TSMC's record quarter came with a cost warning founders shouldn't ignore: the AI supply chain is getting more expensive at the base, and some of that pressure will work its way up.
The numbers from TSMC's second quarter looked almost absurdly good. Revenue hit $40.2 billion, up 33.7% year over year, according to the company's July 16 earnings materials. Gross margin reached 67.7%. Operating margin hit 60.3%. For the company making the chips behind Nvidia's AI accelerators, Apple's silicon, AMD processors and a long list of cloud hardware, this wasn't a soft quarter dressed up by accounting language.
Then TSMC started talking about what it costs to keep building.
Wendell Huang, TSMC's senior vice president and chief financial officer, told investors on the July 16 earnings call that overseas fabs are already diluting margins, even with demand running hot. He said TSMC still expects the ramp-up of overseas fabs to dilute gross margin by 2% to 3% in the early stages and widen to 3% to 4% in the later stages over the next several years. That is the number to watch. At $40.2 billion in quarterly revenue, a three-point drag is roughly $1.2 billion before you even get to the harder question of who pays for it.
TSMC's US commitment is no longer theoretical. On the same call, chief executive C.C. Wei announced an additional $100 billion investment in Arizona to build more logic wafer fabs for 2-nanometer and below technologies, plus advanced packaging capacity. The City of Phoenix said the announcement brings TSMC's total Arizona investment to $265 billion across 10 fabs, two advanced packaging facilities and an R&D center. It also said that, once the announced fabs are complete, around 30% of TSMC's 2-nanometer and more advanced capacity will be located in Arizona.
That's a huge shift. It is also expensive by design.
The margin hit is real #
TSMC's third-quarter guidance already shows the squeeze. The company expects revenue between $44.6 billion and $45.8 billion, about 37% growth year over year at the midpoint, but it guided gross margin down to 65% to 67%. Huang tied that near-term move mainly to the steep ramp of 2-nanometer technology, which he said would dilute second-half gross margin by about 3 to 4 percentage points. Overseas fabs are a separate drag layered on top over the next several years.
Capex tells the same story in cleaner language. TSMC raised its full-year 2026 capital budget to $60 billion to $64 billion, up from the $52 billion to $56 billion range it gave in January. C.C. Wei said the increase comes from stronger customer demand and inflation in tool prices. Look, that is not a footnote. When the world's most important foundry spends another $4 billion to $12 billion in the same year it raises guidance, the cost structure of advanced computing is changing in public.
The company has room to absorb some of it. Advanced technologies, defined by TSMC as 7-nanometer and below, made up 77% of wafer revenue in the second quarter. High-performance computing accounted for 66% of revenue. TSMC also raised its full-year 2026 revenue growth outlook to slightly above 40% in US dollar terms, with Wei saying AI-related demand remains extremely strong and that customers continue to send positive long-term signals.
But strong demand is not insulation. It is a reason to raise prices.
During the same call, Wei was asked about pricing strategy. His answer was careful but plain enough. TSMC wants higher margins, but it doesn't want to squeeze customers out of the market, and it needs enough profit to fund long-term expansion. Barron's reported this week that TSMC plans to raise chip prices by up to 10% starting in 2027, citing Nikkei Asia. That lines up with the pressure visible in TSMC's own numbers, even if the company doesn't spell out every customer negotiation on an earnings call.
Foundry costs become AI costs #
TSMC doesn't sell finished GPUs to you. It sells manufacturing capacity to designers such as Nvidia, AMD, Apple, Qualcomm and Broadcom. Those companies sell chips, systems or devices into markets where AI demand has given the strongest players considerable pricing power. Costs don't travel in a straight line from a wafer contract to your startup's inference bill, but they do travel.
If you're building an AI-native product, you should care about this before it shows up in your cloud invoice. Compute is not a vague infrastructure line anymore. For many AI applications, it is the cost of goods sold. Training runs, fine-tuning, retrieval, inference at scale, customer support agents, internal copilots, you name it. The foundry economics underneath those services are not separate from your gross margin just because they sit several layers below your Stripe dashboard. There is a tempting answer here: assume Nvidia, hyperscalers and model providers will eat the pressure because competition will force them to. Don't bother building a plan around that. Nvidia has pricing power. The cloud platforms have pricing power. TSMC has pricing power because no rival is matching its leading-edge volume at 2-nanometer. When each layer has a reason to protect its margin, the final buyer rarely gets a gift.
This does not mean AI startups should panic. It means your unit economics need to be honest. If your product only works when token prices keep falling forever, TSMC's Arizona buildout is a useful warning. Industrial policy may be good for supply-chain resilience, and the US may well need more advanced chip capacity onshore. But resilience has a bill attached to it. This one is showing up first in TSMC's margins.
Also read: How to Build an AI Agent for Your Business Without Writing Code • Sony and Mitsubishi Electric are putting AI inside the sensor itself to rewire factory automation • Wistron opens its first US factory in Fort Worth and starts stamping out Nvidia's most advanced AI chips