TSMC's July record: stop pricing in cheaper tokens
TSMC July revenue hit NT$467.58B, up 44.7% year over year, with capex raised to $60-64B. Underwrite your automation at today's AI price, not tomorrow's.
Every automation business case we get handed has the same silent assumption buried in row 14: inference gets cheaper. TSMC's July revenue, reported August 10, is a good reason to delete that row. The company that makes the silicon under every AI product you use posted NT$467.58 billion — roughly $14.5 billion — up 44.7% year over year.
What actually happened
TSMC's monthly revenue report puts July at NT$467,580 million, a 5.6% increase over June's NT$442,680 million and a record month. CNBC reports the company now expects 2026 revenue to grow slightly above 40% in U.S. dollar terms, a target raised after last month's second-quarter earnings, and has lifted its capital expenditure plan to between $60 billion and $64 billion for the year.
The mix tells you where it comes from. High-performance computing — the segment where TSMC books AI chips — was 66% of second-quarter revenue. Chairman C.C. Wei's summary: "AI-related demand continues to be extremely robust."
Why it matters for your business
Read those two numbers together. Demand is running so far ahead of supply that the world's most important fab is spending up to $64 billion this year to catch up — and the catching up arrives in 2028, not this quarter. Capacity that gets financed now gets billed later, to whoever is renting the output. That is you, four markup layers down, on your API invoice and your cloud line item.
So the token price cuts you have watched all year are not the supply curve moving. They are vendors buying market share with capital, and capital changes its mind. If your automation only pencils out at a promotional rate, you do not have a business case — you have a bet on somebody else's funding round.
The fix is unglamorous. Underwrite the project at today's list price, not the discounted tier you are currently on, and not a projected cut. If it still returns, build it. Measure cost per completed task, not cost per token — a cheaper model that needs three retries and a human review is more expensive, and the token price hides that. Put a per-workflow ceiling in place so a price change shows up as an alert instead of an invoice. And keep the model call behind an interface, so switching providers is a config edit.
The businesses that get burned in 2027 will be the ones whose margins were quietly underwritten by a vendor's land-grab pricing.
Key takeaways
- TSMC July revenue: NT$467.58B (~$14.5B), up 44.7% year over year and 5.6% over June — a record month
- 2026 guidance raised to slightly above 40% USD revenue growth; capex raised to $60-64B
- High-performance computing was 66% of Q2 revenue — AI is the growth engine, and demand still outruns supply
- Capacity financed now gets billed to renters later; today's token discounts are market-share spending, not cheaper silicon
- Underwrite automation at list price, measure cost per completed task, and set per-workflow spend ceilings
If the ROI only works at a promotional token price, it does not work. We scope automation against what it actually costs to run — measured per completed task, with spend ceilings built in. Run the numbers on your workflow or tell us what you are trying to automate.
Sources: TSMC investor relations, CNBC.
- #ai-infrastructure
- #semiconductors
- #compute-costs
- #tsmc
- #automation-roi
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Moody's calls AI vendor dependence a systemic risk
Moody's warned banks that relying on a few AI model and cloud providers creates systemic dependency and pricing power. The same AI vendor risk applies to your stack.
Read itData center backlash is now an election issue
Gallup says 7 in 10 Americans oppose an AI data center near them, and candidates in both parties are running on moratoriums. Plan your cloud regions accordingly.
Read it