Morgan Stanley issues a "tranquilizer": Nvidia is fine, Blackwell chip's fourth-quarter delivery schedule remains unchanged

Facing doubts about the AI bubble, Nvidia received support from Morgan Stanley, stating that there are no issues with the delivery of Blackwell chips in the fourth quarter of 2024.

On August 12th local time, Morgan Stanley analyst Sharon Shih and her team released a research report stating that after conducting a series of investigations on major suppliers of chips, computing boards, servers, etc., Morgan Stanley expects that Nvidia's Blackwell GPU and related server components will not have any changes in their delivery schedule for the fourth quarter of 2024.

Major chip suppliers such as TSMC and King Yuan Electronics, major server suppliers such as Wiwynn and Kingston, and major investors such as Foxconn and Pegatron all have a high probability of benefiting from Nvidia's Blackwell chips, leading to an increase in stock prices.

Morgan Stanley Expects No Change in Blackwell Chip Delivery Time

From the chip perspective, in last week's AI tracking report, Morgan Stanley pointed out:

TSMC's Blackwell chip production is expected to be delayed by about two weeks between mid to late September. It is estimated that around 620,000 Blackwell chips will be produced in the second half of 2024, mainly concentrated between October and December.

King Yuan Electronics also confirmed that its main AI GPU customer Nvidia has requested to expand new testing capacity, further confirming that chip production will not have significant changes in the fourth quarter of this year. Capacity utilization data from King Yuan Electronics is similar to our estimate of TSMC's CoWoS-L (an advanced packaging technology) capacity, with fourth-quarter capacity of around 10kwpm.

The above data indicates that the production volume of Blackwell chips in the fourth quarter should be around 150,000 chips per month.

In addition, many investors are concerned about the yield issue of CoWoS-L. Morgan Stanley's inspection shows that the yield rate is between 95-96%, so it is unlikely that delays in delivery will occur due to yield issues.

From the perspective of computing boards/servers, Morgan Stanley estimates that around 620,000 Blackwell chips will be produced in the second half of this year. It can be seen that Wiwynn and Foxconn have increased the production volume of "Bianca" and "Ariel" GPU computing boards for the first quarter of 2025.

Moreover, B200A will be introduced into the HGX architecture and is expected to be equipped with UBB, supplied by Wiwynn. Morgan Stanley expects B200A UBB to start production in the first quarter of 2025.

From the perspective of thermal component suppliers, Morgan Stanley has not seen any changes in the delivery schedule of GB200 server system thermal components (including cold plate modules and cooling fans).

The newly designed B200A will replace B100/B200 in HGX servers. B200A HGX servers will continue to use 3D VC cooling design, while GB200A will be designed as 2.5D VC (with better production yield than 3D VC)From the perspective of server/rack delivery, Morgan Stanley's inspection shows that server manufacturers ODMs are continuing to prepare for the large-scale production and capacity expansion of NVIDIA's Blackwell GPU servers/racks.

The production of GPU boards and compute boards typically takes 2-3 weeks to be ready for immediate shipment once the GPU kits are prepared. Server assembly and rack installation may also require an additional 1-2 weeks and 1-2 months. Therefore, Morgan Stanley expects the related components to start production at the end of the fourth quarter of this year for the assembly of GB200 server rack systems later on