In late July 2026, news from the supply chain quickly spread in the semiconductor circle: NVIDIA quietly notified some data center clients of delivery adjustments, postponing H200 and B-series GPU orders originally scheduled for Q3 to Q4 or even early 2027. Although NVIDIA has not commented, multiple channel sources confirmed that this adjustment is not a capacity bottleneck but NVIDIA's first proactive control of shipment pace citing "high customer inventory levels."

This move quickly affected the market. On July 28, NVIDIA's stock fell 4.2% in a single day, dragging the Philadelphia Semiconductor Index down 1.8%. The market began to reassess the long-accumulated inventory risk in the AI chip field.

Inventory "barrier lake" is filling

Since 2023, global large cloud service providers (CSPs) have massively purchased AI training chips. GPUs like H100 and H200 were once in extreme shortage, with lead times up to 10 months. However, in the second half of 2025, with the mass production of rivals like AMD MI400 and Intel Gaudi 3, and the gradual rollout of cloud giants' self-developed chips (e.g., Google TPU v6, Amazon Trainium 3), demand growth for NVIDIA's high-end GPUs began to slow.

According to data tracked by the Chip Inventory Analysis column, as of the end of Q2 2026, the global average inventory weeks for major data center GPUs rose to 9.2 weeks, up 3.1 weeks year-on-year, the highest since the chip oversupply cycle in 2022. Among them, NVIDIA's channel inventory was particularly prominent, with some distributors' inventory-to-sales ratio exceeding 12 weeks.

"This is not a cliff-like drop, but definitely an alarm for inventory accumulation," a senior executive of a Singapore semiconductor distributor told reporters. "NVIDIA's order cut is more like a signal of proactive destocking. Other manufacturers may follow."

"Co-op overcapacity" emerges in the supply chain

The AI chip inventory pressure is passing upstream along the supply chain. In wafer foundry, TSMC's CoWoS advanced packaging capacity utilization has loosened. Previously, to meet emergency orders from NVIDIA and AMD, TSMC rapidly expanded CoWoS capacity from 30,000 wafers per month in 2024 to 80,000. However, in July, news emerged that TSMC has begun adjusting its CoWoS expansion pace, potentially delaying its 2027 expansion plan by half a year.

Meanwhile, the HBM (High Bandwidth Memory) market also felt the chill. Samsung Electronics and SK Hynix heavily invested in expanding HBM3E capacity in 2025, but after NVIDIA slowed procurement, the spot price of HBM3E fell about 5% in July, the first monthly decline in nearly two years. Although the long-term AI demand logic remains unchanged, short-term supply-demand imbalance cannot be ignored.

"From logic chips to memory, the entire AI computing chain is transitioning from restocking to destocking," said semiconductor analyst Wang Zhe of Saxo Bank. "The over-optimistic stockpiling in 2024-2025 will take about two quarters to digest."

Not a complete cooldown: Mature-node and automotive chips bottom out

Notably, this round of inventory adjustment is mainly concentrated in the high-end AI chip sector, while the consumer electronics and automotive chip markets, which had been deeply downgraded, are recovering. According to the Chip Supply-Demand Analysis monitor, in Q2 2026, capacity utilization for mature-node (28nm and above) wafer foundry rebounded to 85%, up 10 percentage points from the low last year. Inventory restocking of MCUs and power devices is evident.

"AI chips are structurally overheated, but the recovery in automotive and industrial chips has just begun," said a Morgan Stanley report in July, noting that global automotive semiconductor inventory has fallen from alarmingly high levels in 2025 to normal, with automakers starting to rebuild safety stock.

This brings new opportunities for the Southeast Asian semiconductor supply chain. Recent earnings reports from Malaysian packaging and testing houses Unisem show rapid growth in automotive chip packaging orders, with capacity utilization nearing full capacity. Singapore, as a Southeast Asian semiconductor distribution and logistics hub, is also seeing restocking demand for non-AI chips.

Market implications: Resetting investment logic

NVIDIA's order cut may mark a new phase in AI semiconductor investment logic. In the past two years, the market has given extremely high valuation premiums to AI chip companies, but once inventory corrections occur, stock price pullbacks could exceed expectations. Investors need to be wary of high volatility risks in AI chip-related ETFs while focusing on analog chips and automotive chips with relatively stable gross margins and benefiting from diversified demand recovery.

"AI is not a fake demand, but the supply chain cannot forever run ahead of demand," said independent analyst Chen Feng. "This inventory adjustment will lay the foundation for a healthier demand surge in the next wave."

Overall, the semiconductor market in July 2026 shows a clear "fire and ice" scenario: AI chip inventory rapidly builds up, leading vendors proactively control shipments; while mature-node and automotive chips are recovering from the bottom, shifting supply chain focus. For practitioners in Singapore and Southeast Asia, grasping this structural divergence will be key to decision-making in the second half of the year.