Investment bank Morgan Stanley believes that even though building a data center with NVIDIA's Blackwell GPUs costs twice as much as building one with AI application-specific integrated circuits (ASICs), the computing power efficiency of the NVIDIA chips is significantly greater than the custom chips as well. The high costs of NVIDIA's latest AI GPUs are a hot-button market topic, with CEO Jensen Huang having asserted on multiple occasions that even though his chips are pricey, they offer greater returns over the long term.
NVIDIA's Compute Performance Per Watt Is Up to 8x Ahead Of Custom AI Chips, Says Morgan Stanley
In its recent coverage, Morgan Stanley compares the TFLOPS (Trillion Floating Point Operations Per Second) per Watt performance of NVIDIA's different AI GPUs to custom AI ASICs offered by Amazon and Google. It comments that the capital expenditure by hyperscalers to build a one gigawatt data center with NVIDIA's Blackwell AI GPUs is twice when compared to building the same data center with Google's tensor processing units (TPUs) or Amazon's Trainium chips.
However, it adds that investing in NVIDIA's chips is worth the spend since they offer greater computing power efficiency. According to Morgan Stanley's estimates, the performance per Watt of the NVIDIA chips is "2x-8x ahead of custom ASICs."
NVIDIA's Rubin Chips Significantly Outperform Amazon & Google's Custom AI Chips, Says Morgan Stanley
The investment bank's slide that accompanies its report computes the TFLOPs per Watt performance of NVIDIA's Vera Rubin (FP4), Vera Rubin (FP8), GB300 (FP8) and H100 (FP8) AI GPUs. Naturally, the Vera Rubin (FP4) is the highest performing GPU on the list as its score sits at 19.5. For the other chips, the scores sit at 6.8, 6.0 and 3.1, respectively. On the other hand, the TFLOPs per Watt for Google's TPUv7 (FP8) and Trn3 (FP8) chips are 4.3 and 2.5, respectively, which places their performance either between the Blackwell and the Hopper generation GPUs or below the Hopper chips.
While NVIDIA's chips offer the greatest performance per Watt, users are shifting towards other metrics as well. For instance, according to an expert from AI infrastructure provider Nebius, AI chips are also being evaluated through their cost per million tokens generated over the hourly cost of running a GPU. Estimate from the Nebius shows that Groq's AI chips cost between five to ten cents per token, while NVIDIA's Blackwell chips cost 25 cents per token. The Groq chips are also purportedly capable of delivering up to 800 tokens per second, which is significantly higher than the NVIDIA chips' 450 tokens per second.
About the author: Ramish is a seasoned technology writer and editor with more than a decade of experience. He specializes in semiconductor fabrication and market analysis. With a background in finance and supply chain management - via his bachelors in Finance and a micromasters in supply chain management from MIT - Ramish combines financial rigor with deep industry insight to deliver accurate and authoritative coverage.
Follow Wccftech on Google to get more of our news coverage in your feeds.
