Summary
Highlights
The video starts by addressing Ben Thompson's article, 'TSMC Risk,' which claims TSMC's conservatism is costing American hyperscalers billions. The speaker challenges the notion that TSMC, a trillion-dollar company investing heavily, is holding back the AI boom. While agreeing that TSMC's 90% market share in AI chips is unhealthy, the speaker argues that 'shortages' are inherent to the complex semiconductor supply chain, not a failure of TSMC.
The 'boba game' (renamed from the 'beer game') is introduced as an analogy to explain system dynamics and supply chain volatility. Players act as retailers, wholesalers, distributors, and factories. Unannounced demand spikes create delays, shortages, or overproduction, illustrating the complexity of supply chains where each step has time delays.
The AI supply chain is incredibly complex, involving thousands of specialized suppliers for equipment and materials. These are not interchangeable, and each niche has few major players. Non-semiconductor factors like power, water, land, and labor also play a crucial role. The 'bullwhip effect' explains how demand signals amplify up the supply chain, leading to extreme volatility in semiconductor and equipment manufacturing, exacerbated by long lead times (e.g., 18 months to 2 years to build a fab).
The video recalls the COVID-era PC and remote working boom, where car manufacturers cancelled orders but then demanded chips back, leading to confusion and double-booking. TSMC responded by investing heavily, but by late 2022, the macro environment shifted, and demand for automotive, smartphone, and PC chips decreased, leading to low fab utilization and financial losses for TSMC's N7 fabs. This demonstrates why having slack advanced node capacity is financially risky.
Intel's struggles during the same period are highlighted: massive hiring, new fabs, and aggressive process node rollouts, only for the boom to end and hyperscalers to shift to GPUs. This resulted in layoffs and competitive setbacks for Intel. TSMC, however, pivoted to AI and survived, showing the effectiveness of its product diversification strategy.
The speaker challenges the criticism that TSMC didn't invest in AI earlier. Despite ChatGPT's release in November 2022, TSMC's CEO in April 2023 was still uncertain about AI's impact. Initial orders for co-ass capacity only emerged then, and revenue projections for AI accelerators were modest. Technical issues in 2024, such as coexistence capacity bottlenecks and yield problems, further delayed scaling. Concerns about whether AI chips would be profitable also made foundries hesitant to commit billions to new fabs.
The video refutes the idea that chips are the primary bottleneck, citing a TSMC customer's comment about power being addressed years ago, but now silicon is the bottleneck. The speaker interprets TSMC's stance as focusing on what they can control (chips). A deleted LinkedIn post from TSMC Arizona's CFO explicitly stated, 'AI's real bottleneck isn't chips, it's power.' The example of Elon Musk bringing in truck-mounted gas turbines due to power shortages, and the limited availability of new gas turbines until 2029, further supports this argument.
The video concludes by discussing the widening gap between the hardware and software worlds. Many in Silicon Valley, particularly those in software and AI, have little understanding of how their hardware is made. The speaker urges patience, emphasizing that people are working hard to produce chips, build racks, and start data centers, despite the perceived slowness of AI models.