Research on what AI compute actually earns

Two strands, one standard: numbers that can be checked against a public document. The Call is the interview itself, edited and anonymised. The data pages are the argument that comes out of it. The glossary defines the contracts and the company files read the filings. The full argument runs in Two Tides.

The Call 01 Who gets TSMC's CoWoS capacity in 2028 260,000 wafers a month, 60% of them Nvidia, and one yield the street has five to fifteen points too high. Read the data page → Flagship data page What a gigawatt of AI compute actually earns Six annualized prices for the same gigawatt, from build cost to the figure everyone quotes, each with the source that produced it. $110B quoted → $10–15B realized Company file Anthropic: $518 billion of commitments against $4.6 billion of revenue What the prospectus reviewed by Reuters implies about realized revenue per gigawatt, and why the covenants matter more than the headline run-rate. Read the file →

Latest

Newest first. Dated entries are posts in the letter, from either strand; the undated pages below are reference entries that get revised rather than replaced.

13 Nov 2026 The Call 06 Chinese accelerator delivery: 180,000 planned, 20,000 shipped Advanced packaging first, logic wafer second, memory third. The delivered number is set below the fab, and next year memory climbs to second as the stockpile runs down. 12 Nov 2026 Tide 010 DRAM 2028: the shortage that is really a routing decision One module carries two prices, 1,500 from the factory and 3,500 on the street. The 15 to 20 percent capacity increment due in 2027 is re-routed monthly, and where it lands decides whether 2028 is a shortage or a price war. 11 Nov 2026 The Call 05 Data centre water: 0.12 to 1.8 litres per kilowatt-hour, and the account nobody publishes Microsoft and Meta stopped evaporating cooling water. Google and Amazon kept it, and the water in the purchased power runs thirteen times the water at the fence. 10 Nov 2026 Tide 009 CXMT packaging: 40 / 30 / 20, and the plant that takes the stack in-house Three houses package its commodity DRAM. The HBM goes into a plant of its own: 1.3 percent of wafers at a reported 25 percent yield, and the only stack price. 9 Nov 2026 The Call 04 Liquid cooling: from 30 kW to 218 kW a rack, and the bill inside the loop The rack went up 7.3 times and the room did not. One rack's loop is $60,600 before the CDU; the 2027 plan works out to 43.6 GW. 6 Nov 2026 The Call 03 Agent CPU costs: 0.2 to 180 vCPU, and the two overcommits in between Providers sell 8 vCPU per seat and an active user averages 1 to 2. Ten times the agents buys only three to five times the CPU. 5 Nov 2026 Tide 008 Domestic AI packaging: 210,000 wafers a year, twice what the 2027 forecast needs At 25 to 40 good dies per wafer the lines release 5 to 7 million dies against a 2.8–3.0 million chip forecast. The constraint is the seam, not the queue. 4 Nov 2026 The Call 02 Ulanqab's data center power: 2 GW live and no grid bays left to allocate The cluster has 2 GW energized and roughly 6 GW in build. The missing input is not electricity; it is connection bays. 3 Nov 2026 Tide 007 Ascend's 2027 shipment forecast: three scenarios and the yield line that picks one The near-doubling to 2.8–3.0 million chips is a yield forecast, and one point of yield is worth 50,000 chips. 2 Nov 2026 The Call 01 TSMC's CoWoS capacity in 2028: 260,000 wafers and who gets them The customer split in wafers, the CoWoS-L ramp, and the TGV yield that decides whether CoPoS arrives in 2027 or 2029. 31 Oct 2026 Data page What a gigawatt of AI compute actually earns Six annualized prices for the same gigawatt: the quoted $110B is a rent estimate with its conditions removed, and the realized figure is a third of it.

Glossary

Six terms carry most of the disagreement in this market. Each entry is a page of its own, because each one is a search query and a recurring mistake.

Gigawatt-year: The unit that makes annualized rent comparable to build cost — and the unit most slides omit.
Take-or-pay: The clause that turns a demand forecast into a balance-sheet obligation.
Spot premium: Why on-demand capacity costs more than twice a three-year commitment, and what the gap is really buying.
Utilization: The denominator in every revenue-per-gigawatt claim, and the one nobody publishes.
Die yield: The only variable that scales a fab's whole output at once, and the one the 2027 Ascend forecast rests on.
Nominal capacity: What lines can release versus what customers receive, and the 300,000-chip gap between them.

Company files

Two files so far, both from documents rather than briefings: what a company discloses about capacity, contracts and revenue, set beside what is said about it.

Anthropic: Run-rate, commitments, and the lease that priced urgency rather than income.
SpaceX: What the Colossus capacity lease annualizes to, and why the top of the rent ladder is a schedule, not a yield.

The letter runs on two tracks: the call itself on most days, the analysis on Tuesdays and Thursdays, with the sources behind both.

Read it in Two Tides

New here? The Call 01 on TSMC's packaging allocation is a good place to start.