The global artificial intelligence arms race is colliding with a formidable physical barrier: the production lines of Taiwan Semiconductor Manufacturing. As the world's dominant contract manufacturer of the most advanced chips, TSMC is confronting an almost frantic demand for production capacity from Silicon Valley tech giants. Its existing manufacturing capabilities are already unable to keep pace with this explosive growth, leading to a rapidly widening supply-demand gap that is beginning to reshape the global semiconductor supply chain landscape. According to The Information, this capacity crisis was laid bare in recent high-level interactions. During a joint media appearance in November last year, NVIDIA CEO Jensen Huang and TSMC CEO C.C. Wei directly confirmed that Huang's trip to Taiwan was to seek "more chips." Simultaneously, Broadcom, representing Google's quest for capacity, has repeatedly applied to TSMC in recent months to increase production for Google's custom TPU chips. However, informed sources revealed that TSMC explicitly stated it could not meet the full extent of this demand. This "capacity run" is having a direct impact on the market. In an October earnings call last year, C.C. Wei disclosed that current demand for TSMC's most advanced chips has reached three times its available production capacity. As TSMC prepares to release its latest financial results, management is expected to face further inquiries on how to alleviate this bottleneck. The strain is not limited to AI chips; the data center construction boom is also driving up demand for memory and connectivity chips, causing the capacity crunch to spread further. To cope with this predicament, some clients have been forced to seek alternatives, even considering changes to the long-relied-upon foundry model. Tesla entered into a $16.5 billion agreement with Samsung in July last year to produce its next-generation chips, with Elon Musk even hinting that Tesla might build its own chip factory. However, for companies like NVIDIA and Google, which are heavily dependent on TSMC's advanced processes, there are few other viable options in the short term. This capacity crisis is becoming the single biggest variable constraining the expansion speed of the AI industry. Demand Surge and Allocation Challenges TSMC is navigating an extremely difficult balancing act, striving to maintain stability for existing clients while responding to the unpredictable nature of the AI wave. As TSMC's largest customer, Apple's annual orders for iPhone and iPad chips typically fall within a predictable range. In contrast, the chip demand fueled by the AI arms competition is highly volatile and difficult to forecast. The current production line scheduling exacerbates this conflict. TSMC uses the same production lines to manufacture chips for both Apple's iPhones and iPads, as well as AMD's AI chips. Similarly, production resources are shared for Apple's server chips, NVIDIA's latest-generation Rubin chips, and Google's TPUs. Pressure from the demand side is multifaceted: OpenAI's planned super data center requires millions of chips; Google is aggressively purchasing as many NVIDIA GPUs as possible, while simultaneously pressuring TSMC through its design partner, Broadcom, to produce more of its custom TPUs. Yet, according to sources, TSMC adheres to a strict annual schedule for negotiating capacity and pricing with clients, typically not discussing orders beyond a one-to-two-year horizon. The company has made it clear that clients cannot "jump the queue" by paying a premium, nor can they arbitrarily cancel orders if business slows down. Expansion Plans That Offer No Short-Term Relief In response to the capacity shortage, TSMC has begun adjusting its global footprint, but these measures are unlikely to provide immediate relief. Reports indicate that TSMC has decided to shift the focus of its new factory in Japan from originally planned automotive chip production to the most advanced 2-nanometer process, with completion expected in 2027. Additionally, TSMC is accelerating construction of its second factory in Arizona, planning to move up the start of 3-nanometer chip production by one year to 2027. However, neither the Japanese nor the American expansion plans can address the current urgent need. To meet the production demands for NVIDIA's next-generation flagship Rubin chips and Google's latest TPU, Ironwood, TSMC's primary solution involves redesigning existing factory space, converting older chip production lines to 3-nanometer production. Investment Prudence Under the Shadow of Cyclicality Despite the booming AI demand, TSMC has not committed to building a new factory dedicated solely to it. Two senior TSMC employees revealed that this caution stems from the company's deep understanding of the semiconductor industry's cyclical nature. Constructing a state-of-the-art wafer fab costs tens of billions of dollars and takes several years, yet chip demand can shift much faster. TSMC previously increased investment during the pandemic due to surging demand for work-from-home and gaming equipment. However, as life returned to normal, demand quickly receded, leading to an 8.7% year-on-year revenue decline in 2023, even as demand for AI chips was beginning to take off. Furthermore, the pure-play foundry model established by TSMC founder Morris Chang dictates its investment discipline. Unlike Intel and Samsung, TSMC does not design or sell its own branded chips, relying entirely on customer orders. If clients cancel orders after capacity is expanded, TSMC faces the risk of costly idle production lines. A "Secondary Blockage" in the Packaging Stage Beyond wafer fabrication, advanced packaging has emerged as another critical bottleneck. This complex process, which involves assembling and connecting multiple processors into a finished product, is crucial for high-end AI chips. Informed sources indicate that TSMC has reallocated capacity from some older chip production lines to advanced packaging components. However, due to the complexity and precision of the process, even chips produced at TSMC's Arizona factory for NVIDIA's Blackwell architecture still need to be shipped back to Taiwan for final packaging. NVIDIA experienced the sting of packaging capacity shortages in 2023 when TSMC could produce enough Hopper processors but lacked sufficient packaging capacity. To secure supply for its latest products, NVIDIA has reportedly locked in the majority of TSMC's advanced packaging capacity from early 2025 onwards. This directly led to other clients being turned away—when Broadcom, on behalf of Google, attempted to increase TPU packaging orders late last year, TSMC refused. Currently, both AMD and Broadcom are testing whether other suppliers can handle part of their chip packaging needs.
Comments