Energy Isn’t the Sixth Pillar. It’s the Fifth One in a Hard Hat

The Fifth Pillar with a Hard Hat

A reader called me out on the energy question. He was right, and the correction is more interesting than the mistake.

The callout

This week, under a post on sovereignty scoring, a reader made a point I brushed off too quickly. Their argument: the kill switch has a hardware equivalent. In Texas, large loads interconnecting after December 2025 must have equipment in place that lets the grid cut them during an emergency. Shutoff hardware as a condition of service, written into state law.

They’re right about the substance. Texas SB6 (2025), codified at PURA §39.170, requires utilities to develop a curtailment protocol — including installing any necessary equipment before interconnection — so that large loads of 75 MW or more connecting on or after December 31, 2025 can be curtailed during firm load-shed events. (Two precisions: the utility executes it on ERCOT’s instruction, and the PUCT’s implementing rule isn’t final yet.)

My reply was that energy is one of only two external dependencies I don’t grade harshly. The other is not owning a chip fab.

My reasoning was the same for both. At the accelerator layer, Nvidia is effectively 1:1 — everyone’s models run on the same silicon, out of the same handful of fabs, and nobody in the enterprise market is going to change that. A dependency every competitor carries identically isn’t a differentiator. Grading it marks the whole field down by the same amount and tells you nothing you can act on. Power, I assumed, was that kind of dependency too: universal, un-ownable, therefore nice-to-have rather than must-have.

Two things came to my attention once it was brought up.

On August 3, 2026, Governor Abbott directed an audit of data center interconnection requests in ERCOT. ERCOT delayed its Batch Zero eligibility classifications while it builds the audit process — roughly 300 data centers of 75 MW or larger, out of a queue of about 474 GW that is ~90% data centers. ERCOT’s all-time peak demand record is 91,089 MW, set July 22, 2026. The queue is 5.2x the largest hour the grid has ever served.

On August 13, 2026, PJM filed at FERC (Docket ER26-3515) to create an Interim Resource Adequacy Service. New large loads — 50 MW or more of cumulative peak at a single site — that aren’t in service with their own secured generation by June 1, 2027 become subject to IRAS curtailment, triggered before PJM calls its paid demand-response program. New load goes to the front of the cut line.

So: the two largest grids in the United States both moved, in the same month, to convert data center power from an entitlement into a conditional, revocable service.

I was wrong that this is nice-to-have. But I was wrong in a specific way that matters, and it isn’t the way the reader thinks.

Energy is not a sixth pillar

The temptation is to bolt an Energy pillar onto the framework and move on. Don’t. Energy fails the test for being its own pillar, because it isn’t a distinct control surface — it’s a load-bearing wall running through all five.

Territorial — where the workload physically sits. Power decides this, not your architecture diagram. Ireland’s CRU decided on December 12, 2025 that new data centres must provide generation or storage matching their maximum import demand — and bid it into the wholesale market — plus source at least 80% of annual demand from newly built Irish renewables. Existing REFIT, RESS and ORESS capacity doesn’t count. Ireland doesn’t just require you to generate; it conscripts you into the market. Singapore has rationed in tranches since its 2019 moratorium, most recently 300 MW general plus 200 MW reserved for green operators. And Israel’s Electricity Authority froze grid-connection processing for data centers of 8 MVA and up for 140 days from late July 2026 — against ~27,000 MW of pending requests on a grid whose average consumption is ~9,000 MW.

Operational — who dispatches your uptime. Increasingly, not you. DOE Emergency Order 202-26-23, effective May 18–20, 2026, authorized PJM to call on backup generation at data centers and other large loads before rolling blackouts. And in late July 2026, in PUCT Docket 59220, Texas held that curtailment obligations are not capped by the nameplate of the generator behind your meter: ~525 MW of co-located load paired with a 265.5 MW wind facility must be fully curtailable within 30 minutes, uncompensated — and that load is barred from paid demand-response participation. Building your own generation does not buy you out. And NERC is now drafting a “Computational Load Entity” registration category — 1 MW of IT equipment inside a 20 MW connected load at 60 kV or above — with ride-through and ramp-rate obligations. NERC has historically registered asset owners. It is now moving to register the load.

Technological — which stack you’re allowed to run. A GB200 NVL72 rack is specced at 132 kW. Uptime Institute’s 2026 survey puts modal installed rack density at 11 kW. Your power envelope, not your procurement team, decides which hardware generation you can adopt at all.

Legal — who writes your terms, and how often they rewrite them. SB6. The CRU decision — where, despite the “Large Energy User” title, the 80% rule applies to data centres specifically. Dominion’s GS-5 rate class, approved by the Virginia SCC in Case No. PUR-2025-00058. In the EU, Delegated Regulation 2024/1364 puts any data centre at 500 kW or more of installed IT power on an annual reporting obligation to a central European database.

Financial — where all of it lands. Every constraint above reaches your company as a number on an invoice or a clause in a contract.

That last line is the whole point. Energy touches all five pillars, but it settles in the fifth. Because the only lever you actually hold — the only place you get a signature — is the contract.

Why this is a Pillar 5 problem, in numbers

Start with the objection: electricity looks small. Epoch AI’s model of a 1 GW AI data center puts energy at roughly 7% of annualized TCO — about $0.6B of $8.5B — at 8.34¢/kWh with a 1.14 PUE. Servers are 60%. On that math, power is a rounding error.

That math is a trap, for two reasons.

First, the denominator is GPU capex. Strip it out and electricity becomes the dominant operating line. No hyperscaler audits this figure out publicly, but the best independent modelling — David Leitch’s at ITK Research — puts electricity at roughly 30% of facility cash opex at 50–60% utilization, rising to 59–85% at high utilization and hyperscale depending on local power price. On his estimate for a Meta 24,576-H100 cluster, energy is ~9% of four-year TCO but ~32% of cash opex. A model, not a disclosure — but nobody has a better one.

Second, and this is the part that should keep a CFO up: the chip is a depreciating asset on a schedule you control. The electron is an escalating liability on a schedule someone else controls.

Look at what re-priced without anyone at your company signing anything. PJM capacity cleared at $28.92/MW-day for 2024/25. The next auction cleared at $269.92 — a 9.3x reset in a single cycle. Then $329.17, then $333.44, then $325.00 for 2028/29 in the auction held July 14, 2026 — 11.5x above 2024/25, and holding at a plateau. Here’s the detail most coverage skips: the last three of those auctions cleared at the administrative price cap, because the RTO failed PJM’s three-pivotal-supplier test and market power mitigation kicked in. PJM’s own 2026/27 report states that without the cap, the system marginal price would have reached $388.57. That isn’t a market finding a price. That’s a regulatory ceiling holding one down.

PJM’s Independent Market Monitor, Joseph Bowring, attributes $29.4B of $63.6B across those four auctions — 46% — to data centers. That’s a modelled counterfactual, not a line item on anyone’s bill, and worth stating as such. Meanwhile US industrial retail electricity hit 8.71¢/kWh in May 2026, up 5.1% year over year, and EIA modelling found that under a high-data-center-demand scenario, 2027 ERCOT North wholesale prices would run $37/MWh — about 79% — above its February base case.

Now make it concrete. One GB200 NVL72 rack at 132 kW and a 1.15 PUE burns about 1,330 MWh a year — squarely inside Eurostat’s 500–2,000 MWh non-household band, so the published comparisons are apples to apples. In Eurostat’s H2 2025 figures, that band runs €74.8/MWh in Finland and €255.2/MWh in Ireland, the EU’s highest. Same rack, same silicon, same tokens: roughly $116K/year in Finland, $116K at the US industrial average, $397K in Ireland.

A 3.4x spread, decided by a line on a map. Against roughly $780K/year of amortized rack capex — SemiAnalysis puts a GB200 NVL72 rack at ~$3.9M all-in, over five years, though most hyperscalers now depreciate over six — that swing is the difference between a viable unit economic and a dead one.

One caveat: those are published tariff bands, not what a campus pays under a negotiated PPA. A large buyer in Dublin does considerably better — and that gap between published and contracted price is the argument. It’s the case for the second rung of the ladder below.

Now connect it to the thesis I’ve been running all summer. If you baseline your unit economics on open-weight models to get off a frontier lab’s per-token meter, you haven’t escaped the meter. You’ve swapped a token meter for a kWh meter. That’s still the right trade — but only because of one property: a kWh meter can be contracted for seventeen years. A token meter re-prices when a vendor publishes a blog post.

Financial sovereignty isn’t about having no meter. It’s about choosing the meter you’re allowed to negotiate.

The ladder: what each rung of energy control costs and buys

Sovereignty is a scale, so here’s the scale, priced.

And the rung most people skip, which is the cheapest on the board:

Flexibility as an asset. Duke’s Nicholas Institute found the 22 largest US balancing authorities could absorb 76 GW of new flexible load if it can be curtailed for just 0.25% of maximum uptime — about 85 hours a year — and 126 GW at 1%. For 88% of that curtailment time, half the load keeps running. Google had 1 GW of demand response in long-term contracts with five utilities by March 2026. Emerald AI, with Oracle and Salt River Project, cut a Phoenix cluster’s draw 25% and held it three hours during a real grid event while keeping workloads inside SLA. (Vendor-reported pilot results, not audited.)

Being voluntarily curtailable is how you avoid being involuntarily curtailed. Under PJM’s August filing, it’s the difference between sitting ahead of the paid demand-response program in the cut order and sitting behind it.

Buy or build

For 95% of enterprises: buy — but buy the contract, not the kilowatt.

Rung 2 is where the actual sovereignty is. A fifteen-year PPA is not an energy purchase; it’s a financial instrument that fixes the most volatile line in your model — and it’s the only rung most companies can execute without becoming an energy company. Note what AWS and Talen did not publish: the price. The widely quoted ~$18B value and 2%/year escalator are analyst estimates, absent from Talen’s own filing. In this market the strike price is the trade secret — which tells you how much of the value sits in the contract.

But the reader pushed back on exactly this, and their challenge went straight at the fab analogy I opened with. A fab really is un-ownable — almost nobody can build one, so exempting it costs the framework nothing. On-site power, they argued, is different: available today at standard project cost, plenty of operators already running it. One is a universal constraint. The other is a decision, and frameworks grade decisions.

That’s the right test, and it deserves a real answer. Two-thirds right — and the missing third matters.

The base rate. The best number available comes from Bloom Energy’s own survey of ~100 operators, fielded April 2024 to April 2025: 1% of facilities were fully powered by on-site generation, 13% used it for any primary power. Bloom projects 27% and 38% by 2030. A fuel-cell vendor publishing a number that undercuts its own pitch is a number I trust. A fast-growing minority of the best-capitalised operators are doing this. It is not yet what most operators do.

The cost. A modern AI campus builds at roughly $17.6M/MW (Cushman & Wakefield), past $25M/MW with AI fit-out. Layer prime generation on at $1,000–1,200/kW for aeroderivative turbines, $1,200–1,800/kW for recip engines, or ~$2,900/kW implied by Bloom’s AEP deal, and on-site generation adds roughly 5–17% of total project capex — a real number, not a rounding error and not a doubling. And it sits on top of the standby diesel already priced into that $17.6M/MW. Backup gensets and a continuous-duty power plant are different equipment.

The availability. The cheapest, most proven option isn’t available. What you can get inside twelve months is fuel cells at a premium, recip engines at 12–18 months, or rented bridge power that is temporary by design.

Their second point is the sharper one: separate generating your own power, a purchase, from running a grid, a utility. They’re right that these differ, and my “you’ll never own the grid” framing papers over it. So let’s draw the line properly. In US law, self-generation stays a purchase as long as no sale occurs — you generate, you consume, no third party is served. No FERC jurisdiction, no state public-utility status. The EU codifies the same idea as the “active customer” right in Directive 2019/944.

Here’s what changed in 2025–2026. The line is no longer drawn at whether you own generation. It’s drawn at whether you’re connected.

Three regulators, eighteen months, one direction. Texas, in Docket 59220, sized the curtailment obligation to the load — 525.5 MW — not to the 265.5 MW generator behind it. FERC, in its PJM co-location proceeding (EL25-49, December 18, 2025), directed that the netting threshold be measured by load, not generator nameplate. NERC is drafting registration for the computational load itself. And Ireland requires your matching generation to bid into the wholesale market — which is to say Ireland has decided the distinction shouldn’t exist for data centres at all.

Owning generation buys resilience, price certainty and leverage. It does not buy you out of anyone’s authority. Only true islanding does — no synchronous path to the grid — and that runs a further 15–30% generation overbuild for N+1 and maintenance rotation, plus firm fuel transport, black start and frequency regulation you now perform yourself. I could not find one confirmed, permanently islanded data center above 10 MW operating today with published reliability results.

So build clears the bar in three cases, and only three:

1. Scale. Sustained 50–100 MW+. Below that, the 17–24% year-five premium on behind-the-meter gas buys you speed you can lease more cheaply.
2. Queue exclusion. You cannot get grid power at any price on your timeline — Dublin, Northern Virginia, Israel, now Texas.
3. Jurisdictional risk. Your host government or utility has shown willingness to use power as a lever, and the workload is genuinely strategic.

And one trap on the buy side. Dominion’s GS-5 class takes effect January 1, 2027: loads at or above 25 MW running at 75%+ load factor, minimum 85% of transmission and distribution costs, 60% of generation costs, and a 14-year minimum contract term. Duke Energy has proposed similar terms in North Carolina — 75% minimum take, 10- or 15-year terms, 50 MW threshold — though the NCUC hasn’t ruled; a July 2026 settlement punted the design to a separate proceeding with a tariff filing due end of September 2026.

A 14-year minimum term is not a tariff. It’s a decade-and-a-half take-or-pay obligation landing on your balance sheet — exactly what financial sovereignty is supposed to be about, and exactly where most sovereignty reviews aren’t looking.

Scoring it: the Energy Control Index

Five questions, scored 0–4, weighted. Max 44.

Bands

• 0–14 — Grid tenant. You don’t have an energy strategy. You have an energy bill.
• 15–25 — Contracted. The correct answer for most enterprises.
• 26–35 — Hedged. Energy is a managed risk, not an open one.
• 36–44 — Sovereign-adjacent. You own the electrons.

Dimension 2 is where the reader’s distinction earns its place in the framework. There are three states, not two:

• Grid tenant — someone else’s generation, someone else’s dispatch. Score 0–1.
• Self-generating, interconnected — your generation, still their dispatch. This is where essentially everyone who “builds” actually lands. Score 2.
• Islanded — your generation, your dispatch, and a 15–30% overbuild plus a compliance burden to pay for it. Score 4.

The gap between rungs 2 and 4 there is the entire finding of the last eighteen months. Most people who think they bought a 4 bought a 2.

And the part that keeps this framework honest: most enterprises should be targeting 20–26, not 44. Above 30 costs more than it saves unless you’re at 100 MW+ sustained, or you’re operating somewhere a government has already shown it will use your power as a lever. Score 44 in a stable jurisdiction at 5 MW and you didn’t buy sovereignty. You bought a power plant you didn’t need.

Where I landed

The reader was right twice: the grid connection is a control plane, and generating power is not the same thing as running a grid.

And they were right that power doesn’t belong in the same bucket as the fab. The fab stays ungraded, because the exposure is identical for everyone and no enterprise decision changes it. Power isn’t like that. It varies 3x by jurisdiction, it’s contractable on terms you negotiate, and you can choose how much of it to own. That makes it a decision, and decisions get scored.

Where I’d push back is on what follows. The distinction between purchase and utility is real, but regulators have moved it. It no longer sits where you’d expect — between renting power and owning generation. It sits at the interconnect. On one side, you own your generation and still answer to a grid operator. On the other, you’re islanded, and almost nobody is.

So for the overwhelming majority of enterprises the lever isn’t a substation. You can price your power, term it, diversify it across jurisdictions, and get paid for flexing it. Every one of those is a Pillar 5 move, executed with a signature.

The unpriced dependency is the dangerous one. And right now, for most enterprises running AI at scale, energy sits on the books at last year’s number, with this year’s clauses attached and next year’s tariff pending.

Score it. Then decide what you can live with.

— Amit

Sources

Grid as control plane

• Texas SB6 (2025), PURA §39.170 — enrolled bill text; PUCT implementing rule still pending — DLA Piper
• Gov. Abbott directs data center audit, Aug 3, 2026 · ERCOT audit process and Batch Zero delay — Utility Dive · ERCOT all-time peak demand records
• PJM Interim Resource Adequacy Service, FERC Docket ER26-3515, filed Aug 13, 2026 — PJM Inside Lines · FERC combined notice of filings
• DOE Emergency Order 202-26-23, May 18–20, 2026
• PUCT Docket 59220, late July 2026 — White & Case · Utility Dive
• CRU Large Energy User connection policy decision, Dec 12, 2025 · decision paper CRU2025236
• Singapore EDB — capacity expansion, May 30, 2024 · 2022 sustainability pilot
• Israel Electricity Authority 140-day connection freeze — DataCenterDynamics · Calcalist/CTech · Jerusalem Post
• Virginia SCC order, Case No. PUR-2025-00058, Nov 25, 2025 · GS-5 terms — Inside Climate News
• Duke Energy NC large load tariff proposal — Canary Media · July 2026 settlement defers tariff design — EDF Climate 411
• Commission Delegated Regulation (EU) 2024/1364 — EU data centre reporting

Cost and unit economics

• Epoch AI — 1 GW data center TCO breakdown
• David Leitch, ITK Research — data centre operating cost structures (independent model, not an audited disclosure)
• PJM Base Residual Auction results — 2024/25 · 2025/26 · 2026/27 · 2027/28 · 2028/29
• IMM attribution of capacity costs to data centers — Utility Dive (Bowring’s public remarks; modelled counterfactual)
• EIA Electricity Monthly Update — industrial retail price, May 2026
• EIA Today in Energy — Texas data center demand and 2027 ERCOT prices, Mar 12, 2026
• Eurostat — non-household electricity prices, H2 2025
• HPE QuickSpecs — NVIDIA GB200 NVL72 rack power · SemiAnalysis — GB200 NVL72 rack cost
• CBRE — North America data center pricing and vacancy, YE2025
• LevelTen North America PPA Price Index, Q2 2026
• Talen Energy 8-K — AWS Susquehanna PPA, June 11, 2025 (price and total value not disclosed)
• Rabobank, “The Marathon” — behind-the-meter power economics, Mar 2026

Build-your-own: base rate, cost and availability

• Bloom Energy 2026 Data Center Power Report — 1% fully on-site powered, 13% partial (survey Apr 2024–Apr 2025) · full report
• Cushman & Wakefield — data center development cost per MW · JLL 2026 outlook — AI fit-out beyond $25M/MW
• GE Vernova Q2 2026 — 116 GW turbine backlog, conversations to 2032 · Caterpillar engine lead times, per Raymond James, Jan 2026
• FERC Docket EL25-49 — PJM co-location rules, order Dec 18, 2025
• NERC proposed “Computational Load Entity” registration
• EU Directive 2019/944, Article 15 — active customers and self-generation

Solutions and supply chain

• Bloom Energy–Oracle: up to 2.8 GW, 55-day first deployment
• Kairos Hermes 2 groundbreaking, Apr 17, 2026 — ANS · RoPower conditional FID, Feb 2026 — POWER · NuScale/UAMPS termination at $89/MWh — Utility Dive

Flexibility

• Duke Nicholas Institute, “Rethinking Load Growth” — curtailment-enabled headroom
• Google — 1 GW of demand response across five utilities, Mar 19, 2026
• Emerald AI / NVIDIA — Phoenix and London pilot results (vendor-reported)
• Uptime Institute 2026 Global Data Center Survey — rack density