Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck. Not investment advice.
Explained
Wafer fab equipment is the set of machines that turn blank silicon into chips. Five toolmakers sell most of it. They are paid before a fab makes its first wafer. The EUV scanner is the tightest tool.
Wire
TeraWulf doubled contracted power at its Muskie campus in Kentucky to 1 GW and pulled the second 500 MW into 2029. The power is signed; the renter is not. The stuck step is the tenant.
Wire
The $27.3B bond behind Meta's Hyperion data center hit a record low near 91.6 cents on 7 Oct. It runs to 2049; Meta's lease is four years at first. The stuck step is Meta's renewal.
Explained
GPU utilization is the share of time a GPU does useful work. A GPU earns only when it is busy, so idle hours are the stuck step. The contract decides who pays for them.
US
Texas's new large-load rule took effect on 8 Oct: $50,000 of security per megawatt and a $100,000 study fee before a big data center can join a grid study. Wire companies get paid first. ERCOT's approval to switch on stays paused while it checks projects; its reports are due 10 Dec.
Analysis
Every AI chip sits on a small board, and the boards are short. The board is about 0.26% of a Rubin rack, but a month of waiting costs more than the board. Who gets paid while it lasts?
Wire
TSMC (8 Oct): September sales of NT$511.86B bring Q3 to about NT$1.49 trillion, above the top of its July guidance at its own exchange rate. The sales are in. The margin, hit by the 2-nanometer ramp, is known on 15 October.
Korea
Samsung guided ₩107.4 trillion of third-quarter operating profit on 8 Oct, its first quarter over ₩100 trillion. Memory makers get paid first; phone, PC and cloud buyers pay more. Profit is not cash yet, and new cleanroom space sets the pace.
Explained
A TPU is a chip Google designs for AI matrix math only; a GPU is a general-purpose parallel chip. Different designers and builders get paid, but both chips wait for the same memory.
Wire
Volta (5 Oct): a $5B loan at about 11%, with $3.7B for 36,000 Nvidia GPUs and $1.3B to back its Norway lease. The AI lab renter is signed. The loan is not. Commitments are due 14 Oct.
Wire
Lisa Su (6-7 Oct): AMD will make substantially more chips in 2027 and plans supply 3-5 years out. Each MI455X carries 432 GB of HBM4, about 31 TB per Helios rack, so Samsung and SK hynix set the pace.
Explained
A TPU, or tensor processing unit, is a chip Google designs for one job: the large matrix math that trains and runs AI models. Broadcom builds and ships it. Google earns only once the chip runs in a powered, finished hall.
