Is AI demand real or just market-driven hype? Rachel Chen, founder of Financial M Square, pointed out on 'After Work International' that the answer may lie within semiconductor company warehouses. She noted that while the Philadelphia Semiconductor Index dropped about 10% in July, it has still risen approximately 60% year-to-date, indicating an overall strong trend. The key data to watch is not short-term declines, but whether there's a turning point in semiconductor fundamentals and the manufacturing cycle. Therefore, the sustainability of Taiwan's stock market rally may directly depend on inventory turnover days at TSMC, NVIDIA, and Broadcom.

Chen observed that over the past 10 to 20 years, global manufacturing and the semiconductor industry have formed cycles roughly every 3 to 4 years, typically experiencing 1.5 to 2 years of growth followed by 1.5 to 2 years of decline. Similar cycles were seen around 2008, 2012, 2015, 2018, and 2021. This semiconductor upcycle began in 2023, and following historical patterns, it was expected to weaken by 2025.

However, the emergence of massive AI demand disrupted this traditional cycle. Manufacturing did slow slightly in the first half of 2025, but a second growth phase emerged in the second half. Chen noted that export, trade, and manufacturing data show AI has pulled the cycle upward again, suggesting the cycle hasn't ended yet and could remain relatively strong into the second half of 2026.

Chen stated that as peak season approaches and NVIDIA's Vera Rubin chips begin shipping, if companies can successfully convert inventory into products and then into revenue, the semiconductor manufacturing cycle could continue to rise.

Nonetheless, a yellow warning light has appeared for this 'super cycle'—inventory levels. Chen pointed out that at the end of last year, when manufacturing demand was very strong, corporate inventories remained low. Six months later, however, inventories have gradually built up. NVIDIA was the first to show this trend, with inventory turnover days rising from below 100 to over 100 days starting in Q3 and Q4 of 2025. Broadcom followed in Q1 2026 with rising inventory, and TSMC saw an increase in inventory turnover days in Q2.

Chen analyzed that 'inventory buildup' could mean two things: first, strong demand leads companies to proactively stockpile raw materials to avoid future shortages; second, demand suddenly weakens, leaving products unsold and piling up passively in warehouses. While both scenarios appear as 'inventory increase,' their implications for stock prices are completely opposite.

Thus, Chen believes recent volatility in the Philadelphia Semiconductor Index largely reflects market uncertainty about which type of inventory buildup is occurring. If warehouses are filled with unsold GPUs and chips, it may signal that the AI boom has peaked. But if the buildup consists of raw materials and work-in-progress, it suggests companies are preemptively securing supplies for the next demand wave.

Based on current data, Chen leans toward the first scenario—'proactive inventory buildup.' One reason is that order volumes still exceed inventory levels. Additionally, even as inventory turnover days increase, semiconductor companies' gross margins continue to rise, indicating pricing power. It's rare to see both rising inventory and rising margins if demand were collapsing.

More importantly, NVIDIA and Broadcom are primarily increasing raw materials and work-in-progress, not unsold finished goods. Chen noted this more likely indicates companies are seeing strong demand and are purchasing components and materials early.

However, the real stress test will come in Q3 and Q4 of 2026. As peak season arrives and NVIDIA's Vera Rubin chips start shipping, if companies can successfully convert inventory into products and then into revenue, the semiconductor cycle could continue upward. Conversely, if shipment speeds fall short of expectations, what is now interpreted as 'preemptive stockpiling' could quickly turn into 'unsold inventory.'

FACT BOX

  • Source: PR Times
  • Category: Survey
  • Organizations: TSMC / NVIDIA
  • Products / services: GPU