This is not investment advice. The author has no position in any of the stocks mentioned. Wccftech.com has a disclosure and ethics policy.
[Update - 8/28/2025] - Morgan Stanley has retracted their article covering the inference performance of various AI accelerators, including those from NVIDIA and AMD. The firm has stated that the methodology used for these tests overstated NVIDIA's advantages and understated the real-world performance offered by AMD's Instinct lineup. Due to several inaccuracies within the post, it has been removed, and a revised test will be published later.
With GPU economics driving a lot of consternation in the financial circles these days, Morgan Stanley has come forward with a fairly convincing note on the unrivalled efficiency advantages of leveraging NVIDIA's GB200 NVL72 GPUs for large-scale AI factories.
For the benefit of those who might not be aware, each NVL72 AI rack contains 72 NVIDIA B200 GPUs plus 36 Grace CPUs, connected via the NVLink 5 high-bandwidth, low-latency interconnect. Do note that each such server rack currently costs around $3.1 million vs. just around $190,000 for an H100 rack.
However, Morgan Stanley believes that it makes much more economic sense to go with NVIDIA's latest rack-scale offering as opposed to the older gen H100k
As per the Wall Street giant's calculations, NVIDIA's GB200 NVL72 systems currently lead the proverbial pack in terms of their ability to generate revenue and profits, followed by Google's TPU v6e pod.
Specifically, a given 100MW AI factory can achieve a 77.6 percent profit margin with NVIDIA's GB200 NVL72 AI racks, while Google's TPU v6e pods rank a close second with a profit margin of 74.9 percent.
Do note that the pricing of Google's TPU v6e pods is not publicly available. However, on average, it costs between 40 and 50 percent less to rent a pod as opposed to the NVL72 rack.
Interestingly, as per Morgan Stanley's calculations, AI factories employing AMD's MI300 and MI355 platforms have negative profit margins, to the tune of -28.2 percent and -64 percent, respectively.
Morgan Stanley's report assumes that a 100MW AI data center entails infrastructure costs of $660 million, depreciated over a 10-year period. GPU costs can vary between $367 million at the low-end to $2.273 billion at the high-end, with depreciation over a 4-year period. Finally, the bank calculates operating costs by applying the power efficiencies delivered by various cooling systems to an average of global electricity prices.
As such, NVIDIA's GB200 NVL72 systems have the highest Total Cost of Ownership (TCO) at $806.58 million, followed by the MI355X platform at $774.11 million.
Follow Wccftech on Google to get more of our news coverage in your feeds.
