The world's largest and perhaps most important chip foundry told Wall Street that its AI-related sales exceeded expectations. However, the company did not specify the extent of the improvement or whether its long-term outlook through 2029 had significantly changed.

Last night, while Europe and the US were still sound asleep, TSMC executives released their financial results for the third quarter ending in September. As expected, the results were strong, largely due to the booming GenAI industry. As AI clusters expand from data center halls to entire data center fields, model builders pursue superintelligence (creating models that outperform humans in all cognitive domains), and companies try to scale their applications with less impressive large language models, while inference workloads for these models will generally account for the majority of AI computing in the near future, the GenAI boom is also driving an increase in transistor count in XPU compute engines and switch and router ASICs.

Well, that's the idea of humans and AI models, all of which—we wouldn't dare say universally, because they're not yet human—are just making best guesses. The only way to truly predict the future is to live it, and we're certainly doing that, aren't we?

Regardless, by far the world's second-largest profit pool in the GenAI boom is TSMC, which is etching Nvidia's GPUs and various other compute and networking engines for AI clusters. Nvidia has the largest profit pool in the GenAI market, and according to a rough calculation we did on the back of an envelope, its AI training and inference revenue and profits are roughly six times that of TSMC. (The envelope is actually an Excel spreadsheet called Calc.xlsx on our desk, filled with all sorts of estimates...)

But don't feel sorry for TSMC—it's doing well and has enough financial resources to maintain its lead over its competitor Intel indefinitely, unless the US Department of Defense and Treasury Department acquire it, or a coalition of willing parties (big banks and equity firms) one day invests in domestic chip manufacturing out of patriotism and President Trump's call.

TSMC's revenue in the quarter ending in September was $33.1 billion, a year-on-year increase of 40.8% and a quarter-on-quarter increase of 10.1%. Net profit is growing faster than sales, which is something all companies like to see, as it means their business is expanding well. Net profit for the quarter reached US$15.1 billion, a 50.2% increase from the previous quarter and an 18% increase from the second quarter.

TSMC's business is growing in two ways. First, wafers are becoming richer (more features packed onto smaller transistors), allowing TSMC to charge higher fees for etching them.

Second, after overspending in the smartphone, PC, and traditional server markets during the COVID-19 pandemic caused wafer starts to plummet in the fall of 2022, TSMC finally started rolling out more silicon. (We've updated everything.) Measured in 12-inch wafer equivalents, TSMC surpassed 4 million wafers for the first time in its history—4.085 million good wafers, to be exact. Revenue per wafer was $8,102, a 59.2% increase from September 2022.

Many of the chips TSMC etches are based on older processes, which are more mature and therefore have higher yields and higher profits per transistor (but not necessarily per wafer; don't conflate the two). Chips made with 10-nanometer and larger transistor geometries generated $8.61 billion in revenue in the third quarter, an 18.1% year-over-year increase in revenue from this traditional manufacturing.

Sales of 7nm chips actually grew more slowly, reaching $4.63 billion (up 16%). Revenue from 5nm chips increased 62.8% to $12.25 billion, accounting for 37% of all chip revenue. Revenue from the newer 3nm process chips increased 61.9% to $7.61 billion, accounting for 23% of sales.

TSMC has only hinted at the size of its AI inference and training business revenue, beyond providing a firm figure for June 2023, stating that AI inference and training chips contribute 6% of revenue. (Thank you.) The company does have an "HPC" category, which refers to high-performance computing chips used in PCs, servers, and other data center equipment, unlike smartphone chips, which TSMC counts separately.

In the third quarter of 2025, sales in the high-performance computing (HPC) category (again, this is not the HPC we discuss on The Next Platform) reached $18.87 billion, a year-over-year increase of 57.4% and a quarter-over-quarter increase of 4.6%. According to our model, AI inference and training chips account for 53.9% of TSMC's HPC division sales in the third quarter of 2025, a 2.7-fold year-over-year increase to $10.16 billion.

TSMC CEO Wei Zhejia did not update the company's forecast released in June. As we previously reported and Wei Zhejia reminded Wall Street today, TSMC forecasts revenue driven by etched and packaged AI accelerators to grow at a compound annual growth rate of around 40% from 2024 to 2029 (inclusive).

"Demand for AI actually continues to be strong," Wei explained on the conference call. "That's stronger than we expected three months ago, right? As for the current situation, we've communicated this to our customers and their customers' customers. So, we previously reported a compound annual growth rate of around 40%. It's actually a little bit better than that. We'll probably update you early next year so we can have a clearer picture of the future."

Back in July, when the June quarter figures were released, we estimated the company's AI-related chip sales to be $9.1 billion in 2024, a 45% compound annual growth rate, implying sales of $85.5 billion by 2029, a 9.4x growth rate. Assuming TSMC's total revenue grows 25% in 2025 and a more modest 20% thereafter, that would represent a 2.6x increase over those six years, to $233.5 billion, with AI chips accounting for approximately 36.6% of total revenue.

A lot could happen between now and 2029, which is why Mr. Wei hasn't changed his forecast. However, we believe that if TSMC can achieve another 15% quarter-over-quarter growth in Q4 2025, its AI device sales could reach $36.63 billion, significantly higher than the $30.3 billion we forecast three months ago.

It would be great to get some firm data and new forecasts when TSMC reports its 2025 earnings next January. Our TSMC model suggests that AI chip revenue growth will slow over time, but Nvidia's co-founder and CEO believes Nvidia can maintain 50% data center growth between now and 2030. Nvidia and TSMC's forecasts are at odds, so we believe Nvidia will eventually expand into other areas of the data center, perhaps even building its own data centers. Perhaps not by acquiring Arm, but by acquiring Advanced Micro Devices and Vast Data.

In the meantime, TSMC will continue to stick with its etching process. This means focusing on improving chip manufacturing technology and praying that China doesn't invade and that this AI bubble lasts. If the AI bubble bursts, guess what? People will still need to buy chips, and TSMC will still profit from it.

TSMC: Continue to expand production in the United States

TSMC held an earnings conference on the 16th. Regarding the progress of its overseas fab construction, Wei Zhejia stated that with the support of the US government at all levels and customers, TSMC continues to accelerate its capacity expansion in Arizona and remains on schedule. However, due to strong customer demand for AI, TSMC is preparing to accelerate the deployment of 2nm or higher-level processes. It is also seeking a second large-scale land parcel in Arizona to support this expansion, reiterating its commitment to developing Arizona into an independent mega-fab cluster.

Mass production at the Kumamoto Fab 1 in Japan has already begun at the end of 2024, with excellent yield rates. Construction on the second fab has just begun, and the start-up schedule will depend on customer demand and market conditions.

Regarding PSMC in Europe, Wei Zhejia stated that with commitments from the European Union, the German federal, state, and municipal governments, construction has begun on the Dresden fab in Germany and is progressing smoothly as planned. However, the start-up schedule will depend on customer demand and market conditions.

Regarding Taiwan's fab expansion progress, Wei Zhejia noted that TSMC is currently preparing to build 2nm fabs in Hsinchu and Kaohsiung Science Parks. Over the next few years, TSMC will continue to invest in leading-edge process and advanced packaging facilities in Taiwan, continuing to invest in Taiwan while expanding its overseas fabs.

Regarding the progress of the 2nm and A16 processes, which the market is most concerned about, Wei Zhejia emphasized that 2nm will enter mass production as planned in the fourth quarter, with faster capacity growth expected in 2026, primarily driven by smartphones and HPC AI applications.

Through its continuous process enhancement strategy, TSMC will continue to launch N2P, an extension of the 2nm process. N2P offers further performance and power efficiency improvements over the 2nm process, and is planned for mass production in the second half of 2027.

As for the A16, it uses Backside Power Rail (BSPR) technology, which is more suitable for specific HPC products with complex signal paths and dense power supply networks. It will be mass-produced in the second half of 2027. It is expected that N2, N2P, A16 and its derivative technologies will become another large-scale and long-lived process node of TSMC.

Two-track development of advanced packaging in the United States

TSMC's new plant in Arizona, USA, has received a lot of orders, which has also driven strong demand for back-end packaging and testing. TSMC Chairman Wei Zhejia said yesterday (16) that he is working with a large-scale professional outsourced packaging and testing (OSAT) partner, and the other party has started construction in Arizona, which is earlier than the two advanced packaging plants planned by TSMC. He said that the goal of the cooperation between the two parties is to support the localization needs of key customers and accelerate the delivery schedule.

The market speculates that the "large OSAT partner" referred to by Wei Zhejia is the American company Amkor, but TSMC has not confirmed it. Industry analysts believe that TSMC will expand its advanced packaging production capacity in the United States with a dual-track strategy of self-construction and outsourcing, including Amkor's construction of a plant first, with mass production between 2027 and 2028; TSMC's two advanced packaging plants will take over later, forming a local AI supply chain with advanced processes to meet the strong long-term market demand.

The industry pointed out that Aikker's advanced packaging plant in Arizona broke ground on the 6th with an investment of US$7 billion. The first plant is expected to be completed in mid-2027 and mass production will begin in early 2028.

Source: Content compiled from nextplatform

View more at EASELINK