We summarize key data points and updates from the AI supply chain to help investors track real supply-demand situations. NVIDIA issued upbeat October quarter guidance for revenue of US$16bn: That's well above Asia-based investors' expectations at US$13-15bn. 

AI downstream supply chain: 

• Our AI server shipment estimate is higher at 373K units in 2024 (up 91% YoY), vs. our prior estimate of 267K units ( Exhibit 15 ). 
• Per our checks, Wistron is still 100% H100 GPU baseboard assembly now and will see strong ramp in 3Q23, followed by a milder ramp in 4Q23. In 2024, FII may cut into the GPU baseboard supply per NVIDIA's intention to dual-source, but Wistron could also handle GPU module assembly. 
• Per our checks, 15-20% of AI server-related shipments are pushed out to 2024 for Inventec, reflecting stronger demand but component tightness. 
• Inventec may start to do L11 assembly for CSP in Mexico in 2024 in small volume. 
• Wiwynn's ASIC AI projects are mainly for Amazon and volumes may start to pick up from 4Q23 onwards.
 • As the sole supplier for NVIDIA's system level testing (SLT) equipment, Chroma indicates that there has been no order push-out of late, implying that shipments and revenue from this business should grow meaningfully in the remaining months of the year.

AI semi supply chain updates
 CoWoS – bottleneck easing

 • NVIDIA indicates that CoW(by TSMC) would not limit its H100 GPU shipments in the coming quarter. The company "expects supply to increase each quarter through next year." • Exhibit 1 shows CoWoS expansion by vendors. Meanwhile, TSMC is hiking CoWoS prices 20% for rush orders from small customers. • The new NVIDIA GPU revenue forecasts imply 129k CoWoS consumption in 2024e (see Exhibit 4 )

AI GPU supply 

• Our foundry supply chain checks suggest that NVIDIA GPU shipments may grow 30-40% Q/Q in 4Q23, after +10-15% Q/Q in 3Q23.