NVIDIA's new 2U air-cooled MGX GB200 NVL2 server continues to suffer from "thermal issues" says Morgan Stanley.
In a new investment note, Morgan Stanley provides an updates on the AI scene with notes that NVIDIA's MGX GB200 NVL2 with its 2 x Grace CPUs and 2 x Blackwell B200 AI GPUs on the same PCB board, is suffering from "thermal issues".
The full note explained: "NVIDIA MGX GB200 NVL2 houses 2x Grace and 2x B200 Blackwell GPUs on the same PCB board, with the GPU module connecting to the main PCB board using an SXM7 module. All of the servers showcased at OCP were based on a 2U air-cooled form factor. However, our conversations with supply chain partners indicated to us that there are still some thermal issues with the 2U form factor, so this may potentially end up being in a 4U form factor instead".
- Read more: NVIDIA GB200 NVL72: 'highest-power-consuming server in HISTORY' at 132kW
- Read more: NVIDIA 'halting developing' of GB200 NVL36x2 AI servers
AI semi supply chain data points from Morgan Stanley:
- AMD seems to have trimmed some CoWoS wafer booking at TSMC for 2025 given uncertainty around MI325 demand, but capacity was immediately taken by NVIDIA.
- Intel Habana hasn't changed its wafer booking at TSMC.
- Marvell's CoWoS booking at TSMC is 3x higher in 2025 vs. 2024. If the output for Tranium 2 is 200k units in 2024, based on CoWoS capacity booking, we believe there could be 400k units of Tranium 2 in 2025 and 500k-600k units of Inferentia 2.5 in 2025.
- WPG sales grew 25% Q/Q in 3Q vs. previous guidance of only up 5.5% Q/Q. We cite increasing business coming from AMD CPU and "GPU".
- NVDA MGX GB200 NVL2 houses 2x Grace and 2x B200 Blackwell GPUs on the same PCB board, with the GPU module connecting to the main PCB board using an SXM7 module. All of the servers showcased at OCP were based on a 2U air-cooled form factor. However, our conversations with supply chain partners indicated to us that there are still some thermal issues with the 2U form factor, so this may potentially end up being in a 4U form factor instead.