TSMC's AI Order Book Is Full—Its Margin Math Is Another Story

TSMC's latest quarterly report card carries a paradox that analysts are still working through: demand has rarely looked stronger, yet the profitability that demand is supposed to deliver is drifting in the wrong direction. The chipmaker posted a gross margin of 66.4% for the third quarter of 2026 on Thursday, down from 67.7% in the prior three-month period. Record appetite for AI silicon, in other words, did not translate into a richer margin.

Goldman Sachs remains constructive on the stock, pointing to AI-chip demand alongside higher selling prices and efficiency gains as the levers that should support the business. That framing matters because it goes beyond volume. By citing pricing and productivity, the bank is effectively acknowledging the cost side of the equation—and betting TSMC can manage it. Wedbush, for its part, singles out AI-chip demand as the growth engine and reads the quarter's revenue beat as a positive signal for the closing stretch of the year.

Both views underwrite expectations of continued momentum, but they answer different questions. Demand and the top-line outlook are one thing; holding the line on margins as costs climb is another.

Why the Top Line Isn't the Whole Story

The pressure is coming from several directions at once. Media reports point to costlier equipment, materials and energy feeding into production, while the growing share of the most advanced manufacturing nodes adds its own drag on profitability. Whether higher prices can offset those burdens comes down to negotiations with the large AI customers—a variable that sits outside any demand forecast.

Management's own guidance reflects the uncertainty. For the fourth quarter of 2026, TSMC projects a gross margin of 65–67%, a range wide enough to accommodate either an improvement or a further slide from the third quarter. For investors, pricing is therefore more than a growth talking point; it helps determine how much of each additional revenue dollar actually reaches the bottom line. Rising sales alone are not proof of improving profitability.

CEO C.C. Wei described AI-chip demand as extremely robust, yet the same management team struck a cautious tone on the year ahead, citing rising costs and macroeconomic uncertainty. That combination is what tempers the bullish analyst calls: strong demand supports the business but does not erase cost risk—which is precisely why Goldman's case leans on selling prices and efficiency rather than order flow.

A 3nm Nod From Nvidia—With Limits

Separately, a report from Reuters, citing the Seoul Economic Daily, suggests TSMC's advanced process technology could surface inside an Nvidia product. Samsung is said to be planning to use TSMC's 3nm process for the base die of Nvidia's NVHBM custom high-bandwidth memory. The scope is narrow and worth stating plainly: the decision concerns one specific Nvidia product, not Samsung's memory output as a whole.

According to the report, Nvidia asked Samsung, SK hynix and Micron to adopt the relevant manufacturing technology for NVHBM, and Samsung accepted after internal review. Unveiled in August, NVHBM integrates Nvidia's custom memory controller in the base die, while the memory chips themselves are to be built on sixth-generation DRAM. TSMC's process is earmarked for the base component—not for every element of the memory.

That distinction keeps the news in perspective. It does not amount to a wholesale shift of Samsung's HBM production or a broad migration to TSMC. Standard HBM and custom memory products for other customers remain within Samsung's own manufacturing model, and Samsung is reportedly examining the option of producing the NVHBM base dies in-house should demand outstrip TSMC's capacity—an alternative, not a decided switch. For TSMC, the reported technology choice is a signal of positioning rather than a firm production commitment.

The company's third-quarter disclosure offers a sense of scale for its leading-edge work: the 3nm process accounted for 33% of wafer revenue. That figure says nothing about available capacity or how future volumes might be allocated among customers, so it is no yardstick for the potential size of the Nvidia project.

Read together, the two threads point in the same direction. Nvidia's reported preference affirms the standing of TSMC's process technology in advanced memory designs, while Samsung's continuing in-house production and its contingency plan cap how far that endorsement extends. The takeaway is technological relevance—not a fresh revenue forecast. And on the core question of whether pricing power and efficiency can outrun rising costs, the margin outlook will tell investors more than any demand headline.