Why Are NVIDIA Blackwell GPUs More Expensive?
Morgan Stanley analyzes why NVIDIA Blackwell GPUs are valuable for data centers with their superior performance per watt despite their high cost.
Investment bank Morgan Stanley states that information centers built with NVIDIA’s Blackwell graphics process units cost twice as much as custom-made chips from Google and Amazon. However, the bank argues that the high process power efficiency offered by NVIDIA chips offsets this cost difference.
NVIDIA CEO Jensen Huang has stated many times that although its chips are high-priced, they provide more returns in the long run. Morgan Stanley’s current report adds a new dimension to the cost debate in the market with technical data supporting this view.
NVIDIA’s Process Power Efficiency Leads Ahead of Custom Chips
Morgan Stanley compares the process performance per watt (TFLOPS) of different AI GPUs from NVIDIA with Amazon’s Trainium and Google’s TPU chips. According to the analysis, in a one-gigawatt data center installation, NVIDIA Blackwell GPUs cost twice the cost of competing custom-built chips.
However, the bank emphasizes that NVIDIA’s performance costs per watt are 2 to 8 times higher than custom-made chips. This efficiency difference puts NVIDIA in a different position in the industry in terms of total cost of ownership and long-term operational success.
According to the road map shared by NVIDIA, the Blackwell architecture will be followed by Rubin architectures in 2026 and Feynman architectures in 2028. These new generation chips reach high performance values of 19.5 TFLOPS per watt, especially with models such as Vera Rubin (FP4).
Alternative Metrics and Cost Discussions
Evaluations in the market do not only focus on performance per watt, and different metrics also stand out. Infrastructure providers like Nebius use a valuation formula based on the cost per million tokens produced rather than the hourly cost of GPUs.
According to Nebius information, the cost per token of Groq chips varies between 5 and 10 cents, while this number remains at 25 cents for NVIDIA Blackwell chips. It is also stated that Groq chips have the capacity to generate 800 tokens per second, while NVIDIA is further behind in this field with 450 tokens per second. This competition between NVIDIA’s high performance efficiency and the token cost advantages of other manufacturers shapes the future of data center investments.
Do you think the total cost or performance efficiency per watt is more critical in information centers?
