BlockBeats News, June 30th, the wave of artificial intelligence capital expenditure is spreading from the GPU itself to a deeper supply chain including wafer foundries, advanced packaging, glass substrates, TPUs, and testing equipment.

Morgan Stanley outlined a more complete AI hardware chain in four recent reports: Cloud providers such as Google are still significantly increasing their computing power, TSMC continues to dominate advanced processes and key capacities like CoWoS, packaging materials and testing equipment suppliers are being repriced, and panel manufacturers are also trying to enter the next-generation advanced packaging market through glass substrates.

The most crucial assessment still falls on TSMC. Morgan Stanley stated that Samsung and Intel's catch-up in advanced processes poses a limited threat to TSMC. The firm believes that TSMC has a larger production capacity, a more advanced EUV equipment supply, and stronger process execution capability. By the end of 2028, TSMC's combined capacity for 2nm and 3nm nodes could reach 400,000 wafers per month, close to 10 times the scale of Intel's 14A/18A and Samsung's SF2.

Driven by demand for AI CPUs, GPUs, ASICs, and photonic chips, Morgan Stanley expects TSMC's revenue growth in 2026 to be close to 40%, capital expenditure to rise to $56 billion, and reach $75 billion in 2027. The firm also predicts that TSMC will be able to raise the prices of advanced process wafers by 5% to 10% in 2027.

Advanced packaging is another bottleneck. Morgan Stanley forecasts that TSMC's CoWoS capacity could expand to 200,000 wafers per month by 2027, and SoIC capacity to 70,000 wafers per month. As AI chip packaging sizes increase, HBM usage rises, and ASIC projects multiply, CoWoS is no longer just complementary capacity for NVIDIA GPUs but a key constraint for the entire AI computing infrastructure.

The demand side is supported by Google. Morgan Stanley projects in the Alphabet report that by 2028, Google will add about 9GW of on-premises computing power, with approximately 7GW coming from TPUs and another 2GW from NVIDIA GPUs. The firm estimates that Google will retain about 5GW of TPUs for internal and cloud services such as search, YouTube, and Google Cloud, and sell about 4GW of TPUs as first-party systems to external customers.

This will significantly change Google Cloud's revenue structure. Morgan Stanley forecasts that Google Cloud's revenue could exceed $300 billion by 2028, with EBIT reaching around $132 billion, accounting for approximately 46% of Alphabet's overall EBIT. Therefore, the firm has raised Alphabet's target price from $375 to $415, while increasing the 2027 capital expenditure forecast from $300 billion to $350 billion.

This also explains why Morgan Stanley remains bullish on the Greater China AI Semiconductor Chain. The bank expects the global semiconductor market to reach $1.5 trillion by 2030, with around half of it coming from AI semiconductors; the AI semiconductor Total Addressable Market (TAM) could be close to $753 billion. Beneficiaries include not only TSMC but also MediaTek, GUC, Alchip, Himax, Unimicron, Aspeed, ASMPT, FOCI, WinWay, MPI, and Hon Precision, covering design, packaging, testing, and optical interconnection.

Meanwhile, panel manufacturers are also being incorporated into the advanced packaging narrative. Morgan Stanley stated that glass substrates have potential in future large-size AI/HPC packaging, as they may bring lower signal loss, better thermal expansion matching, and stronger mechanical stability. Innolux is believed to be making rapid progress with glass core substrates, BOE is attempting to cover the entire process from TGV to build-up layers, while AUO is more focused on LEO satellite antennas and CPO optical modules.

However, Morgan Stanley is more cautious about this segment. The bank believes that glass substrates may not truly enter HPC mass production until 2028 or 2029, and the recent surge in panel stocks largely reflects market optimism. It favors BOE, seeing its scale and valuation as more attractive; while maintaining a neutral view on Innolux and AUO.

Overall, the key message conveyed in Morgan Stanley's recent four reports is that AI investment is entering a deeper supply chain reevaluation phase. GPUs remain at the center, but bottlenecks are spreading to advanced processes, CoWoS, SoIC, HBM, ASIC, testing equipment, optical interconnects, and new packaging materials. Morgan Stanley's core assessment is that the market is underestimating the continued AI computing power demand from 2027 to 2028, as well as the pricing power of TSMC and its related supply chain.