Morgan Stanley reveals new information about NVIDIA GB200: Supply chain initiated, new changes in two segments | AI dehydration

Author: Zhang Yifan

Editor: Shen Siqi

Source: Hard AI

With various market analyses, Morgan Stanley has brought some new insights to the GB200.

The GB200's supply chain has been activated, expecting to drive two new markets: chip testing and glass substrates.

Another area worth noting is Morgan Stanley's continued focus on storage supply and demand, anticipating tight conditions to persist until the second half of 2025, with the upward trend in prices expected to continue.

1. Morgan Stanley: GB200 Supply Chain Update

Recently, Morgan Stanley updated the situation of the GB200 supply chain, mentioning that the supply chain for GB200 DGX/MGX has been initiated. Currently, it is in the stage of design adjustments and testing, with the orders and supplier allocations for 2025 expected to be finalized in the coming months.

1. Morgan Stanley's Shipment Forecast for GB200 in 2024/2025

Based on CoWoS capacity allocation, Morgan Stanley predicts:

• In the second half of 2024, it is expected to deliver approximately 420,000 GB200 super chips to the market;

• In 2025, the production volume of GB200 chips is estimated to be around 1.5 to 2 million.

2. GB200 Drives Two Incremental Markets: Testing and Packaging

Morgan Stanley mentioned two incremental aspects brought by GB200: semiconductor testing and semiconductor packaging.

• Semiconductor Testing: With the strategic use of large chips in GB200, the increase in chip size will lead to a decrease in yield, thereby driving the demand for semiconductor testing. Data shows that in Q2 2024, NVIDIA's AI GPU testing demand increased by 20% compared to the previous quarter.

• Semiconductor Packaging: The advanced packaging process used in GB200 will utilize glass substrates. This is mainly because compared to silicon and organic substrates, glass substrates have advantages such as adjustable strength, low energy consumption, and high temperature resistance. However, the disadvantage is that the cost of using glass substrates is higher than silicon and organic substrates.

3. ASIC Chips Enter the AI Semiconductor Market

Morgan Stanley predicts that ASIC (customized AI chips) will surpass GPU in growth rate in the coming years and may capture 30% of the cloud AI semiconductor market in four to five years.

The main reason is the gradual trend towards personalized demand downstream. Referring to the trend mentioned by Micron and Intel earlier this month that HBM is gradually moving towards customization, it is also due to the trend of personalized demand downstream.

Similarly, ASICs are designed for specific AI tasks, allowing optimization for specific algorithms and models, thereby achieving higher performance than general-purpose GPUs.

Currently, large overseas cloud service providers such as Google, AWS, Tesla, and Microsoft are actively developing and deploying their custom ASICs. The report mentions AWS's Inferentia and Trainium chips, which have a 50% higher performance per watt compared to GPUs.

2. Morgan Stanley: Storage Supply and Demand Tightness Continues

In the second half of 2023, in order to meet the growing demand for products such as HBM, DDR5, and QLC SSD in the market, storage manufacturers increased capital expenditures to expand capacity.

The market believes that manufacturers' active supply will slow down the price growth of memory.

In its report, Morgan Stanley pointed out that despite some easing in the supply chain, the memory market will remain tight until the second half of the 2025 fiscal year. This is based on three factors:

1. Rise of AI Endpoints: AI smartphones and AI PCs will drive the growth of DRAM and RAM capacity in terminal devices.

The report predicts:

• Memory capacity of AI smartphones will increase by 40%, with an average RAM of 28GB;

• AI personal computers (mainly from HP, Lenovo, and Asus) will use 32GB of memory;

• Content on smartphones/laptops will grow at a compound annual growth rate of 14%/12% in the next three years;

1. Impact of Yield Loss: The high yield loss in the CoWoS packaging process used in storage, about 10-15%, will directly affect the effective supply of HBM.

2. NAND Demand Growth: Benefiting from the adoption of QLC SSD (Quad-Level Cell Solid State Drive) in AI servers. QLC SSD, with its high storage density and lower cost, will gradually replace traditional SSDs.

Based on these findings, the report has raised the Total Addressable Market (TAM) for memory in the 2024/2025 fiscal year by 15%-18%