The most useful spreadsheet I have read this year does not forecast a single token of demand. It counts boxes. Take the semiconductor team's shipment schedule — how many B100s, Rubins, TPUs, and Trainiums leave the loading dock each year — divide by eight to get servers, and multiply by the nameplate draw of the actual Dell chassis they ship in: 8,419 watts for an 8×H100 box, 11,787 for Blackwell, 14,733 for Rubin, 23,021 for Feynman. Add 10% for storage, gross up for cooling. No narrative anywhere in the chain — just physics times a purchase order. The output: the global AI fleet needs ~109 cumulative gigawatts by 2028, consuming more than 800 TWh a year, with roughly 60% of it trying to plug in inside the United States.
Here is the issue's framework, and it fits on a napkin: the Watts Ledger. Demand is a purchase order; supply is a construction schedule; the difference is a price. Chips arrive in 12 months. Transmission arrives in 5 to 7 years. When one side of a ledger moves six times faster than the other, the gap doesn't close — it gets priced.
Now the American ledger. Demand 2026–28: 67.6 GW. Subtract what's actually being built — 19.8 GW under construction, haircut to 14.85 at a 75% completion rate — and 15 GW of available grid interconnection. The hole: 37.7 GW before heroics. Stretch the window to 2030 and the hole grows to 121.7 GW. The analyst then prices the heroics, each with a probability attached — and this is where the issue gets interesting, because the largest single rescue on the list is not a turbine. It is the bitcoin-mining industry's real estate.
Why does a missing gigawatt command such a price? Because the thing waiting for power is not a warehouse — it is a depreciating bar of silicon. A Blackwell server burns about $10,000 of electricity a year and $40,000 of depreciation. Idle time costs four times more than energy. The workbook runs the trade explicitly: save 30% on power forever, but wait 24 extra months to energize, and you destroy three dollars of GPU depreciation for every dollar of power saved. A rational developer will therefore pay nearly double the price of electricity to get it years sooner — and the value of jumping a three-year queue computes to $13.33 per watt, against roughly $3.50 per watt for what a bitcoin mine is worth as a bitcoin mine.
Multiply the 2026–30 hole by the value of filling it fast and you get the number that should be on the cover of the report: 121.7 GW × $13.33/W ≈ $1.62 trillion of value waiting for whoever can deliver energized capacity early. That is the prize pool that has gas-turbine order books sold out, fuel-cell makers re-rating, nuclear owners hosting hyperscalers behind the fence, and — the subject of No. 005 — bitcoin miners trading like infrastructure REITs. America has quietly decided the marginal electron belongs to the machine that talks; the only argument left is who collects the toll.
| 2026–2028 | 2026–2030 | |
|---|---|---|
| Power needed | 67.6 GW | 151.5 GW |
| Under construction | −14.85 GW (19.8 GW × 75% completion) | −14.85 GW |
| Grid headroom | −15 GW | −15 GW |
| Raw shortfall | 37.7 GW | 121.7 GW |
| Rescues, prob.-weighted | 27–46 GW (mid 36.5) | not yet visible at scale |
| Net position | −10.7 GW low case; −1.2 mid | −121.7 GW × $13.33/W = $1.62T |
The shortfall is a forecast of a forecast: it inherits every assumption in the chip schedule. If NVIDIA's 2027–28 volumes slip even 20%, the gap halves. If chip designers pivot hard to efficiency — the report's own Risk #2 — watts per token fall faster than tokens grow. The 75% utilization assumption flatters demand; real fleets idle. The 15 GW grid-headroom figure is deliberately conservative, and ERCOT alone has surprised to the upside before. And the rescue list is not exhaustive: behind-the-meter gas, imports from Canada, and demand response don't appear at all.
If several of those break friendly, the "shortage" resolves into a glut of contracted capacity coming online into decelerating demand — at which point $15/W conversions and $13/W queue-jumping premiums mean-revert violently toward construction cost. Scarcity pricing is the most perishable pricing there is. That is precisely what a holder of miner equities at +400% implied upside needs to keep in the front pocket.
JLL counts 5.55 GW under construction and 12.25 GW planned. Demand needs ~16–29 GW/yr of increments. The pipeline must roughly double, every year.
Risk #1 in the workbook: data centers can't absorb the chips. The tell is upstream — optical and hardware suppliers (COHR, LITE, DELL, STX) flagging push-outs.
Gas turbines are rescue #1 (15–20 GW at 90%). If aeroderivative lead times stretch past three years, the low case becomes the base case.
Nobody in this story wants electricity. The hyperscaler wants to never be the one who ran out of compute while a rival trained the next model; the miner wants, after a decade of being despised for burning power on nothing, to be told his megawatts were the point all along; the utility wants to matter again. The desire under the data is the oldest one — to not be left behind — and it is mimetic all the way down: each buyer's appetite is set by watching the other buyers, not by watching their own customers.
Which is why the ledger will overshoot. Gaps priced by rivalry always do. The grid is the one counterparty in the trade that does not feel envy, and it will deliver its watts on the schedule of concrete and copper, indifferent to what the chips cost whoever is waiting.