The Long Read · 24 Aug
An 8 June 2026 letter kept 1.9 kilograms of lean mass at week 24. The product clock is a Phase 3 NDA by 31 December 2028, not another tablet.
Thesis map
A GPU rack makes heat that still needs a 2.3 megawatt cabinet, a transformer, and grain-oriented electrical steel.
Hitachi Energy broke ground in South Boston, Virginia, on 29 June 2026 for 457 million dollars, completion still 2029. Vertiv put a 2.3 megawatt CoolChip CDU 2300 into EMEA on 26 May 2026. Cleveland-Cliffs is the only US producer of grain-oriented electrical steel at Butler Works, with about 25 percent more GOES in 2028. A rack that arrives this quarter still waits on those three objects.
The chain
GPU rack
A rack can ship on a GPU cycle. The sink that rejects its heat cannot. Flatten a token price into a cooling plant and you get a false headline.
Heat loop
Vertiv, 26 May 2026, London: CoolChip CDU 2300 in EMEA, 2.3 megawatts, liquid-to-liquid, in a row or a mechanical room. A cabinet moves heat. Someone still has to reject the loop.
Transformer
Hitachi Energy South Boston: more than 300,000 square feet, bays taller than 120 feet, a 375-ton overhead crane, rail out the back. Greg Callahan told ENR the project was a little behind. Completion is still 2029.
Electrical steel
Cleveland-Cliffs Butler Works in Lyndora, Pennsylvania, is the only US producer of grain-oriented electrical steel. TRAN-COR is the only high-permeability electrical steel made here. You cannot wind a core without this coil, or an import.
Long side
Owns the scarce object. Named filing, named plant, named queue.
long
29 June 2026 groundbreaking
Hitachi Energy
457 million · 2029
About 825 jobs on about 850 already there. The campus has been a plant since 1968, when it was Westinghouse. A press release does not ship a unit. A height variance is not a nameplate.
VRT
26 May 2026, London
Vertiv
2.3 megawatt CDU 2300
Microsoft already wrote liquid-to-chip, 2026 and beyond, closed loop. Goodyear 157th Avenue had no room above 500,000 gallons a day. Microsoft put 36 million toward a 90 million expansion and agreed later halls would be air-cooled.
CLF
induction reheat 2028
Cleveland-Cliffs
170 million at Butler
Lourenco Goncalves, on the second-quarter 2026 call: induction reheat done in 2028, about 25 percent more GOES at that plant. WPXI: 170 million dollars, 75 million from the Department of Energy, four induction furnaces in. About 25,000 extra tons in the original DOE write-up. That is not a second mill.
long
17 August
Hefei bushing shop
300 million dollars
On 17 August Hitachi Energy put 300 million dollars into Hefei, Anhui, for a transformer plant, an ultra-high-voltage bushing facility, and a digital tap-changer line. Bruno Melles called it part of a 9 billion dollar plan. It is not a winding that left Virginia.
Short side
Constrained by the chain, or a substitute that is not the product.
NVDA
A rack without a sink
A GPU rack without a transformer and a heat loop is a crate. The leftover of a megawatt slide is Hitachi 375-ton crane, not the chip.
short
Cheap token
Vertiv 2.3 megawatt cabinet is the leftover of a token. A list price that falls does not reject heat. Someone still has to reject the loop.
short
A 457 million announcement
You can announce a 457 million dollar plant. You cannot wind a core without Butler coil, or an import. A groundbreaking is not a nameplate in 2026.
How they interconnect
GPU rackmakesHeat loop
A GPU rack makes heat that a 2.3 megawatt liquid-to-liquid cabinet has to take before anyone talks about a token.
Heat loopstill needsTransformer
A cabinet moves heat. A transformer still has to stand. South Boston completion is still 2029.
TransformerneedsElectrical steel
You cannot wind a core without grain-oriented electrical steel, or an import. Butler is the only US mill.
Hitachi EnergybuildsTransformer
Hitachi Energy South Boston is the 2029 bay with the 375-ton crane. The Hefei note funds bushings. It does not ship that bay.
VertivtakesHeat loop
Vertiv sells the 2.3 megawatt CDU. Microsoft already wrote the closed loop. Goodyear still had a 500,000 gallon-a-day ceiling.
Cleveland-CliffsrollsElectrical steel
Cleveland-Cliffs rolls TRAN-COR at Butler. About 25 percent more GOES in 2028 is a better furnace on the only mill, not a second mill.
Hefei bushing shopdoes not shipTransformer
A 300 million dollar Hefei bushing shop on 17 August is a porcelain insulator. It is not a winding that left Virginia.
TransformergatesA rack without a sink
A transformer that has not shipped gates the rack. Hitachi cannot ship as fast as GPU racks arrive.
Vertivis notCheap token
A 2.3 megawatt brochure is not a rejected loop. The cabinet is what you buy first. The basin still has to take the water.
Cleveland-CliffsoutrunsA 457 million announcement
A 457 million dollar announcement does not reprint Butler coil. The mill is the leftover of a 2029 bay.
If you believe the thesis
Heat rejection and electrical steel stay slower than GPU racks through 2028. A 375-ton crane in a 2029 bay, a 2.3 megawatt cabinet, and one GOES mill cannot be parallelized. Hitachi putting 300 million dollars into Hefei bushings on 17 August funds a shop. It does not ship a winding from Virginia.
If the thesis is incorrect
Transformer lead times shorten by a year across two quarters while rack shipments still rise. Then the rack was the constraint. Cleveland-Cliffs disclosing that the Butler induction line is complete and GOES shipments from that mill are up versus the 2026 baseline, by 31 December 2028, would spend the steel leg.
Hostess is a bag on a shelf. Foundayo is a 1 April tablet. EMBRAZE is an 8 June paper about a kilogram that was not fat.
A third of the weight off tirzepatide alone in that trial was lean mass.
The dates are already on paper. 1 April on a Lilly approval. 8 June on a Nature Medicine letter. 18 June 2025 on the same 1.9 kilogram sentence, printed by the sponsor eleven months before the journal. A date is not a label.
The letter the catalog actually produced is a Phase 2 composition result, not a product. Pratley, Denham, Trivedi and colleagues, Nature Medicine, 8 June 2026: EMBRAZE, 102 adults, tirzepatide plus apitegromab versus tirzepatide plus placebo. Apitegromab is a fully human monoclonal that selectively inhibits myostatin activation. Dose: 10 milligrams per kilogram every four weeks, intravenous. At week 24, least-squares mean 1.9 kilograms less lean-mass loss than placebo (80 percent CI 1.2 to 2.7; p = 0.001) despite similar total weight loss. Relative retention of lean mass: 54.9 percent. Adverse events 76 percent versus 71 percent. One serious event in each arm of 51. ClinicalTrials.gov NCT06445075. Scholar Rock funded it. The company had already printed the same 1.9 kilogram / 54.9 percent sentence on 18 June 2025. The journal made it a paper eleven months later.
The antibody did not make people lose more weight. It changed the composition of the same weight. That is the leftover. You can clear a calorie and still leave a sarcopenic year, a fall, a joint, a nurse. Bimagrumab BELIEVE, an activin-receptor antibody plus semaglutide, is the cousin in the same drawer. Enobosarm and the other myostatin and SARM notes are the rest of the drawer. None of them is approved for this use. Apitegromab is an SMA story that picked up an obesity sidecar. A Phase 2 of 102 people is a letter, not a cluster. It is also the first object that makes the leftover measurable.
Do not medicalise a ticker. Scholar Rock is a small listed name. Lilly and Novo are on the board because they sold the weight that created the kilogram. Stryker and Intuitive are the listed body-hardware cousins if you need a hip or a console. They are not EMBRAZE.
The food-demand page already named Hostess and Gallup's 11 percent. Units on a shelf are still the print. A 1.9 kilogram lean-mass delta is a different object. Foundayo shipped 1 April. That tablet is access. Figure's Forge print on 22 August was $174.00, down $20.93. That is a used share. The sink page already named Hefei, South Boston, Foundayo, and this 1.9 kilogram letter. This piece is the kilogram.
Tokens do not have lean mass. Watts do not. Orbits do not. Appetite did, and the industry is now writing antibodies against the waste product of its own drug. The tablet shipped. It is whoever owns the year the lighter body still has to live: muscle, bone, balance, care. A warehouse robot is the industrial version of the same sentence. One is a bid on replacing the hour. The other is a bid on keeping the hour from collapsing. Only one has a p-value.
Cross-asset does not care which kilogram was kept. AI abundance cares only if the middle's hosting cost becomes a care cost. It has not, on the invoices those pages already named.
A refutation would look like this: a Phase 3 lean-mass preservation program, with a named antimyostatin or activin antibody plus an incretin, that has an NDA or a pre-announced filing date, by 31 December 2028, confirms the leftover became a product. A 2028 of Phase 2 letters and no filing confirms the kilogram stayed a paper. EMBRAZE is a trial, not a recommendation. The listed peers on the board for the weight are the incretin names, not the SMA name, until a 10-Q puts apitegromab in a sentence that can be dated.
The honest uncertainty is which object gets a number first. If a named Phase 3 files and Hostess is still a mix print, the kilogram moved and the aisle did not. If Hostess units fall and no antibody files, the aisle moved and the letter stayed a letter. If Foundayo prints a named line and EMBRAZE is still n=102, access shipped and the leftover did not. Those three sentences can live in the same year. One point nine kilograms are a fact about one paper. Eleven percent of adults are a fact about one survey. Neither is a label.