TSMC's September Sales Show AI Chip Demand Still Outrunning Supply

Taiwan Semiconductor Manufacturing Co., the contract chipmaker that fabricates the processors behind most of the world's artificial intelligence systems, told investors on Thursday that September revenue rose 54.6 percent from a year earlier, extending an unbroken run of outsized growth that shows little sign of cooling. The company's monthly sales disclosure, filed as a report with the U.S. Securities and Exchange Commission, put consolidated revenue at roughly NT$511.9 billion, or about $16 billion at the company's planning exchange rate. The figure is essentially flat against August, which was itself a record month, meaning TSMC closed the third quarter running at a pace its own guidance had not fully anticipated.

The release matters beyond one company's bookkeeping. TSMC manufactures the advanced logic chips designed by Nvidia, Advanced Micro Devices and Broadcom that underpin the current data-center building boom, and its monthly sales are widely read as the clearest real-time gauge of how much of that boom is actually being paid for in silicon, rather than promised in press releases. Unlike most large companies, which disclose revenue only once a quarter, TSMC reports consolidated sales every month in a steady stream of filings lodged with American regulators, giving investors a running read on demand well before the company walks through the numbers on its next quarterly earnings call.

September's print is notable less for the headline growth rate itself, which has accelerated through the year from 33 percent growth in the second quarter to above 50 percent in August and September, than for what it says about momentum. Revenue did not fall off after August's record; it held almost exactly at that level, a pattern analysts have come to read as evidence that factories are running close to full capacity rather than demand easing.

The numbers

September's NT$511.86 billion in revenue was down 0.6 percent from August's record NT$514.81 billion but up 54.6 percent from NT$330.98 billion in September 2025, according to the filing. Revenue for the first nine months of 2026 reached NT$3.90 trillion, up 41.1 percent from the same period last year. Added together, July, August and September revenue comes to roughly NT$1.49 trillion, which converts to about $46.7 billion at TSMC's assumed exchange rate of NT$32 to the dollar — above the top of the $44.6 billion to $45.8 billion third-quarter guidance the company gave in July. Taiwan's state-run Central News Agency reported after the August figures landed that the combined July-August total already had analysts saying TSMC was on track to meet or beat its own forecast; September's number confirms it did.

An expansion still behind demand

The September print extends a pattern that has held for more than a year: TSMC raising its own expectations and then clearing them. The company has told investors it expects 2026 revenue to grow by slightly more than 40 percent in dollar terms, and it has lifted planned capital spending for the year to a range of $60 billion to $64 billion to add capacity, chiefly for the advanced nodes used in AI accelerators. A preview of the company's upcoming earnings call notes that high-performance computing chips — the category that includes AI accelerators and data-center processors — made up 66 percent of TSMC's wafer revenue in the second quarter, up from a much smaller share only a few years ago.

Chief executive C.C. Wei has been blunt about the mismatch between what customers want and what TSMC can build. Speaking to shareholders at the company's annual meeting in Hsinchu in June, Wei acknowledged that even with new fabs coming online in the United States, the company would not catch up with orders anytime soon.

"It will be a long time before we can meet customer demand," Wei told shareholders in June.

That shortfall is industry-wide, not specific to one company. Trade publication Bits&Chips has reported, citing the WSTS semiconductor forecasting group, that the global chip market is expected to reach $1.5 trillion in 2026, with hyperscale cloud operators alone projected to spend around $725 billion on AI infrastructure this year, a large share of it flowing to chipmakers and their suppliers.

Who feels the squeeze

The immediate winners sit upstream and downstream of TSMC's fabs. Nvidia, AMD and Broadcom depend on TSMC's leading-edge nodes to build the GPUs and custom accelerators that cloud companies are racing to install, and any slippage in TSMC's output effectively caps how many AI chips reach the market regardless of demand. Cloud operators such as Microsoft, Amazon and Google, which have committed hundreds of billions of dollars to AI data centers, remain dependent on allocation decisions made in Hsinchu and at TSMC's newer sites in Arizona and Japan. Taiwan's broader economy also rides on these numbers: semiconductors are the island's dominant export, and TSMC's monthly filings are tracked by its government as a proxy for industrial output.

Investors have their own stake. TSMC shares are among the most closely watched proxies for the durability of the AI investment cycle, and the September figures will feed directly into expectations for the handful of companies — Nvidia and AMD especially — that report their own results in the weeks following TSMC's quarterly call.

What comes next

TSMC will not comment further until its formal third-quarter earnings call, scheduled for October 15 at 2 a.m. Eastern time, the start of a quiet period the company began observing on October 5. Analysts' consensus, per the Crypto Briefing preview, points to revenue near $45.8 billion, the top of TSMC's guidance range, with earnings per share around $4.39. Investors will be watching whether high-performance computing's share of revenue holds near two-thirds of the total, whether gross margin lands inside the company's 65 to 67 percent target band, and whether management raises capital-spending plans yet again — a move that would signal TSMC still sees the current AI buildout as a multiyear phenomenon rather than a cycle due to peak. Nothing in Thursday's monthly filing resolves the broader question hanging over the sector: whether spending at this scale is sustainable once cloud companies start billing customers for the AI services all this hardware is meant to run. For now, the order books, and the monthly numbers that trace them, keep climbing.