Thirty thousand truckloads
The $886 billion print was not a one-year spike. The same tracker now has 2027 at about $1.3 trillion. The power is still the constraint. The copper, the turbines, and the mines are why that constraint does not clear.
Jensen Huang put a 500-megawatt data center at 30,000 truckloads — and that was before anyone counted the power plant. He called it the largest infrastructure build-out in human history and told the electricians and the plumbers that this was their time. The dollar figure behind those trucks has already moved. Amazon, Microsoft, Google, and Meta now sit around $735–750 billion for 2026. TrendForce’s top nine, the print that used to be $830 billion and up 79 percent, is now more than $886 billion and up about 90 percent — and the same house has those nine at about $1.3 trillion in 2027. David Sacks said AI capex is already more than 2 percent of US GDP, and was 75 percent of GDP growth in a recent quarter.
The 1990s telecom cycle bought a lot of dark fiber that sat unlit. Copper in a hall does not work that way. Once the busway is in the building, it is used. Lewis Hart, who finances commodities at Brown Brothers Harriman, said the under-followed story in compute is copper — “as important” as the power everyone already talks about, and visible in the credit flows, not just the slide decks.
Everything in this market is power constrained. The reason that these folks may miss a number or a forecast have nothing to do with demand. It is entirely 100% due to the supply of the power necessary to generate the output token.
David Friedberg, All In, May 2026
The part that will not ship
Of a 16-gigawatt announced pipeline for 2026, about 5 gigawatts is actually under construction. Gas-turbine and step-up-transformer lead times have stretched to three to four years, against about eighteen months in the old world. Christian Bruch, who runs Siemens Energy, told Columbia he expects “some shortfalls in the industry of just digesting this growth” over the next two to three years. Cembalest’s version: five years for a combustion turbine, and people scrambling for Bloom fuel cells instead. Bloom’s July quarter put a number on that scramble — $1.065 billion of revenue, and 1.2 gigawatts of Oracle already contracted under a 2.8-gigawatt umbrella. GE Vernova’s gas book is larger, 116 gigawatts, but only about a fifth of those gigawatts are data-center, and the $5 billion first-half “data center” print is electrification kit, not turbines. The copper story does not turn on that mix. The bottleneck is still the metal and the machines that will not ship.
A single major AI data center is put at 40,000 to 50,000 tons of copper. Mine projects take fifteen to twenty years from discovery to production. Existing and planned mines cover only about 70 percent of projected 2035 demand. AI halls alone are headed toward 500,000 tons of copper a year by 2030 on one forecast. Escondida, the world’s largest copper mine, does about 1.2 million tonnes a year. A separate print called a 304,000-tonne deficit for 2025, widening in 2026. The United States depends on China for 100 percent of fifteen critical minerals. That stack is a different article. This one is the conductive metal in the truck.
Audrey Robertson, at Energy, named the American half of that metal as a smelter problem. “We only have three copper smelters in this country in the last decade. China’s built 42, I believe.” Processing, not ore. Color on a bottleneck this page already tracks — not a new infrastructure chain.
Three desks, one balance
The International Copper Study Group, J.P. Morgan, and Goldman Sachs do not agree on 2026. ICSG flipped from a 209,000-tonne surplus last October to a 150,000-tonne deficit — its first deficit call since 2009. J.P. Morgan sees a 330,000-tonne shortfall. Goldman still sees a 160,000-tonne surplus, on China demand weakness overwhelming the AI bid. The market, as of May, was trading closer to the deficit camp: LME around $13,100 to $13,400 a tonne after a record. That is a price. It is not a verdict. All three figures still stand.
Copper balances the desks do not share
Not a reconciliation — the three 2026 balances sit side by side. The 304kt / 500kt-by-2030 pair is a separate forecast.
Henry Hub, the US gas benchmark, was about $3 against $12 to $20 abroad when Cembalest did the math — cheap enough that behind-the-meter turbines and reciprocating engines pencil while the interconnect does not. That gas story has its own page. The materials point is simpler: the bottleneck has moved upstream of the GPU, into transformers, turbines, engines, and the mines that cannot open on a data-center calendar.
The orders land in industrials, not semis
Neil Dutta and Skanda Amarnath, on The Compound in June, ran a one-year correlation of every S&P 500 name against the semiconductor ETF SMH. Fifteen non-tech companies worth about $2 trillion now move with semis at correlations of 0.5 or higher. Twelve are industrials. Caterpillar sells backup generator sets into hyperscaler halls. Vertiv sells cooling and power management. Eaton sells switchgear. GE Vernova sells gas turbines. Generac sells backup power. Their order books, in Dutta’s framing, have become AI capex order books even though GICS still files them as industrials.
Skanda Amarnath’s caveat cuts both ways: correlations go to one in risk-on tapes too, so part of the 0.5 may be beta rather than fundamentals. And if hyperscaler capex rolls over, the basket de-rates with semis — watch credit spreads. The June read that put the Mag 7 up 7.3 percent and the other 493 up 12.6 percent is stale. As of September 11 the equal-weight S&P and the cap-weight sat at the same +11.9 percent; the Mag 7 were at +6.2 percent. S&P and MSCI are consulting on GICS through October 30. They may or may not change the structure. The topics on the table are semiconductors, HPCaaS, and foundation models — not electrical equipment into IT. None of Vertiv, Eaton, Caterpillar, Cummins, GE Vernova, Generac, Hubbell, or Comfort Systems has been reclassified out of industrials. The RenMac 0.5-or-higher SMH table has not been replicated. The basket is not a re-rate.