TSMC posted record Q1 revenue of $35.6 billion with gross margins approaching sixty-five percent. Every AI chip flows through their fabs. The foundry's margin is the most honest demand meter in the stack.
Every earnings season, the AI infrastructure debate restarts from the same question: is the spending justified? The hyperscalers announce capex. Analysts model returns. The market tries to decide whether six hundred and fifty billion dollars a year is a railroad or a telegraph.
TSMC just answered the question from a place nobody else can. On April 10, the company reported first-quarter revenue of 1.13 trillion New Taiwan dollars — roughly thirty-five point six billion in U.S. dollars — up thirty-five percent from a year earlier. March alone hit 415 billion New Taiwan dollars, a forty-five percent year-over-year surge. The stock rose more than two percent in pre-market trading. The full earnings conference is April 16, but the revenue number already says what matters.
Here is why TSMC's margin is the most honest signal in the AI economy: every chip tells a partial story. NVIDIA reports demand for its GPUs, but NVIDIA designs chips — it does not manufacture them. Broadcom reports custom silicon revenue, but Broadcom is a designer too. The hyperscalers report capex, but capex is an input, not a product. Only TSMC touches every wafer. When NVIDIA, AMD, Broadcom, Apple, Qualcomm, Amazon, Google, and Microsoft all need leading-edge silicon, every order flows through the same fabs in Hsinchu. TSMC's revenue is the integral of the entire AI hardware economy.
The Margin as Demand Meter
Revenue growth tells you volume. Margin tells you pricing power. And pricing power tells you whether demand is real or subsidized.
TSMC guided gross margins of sixty-three to sixty-five percent for the quarter. Some analysts forecast higher — one estimate put it at nearly sixty-seven percent, above the company's own guidance range. These are extraordinary numbers for a manufacturing business. For context, TSMC's gross margin in 2019 was around forty-six percent. The twenty-percentage-point expansion over seven years is not the result of cost-cutting. It is the result of customers willing to pay more per wafer because no alternative exists.
The company holds roughly seventy percent of the global foundry market. Samsung, in second place, holds seven percent. This is not a competitive market. It is a monopoly with a customer list that reads like the index of the AI economy. When TSMC raises wafer prices — as it has, three to five percent on sub-five-nanometer processes this year — every customer pays. The alternative is not manufacturing chips.
The pricing reveals demand quality. TSMC's next-generation two-nanometer wafers are expected to cost roughly thirty thousand dollars each. The angstrom-era A16 process: forty-five thousand dollars per wafer. Customers are pre-booking capacity at these prices because the alternative — waiting — is more expensive than overpaying. That is what real demand looks like. Not projected demand. Not survey demand. Demand denominated in purchase orders for silicon that does not exist yet at prices that would have seemed absurd five years ago.
The Capex Equation
TSMC plans to spend between fifty-two and fifty-six billion dollars in capital expenditure this year, up from forty-one billion in 2025. Seventy to eighty percent is allocated to advanced processes — the two-nanometer and angstrom-era nodes that AI workloads require. This is the foundry's own bet on the persistence of demand. TSMC does not build speculative capacity. It builds to order. The capex figure is a forward-looking indicator because it reflects customer commitments, not management optimism.
The AI accelerator segment — chips designed specifically for training and inference — is now projected to grow at a fifty-four to fifty-six percent compound annual rate through 2029, up from an earlier estimate of roughly forty-five percent. In 2025, AI accelerators represented seventeen to nineteen percent of TSMC's wafer revenue. By 2029, if the revised projections hold, AI could account for more than half of the company's advanced-node revenue.
This is the infrastructure bet resolving in real time. Not in analyst models or hyperscaler press releases. In the gross margin of the company that physically manufactures the chips. Margins expand when demand exceeds supply and customers lack alternatives. Both conditions hold simultaneously at TSMC, and neither shows signs of reversing.
What the Margin Doesn't Say
The quarter that proves demand is real does not prove demand is permanent. TSMC's revenue is a coincident indicator — it confirms what is happening now, not what happens when the AI capex cycle meets its first real downturn. The foundry's fifty-six billion dollar capex plan assumes customers will still want leading-edge wafers in 2028. That assumption is built on commitments, not certainties.
There is also the geographic question. TSMC's manufacturing concentration in Taiwan remains the single largest structural risk in the global semiconductor supply chain. The company's Arizona expansion — now projected at up to a hundred and sixty-five billion dollars for multiple fabs — is an attempt to diversify, but the leading-edge capacity is still overwhelmingly in Hsinchu. A Taiwan Strait crisis would not disrupt a company. It would disrupt the substrate of the AI economy itself.
But the margin does say something that no other number can. It says that the customers building the AI economy — every major cloud provider, every major chip designer, every company betting its future on artificial intelligence — are paying record prices for manufacturing capacity and booking more. The question of whether the AI infrastructure buildout is a railroad or a telegraph will take years to fully resolve. The foundry's margin, today, says the trains are full.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)