TSMC's Record July Is a Supply Signal, Not a Demand Signal

TSMC published its July revenue report on August 10: consolidated revenue of NT$467.58 billion (about US$14.5 billion) for the month, up 44.7% from a year earlier and up 5.6% from June, another monthly record. It arrived two weeks before the other anchor of the AI trade reports: Nvidia, now TSMC's largest customer, delivers its July-quarter results after the close today, August 26. Financial coverage read the TSMC print as proof that AI demand is accelerating, which looks like an all-clear for the whole trade.

That reading has the driver wrong. TSMC's monthly revenue is not a demand survey; it is a shipment and pricing record. TSMC sells essentially every advanced wafer and packaging slot it can produce, and its growth rate is governed by how fast its capital spending turns into capacity, how fast product shifts to higher-priced nodes, and what it charges when it gets there. The scarce input is not orders. It is supply, and TSMC controls both the constraint and the price. That distinction determines what tonight's Nvidia number can and cannot tell you.

The July record is a capacity and pricing achievement

Set the July number against what TSMC committed to at its own second-quarter report on July 16. Revenue was $40.2 billion (NT$1,270.38 billion), up 36% year over year, with net income up 77.4% to a record and gross margin of 67.7% — a remarkable level for a foundry, earned by selling an advanced-node mix in which 7nm-and-smaller nodes contributed 77% of wafer revenue. On that result, management raised 2026 revenue growth guidance to above 40% in U.S. dollars, lifted capital spending to $60-64 billion from the $52-56 billion range it set in January, and guided the September quarter to $44.6-45.8 billion, roughly $2 billion above consensus. C.C. Wei described the supply-demand gap as "very big."

The capex number is the one to underline. It is the mechanism that decides whether the record recurs: revenue is downstream of capacity delivery. Read the monthly revenue figures as a supply ledger, measuring how fast a $60-64 billion budget converts into shippable wafers and packages, not how much demand would like to take.

The constraint has migrated to advanced packaging, the binding bottleneck. CoWoS — the packaging that pairs an AI chip with its memory — carries 52- to 78-week lead times and is sold out through this year, by industry reports. TSMC is scaling monthly CoWoS output from an estimated 75,000-80,000 wafers toward 120,000-130,000 by the end of 2026; the specific figures are analyst estimates, since TSMC does not disclose packaging output, but pricing confirms the direction. A 300mm wafer on TSMC's 2nm node reportedly costs about $30,000, against $18,000-20,000 on 3nm, and 2nm order books reportedly run into 2028.

The July surge sets Nvidia's print, and most of it is already written

Nvidia overtook Apple as TSMC's largest customer, a switch the 2025 full-year numbers confirmed: about 19% of TSMC revenue, or NT$726.9 billion (roughly $23.2 billion), up from 12% a year earlier, with estimates putting the share above 20% for 2026. What matters more than the percentage is what Nvidia already owns. Industry estimates put Nvidia at roughly 60% of TSMC's CoWoS output, with more than half of the 2026-27 packaging expansion allocated to it. That allocation is the mechanism linking the two income statements: TSMC's shipments from April through July become, to a meaningful degree, Nvidia's May-through-July product, and by the time Nvidia reports, the supply has already been converted into sellable units.

The prior quarter establishes the arithmetic. Nvidia reported revenue of $81.6 billion for the period ended April 26, up 85% year over year, at roughly a 75% gross margin. For the July quarter, consensus sits near $92 billion, only about 1% above Nvidia's own guidance.

So what has the July surge actually raised? Not the revenue print, which TSMC's capacity schedule largely pre-set and the market already knows. It has raised the bar on the two variables TSMC's shipments do not determine: Nvidia's guidance for the second half, and its gross margin on the Rubin transition.

Two markets are bidding for the same capacity

Reading the AI trade through TSMC exposes a split that GPU-centric coverage misses. TSMC's advanced capacity now serves two customer groups — merchant GPUs such as Nvidia's and custom ASICs such as Broadcom's and Google's TPUs — bidding for the same CoWoS slots. Nvidia's roughly 60% share is both a moat and a ceiling: it locks in Nvidia's own supply, and it caps Nvidia's growth at the pace TSMC expands. The transition runs on TSMC's schedule as well: Rubin enters volume production in the September quarter on TSMC's N3 process, and the follow-on Rubin Ultra has reportedly hit friction in its CoWoS-L packaging.

The pricing side cuts the same way. Nvidia monetizes its allocation at roughly a 75% gross margin while TSMC keeps raising the price of the bottleneck it controls — $30,000 2nm wafers now and, by report, more on the next node. Both collect while demand exceeds supply; the open question is who gives back margin once the new capacity lands.

The record extends only if the supply side delivers

The cautionary precedent is TSMC itself. At the very report that produced the record and the capex hike, TSMC's stock fell over 5% the next session, as investors focused on the $60-64 billion spend and the cost of U.S. expansion. Extraordinary news was already in the price. Nvidia's setup tonight is analogous: the print is signposted, so the swing is in guidance and margin, not revenue.

None of this denies that AI demand is real; TSMC's guidance and its customers' commitments confirm it. It re-locates the driver. The July surge does not measure how high demand can fly; it measures how fast TSMC can build, and at what price. That is the bar it raises for Nvidia, its biggest customer. The condition to watch is a two-sided margin question: if TSMC's CoWoS expansion and 2nm ramp land with pricing intact, the record extends and Nvidia's supply keeps growing; if the added capacity meets a pause in orders, the giveback appears first in Nvidia's guidance and margins, then in TSMC's. Tonight's report — guidance and gross margin, not the headline revenue — is the first data point on that question.