全球 AI 的咽喉：為何台積電的產能跟不上世界的野心？

The global artificial intelligence arms race is hitting a hard physical wall: TSMC's production lines.

As the world's leading foundry for the most advanced chips, TSMC is facing an almost frenzied demand for capacity from Silicon Valley tech giants, and the existing manufacturing capabilities can no longer keep up with this explosive growth, leading to a sharp widening of the supply-demand gap and beginning to reshape the global semiconductor supply chain landscape.

According to The Information, this capacity crisis has been laid bare in recent high-level interactions. During a public appearance last November, NVIDIA CEO Jensen Huang directly confirmed to the media alongside TSMC CEO C.C. Wei that his trip to Taiwan was specifically to seek "more chips." Meanwhile, Broadcom, representing Google, has repeatedly requested TSMC to increase capacity for Google's custom TPU chips in recent months, but sources familiar with the matter revealed that TSMC has clearly stated it cannot meet all of their demands.

This "capacity squeeze" is having a direct impact on the market. C.C. Wei revealed during TSMC's earnings call last October that the demand for TSMC's most advanced chips has reached three times its capacity. As TSMC is about to release its latest earnings report, management is expected to face more inquiries on how to alleviate this bottleneck. Not only AI chips, but the boom in data center construction has also driven up demand for storage and connectivity chips, further spreading the capacity tightness.

To cope with this predicament, some customers have been forced to seek alternative solutions, even considering changing the long-reliant foundry model. Tesla reached a $16.5 billion agreement with Samsung last July to produce the next generation of chips, and Musk has even hinted that Tesla may build its own chip factory. However, for NVIDIA and Google, which are heavily reliant on TSMC's advanced processes, there are virtually no other options in the short term, and this capacity crisis is becoming the biggest variable constraining the expansion speed of the AI industry.

Surge in Demand and Allocation Challenges

TSMC is at an extremely difficult balancing point, needing to maintain stability for existing customers while also dealing with the unpredictability of the AI wave. As TSMC's largest customer, Apple typically has a predictable order volume for iPhone and iPad chips each year. However, the chip demand fluctuations brought about by the AI arms race are severe and hard to predict.

The current production line arrangements exacerbate this contradiction. TSMC is producing chips for both iPhone and iPad on the same production line, as well as AMD's AI chips. Similarly, Apple's server chips share production resources with NVIDIA's latest Rubin chips and Google's TPUs.

The pressure on the demand side comes from multiple fronts: OpenAI's planned super data centers require millions of chips; Google is trying to buy as many NVIDIA GPUs as possible while pressuring TSMC through its design partner Broadcom to produce more custom TPUs.

However, sources familiar with the matter told The Information that TSMC insists on strict annual schedules for negotiating capacity and prices with customers, typically not discussing orders beyond one or two years, and clearly states that customers cannot "cut in line" by paying a premium, nor can they casually cancel orders when business slows down

Expansion Plans That Are Hard to Quench Immediate Thirst

In the face of capacity shortages, Taiwan Semiconductor Manufacturing Company (TSMC) has begun to adjust its global layout, but these measures are unlikely to take effect in the short term. According to reports, TSMC has decided to change the production focus of its new factory in Japan from automotive chips to the advanced 2-nanometer process, with the factory expected to be completed in 2027. Additionally, TSMC has accelerated the construction of its second factory in Arizona, planning to advance the production of 3-nanometer chips by one year to start in 2027.

However, neither the expansion plans in Japan nor the United States can address the urgent needs at hand. To meet the production demands for NVIDIA's next-generation flagship chip Rubin and Google's latest TPUIronwood, TSMC's current solution mainly involves redesigning existing factory space to convert outdated chip production lines into 3-nanometer production lines.

Prudent Investment Under Cyclical Shadows

Despite the booming demand for AI, TSMC has not committed to building a new factory specifically for this purpose. According to two senior TSMC employees who spoke to the media, this caution stems from the company's deep understanding of the cyclical nature of the semiconductor industry.

Building a cutting-edge wafer fab costs billions of dollars and takes years, but the speed of changes in chip demand is much faster. TSMC had increased investments during the pandemic due to a surge in demand for remote work and gaming, but as life returned to normal, demand quickly fell back, leading to an 8.7% year-on-year decline in revenue in 2023, even though AI chip demand had begun to take off at that time.

Moreover, the pure foundry model established by TSMC's founder Morris Chang also dictates its investment discipline. Unlike Intel and Samsung, TSMC does not design or sell its own chips and relies entirely on customer orders. If customers cancel orders after expanding production, TSMC will face the risk of expensive idle capacity.

"Secondary Bottleneck" in the Packaging Stage

In addition to wafer manufacturing, advanced packaging has become another key bottleneck. This is a complex process of assembling and connecting multiple processors into finished products, which is crucial for high-end AI chips.

According to insiders, TSMC has reallocated some of the capacity from outdated chip production lines to advanced packaging components. However, due to the complexity and precision of the process, even though TSMC's factory in Arizona has begun producing NVIDIA's Blackwell chips, these chips still need to be sent back to Taiwan for final packaging.

NVIDIA experienced the pain of packaging capacity shortages in 2023 when TSMC was able to produce enough Hopper processors but could not provide sufficient packaging capacity. To ensure the supply of its latest products, NVIDIA is said to have secured most of TSMC's advanced packaging capacity since early 2025. This has directly led to difficulties for other customers—when Broadcom attempted to increase packaging orders for TPU on behalf of Google at the end of last year, it was rejected by TSMC. Currently, both AMD and Broadcom are testing whether other suppliers can share their chip packaging needs