bulbs for 85 customers. But that small beginning meant one thing — *from that moment, the binding variable of the Industrial Revolution became the watt*. Not how big the factory was. How much electricity it could draw.

The 2026 AI capital cycle has arrived at the same inflection. The binding variable has shifted from chips to watts.

Even if NVIDIA ships more GPUs every quarter, no electricity means no quarterly revenue. Even if Microsoft promises $500B to Stargate, no transmission line means the promise stays on a spreadsheet. So the 2026 AI cycle is defined not by Big Tech CFO capex announcements but by *PJM capacity auction results and ERCOT interconnection queue lengths*.

The IEA's 2026 Electricity report quantifies this. Global data center power demand doubles from 485 TWh in 2025 to 950 TWh in 2030. The US and China account for 80% of that growth. EPRI raised its forecast of US data center electricity share from 4% (2023) to 9-17% by 2030 — a 60% upward revision.

What this means: power is no longer an *auxiliary variable* of the AI cycle. It is *the binding variable*.

One number shows that power has become the binding variable. The US interconnection queue.

The queue is — a backlog of projects applying to build new generation and connect to the grid. Lawrence Berkeley National Lab's "Queued Up 2025" measures its length.

As of end-2024: 2,290 GW. That is nearly 2x all currently operating US generation capacity (~1,300 GW). 10,300 projects. Of those, 1,400 GW generation + 890 GW storage.

What this means: *the US has plenty of capacity standing in line to build more electricity*. The line just doesn't move. Median time from interconnection application to COD (commercial operation date) was <2 years in 2000-2007. By 2018-2024 it was 4+ years. In 2024 alone, 112 GW of solar/storage projects *withdrew* from the queue — they couldn't wait that long.

This backlog must unclog for data centers to get new power. But it's not unclogging. *It's getting longer*. FERC, PJM, ERCOT, MISO — each ISO attempts queue reform, but data center demand growth outpaces reform speed.

Once power became the binding variable, Big Tech did one thing — *they started buying nuclear directly*.

September 20, 2024. Microsoft signed a 20-year PPA with Constellation Energy. Target: Three Mile Island Unit 1, 835MW. The plant where Unit 2's 1979 partial meltdown stopped the US nuclear industry for a generation. Unit 1 ran separately until economic shutdown in 2019. MSFT's PPA wakes it. Restart planned 2027 (one year earlier than prior schedule). DOE backed it with $1B in loans.

Three months later. AWS signed a 17-year PPA with Talen Energy. Target: Susquehanna nuclear, 1.92 GW. Separately, AWS bought Talen's data center campus for $650M. "Co-located in-front-of-the-meter" structure — nuclear powers the data center directly, bypassing the grid. (This requires a FERC regulatory workaround. It became a political issue in 2025.)

Google: contract with Kairos Power for 7 SMRs, 500MW combined. Amazon: 5GW SMR commitment with X-Energy. Meta: LOI with Vistra.

What all these deals have in common: *the unit economics of electricity changed*. Typical PPA prices of $40-50 per MWh rose to $80-150 for data-center-dedicated supply. Big Tech is promising to pay more for electricity — more stable, longer term, cleaner. That promise justifies the operator's cost of capital and wakes shuttered nuclear plants.

Industrial policy didn't make this happen. *The electricity-price market did*.

In a power cycle, the first prices to rise weren't at the power plants. They were *the infrastructure that delivers and stabilizes that power to the data center*.

Vertiv (VRT). Data center cooling, UPS, power management. Stock from ~$15 end-2022 to ~$130 by 2025 (8.5x). Revenue $5.7B (2022) to ~$10B (2025). As AI data centers ran 5-10x denser power per square foot than conventional servers, liquid cooling became mandatory. Vertiv is the standard. Margins expanded (2025 operating margin ~22%).

Eaton (ETN). Power management, transformers, data center backup. Revenue $20.8B (2022) → ~$28B (2025). *Transformer lead time is 4+ years*. Order today, receive in four years. Eaton's backlog is a single visibility metric for the cycle.

Quanta Services (PWR). Transmission line construction and maintenance. Revenue $17.1B (2022) → ~$28B (2025). For the interconnection queue to unclog, more transmission must be built. Quanta builds it. Backlog end-2025: ~$36B (+35% YoY).

GE Vernova (GEV). Gas turbines, transmission equipment. Spun off from GE in April 2024. Market cap $30B at spin → ~$95B by 2025. As AI data centers brought baseload natural gas back, GE Vernova's gas turbine backlog grew +40% YoY.

Hitachi Energy. Sole global #1 in transformers. Backlog end-2025: $30B+. Transformer lead time touches 6 years in some quarters.

What these companies share: *they aren't glamorous like NVIDIA, but no data center runs without them, even after the GPU arrives*. They are the second-row beneficiaries of the cycle.

Korea is a miniature of the US power crisis — a smaller territory with even more concentrated demand.

The Yongin Semiconductor Cluster. Samsung and SK Hynix to invest KRW ₩600 trillion+ over 30 years. To operate, by 2053 it will need 10+ GW — the equivalent of 10 reactors. Current Yongin-area data center and semiconductor supply: 0.6 GW. The gap is over 9 GW.

KEPCO is already at the limit of grid load. 55%+ of new transmission projects are delayed (as of October 2025). 2013-2023: transmission grew only +14%, distribution +22% — less than half of demand growth. Time-to-grant for new data center power supply: KEPCO median was 2-3 months, now stretched to 12 months.

Two solutions are in progress.

First, Grid 3-Laws passed (2025). Private capital allowed to participate in transmission construction — moving from KEPCO-sole to KEPCO + private consortium. Short-term acceleration possible.

Second, i-SMR (innovative small modular reactor). Joint KHNP + KAERI development. 170 MWe. Standard design completed end-2025; standard design approval targeted 2028. SMR Special Act passed allowing direct data-center PPAs. This is Korea's version of the MSFT-Three Mile Island structure. Doosan Enerbility is the sole supplier of the SMR pressure vessel.

From a Korean market view: power infrastructure beta works in the Korean capital market too. LS ELECTRIC (transformers), Doosan Enerbility (SMR), Korea Engineering, HD Hyundai Electric (transformers). There's no single megacap like Vertiv or Eaton in Korea — but the same layer rides the same cycle.

The Korean data center market itself is exploding. Mordor Intelligence: $0.58B (2025) → $1.89B (2030). IT load 1.96 → 6.32 thousand MW. That demand is added on top of the Korean grid.

The single proposition of this memo: the binding variable of the AI capital cycle is not chips but watts.

In a quarter where NVIDIA can make more GPUs — if the US grid can't supply more electricity, revenue doesn't grow accordingly. Microsoft restarting Three Mile Island, AWS buying the Susquehanna campus, Google contracting Kairos for SMRs — every one of these deals is *recognition from a different angle of the same fact: the unit economics of electricity changed*.

What Pearl Street did in 1882 — making watts the binding variable of industry — is happening again in 2026 with AI. This time faster, more expensively, more globally.

Two implications for investors.

First, power infrastructure beta can be longer than chip beta. Just as 1.6T optical transceivers lag the GPU cycle by 2-4 quarters, power infrastructure cycles are longer — transformer 4-yr lead time, transmission 5-10 yr build, nuclear restart 3 yr, first SMR 5-7 yr. *A longer cycle means longer revenue visibility*. That's why Vertiv, Eaton, Quanta, Constellation, and Doosan Enerbility's 5-year backlogs matter.

Second, the first crack of the crisis comes from transmission cost spikes. If the next PJM capacity auction explodes another +500%+ — that's *the point where data center capex ROIC math is threatened*. The first quarter Big Tech capex guidance acknowledges "power cost" as a risk factor is the quarter the cycle's cost of capital starts to be pressured.

*Since Pearl Street in 1882, industry has been measured in watts. The 2026 AI cycle is no different.*