Login to Continue Learning
NVIDIA’s Blackwell GPUs have secured the lead in AI inference performance, resulting in higher profit margins for companies using them compared to competitors. According to Morgan Stanley Research, most AI inference “factories” are enjoying profit margins of over 50%. NVIDIA is at the forefront with a 77.6% profit margin and an estimated revenue of $3.5 billion.
In comparison, Google’s TPU v6e pod secures a 74.9% profit margin, while AWS Trn2 Ultraserver has a 62.5% profit margin. AMD’s latest MI355X platform yields a negative 28.2% profit margin, and the older MI300X platform has a negative 64.0% profit margin in AI inference.
NVIDIA’s GB200 NVL72 chip generates $7.5 revenue per hour, second to HGX H200 at $3.7, while AMD’s MI355X yields only $1.7 per hour. Other chips typically generate between $0.5 and $2.0 in revenue per hour.
NVIDIA’s massive lead is attributed to its FP4 support and continued optimizations of the CUDA AI stack. The company has shown “Fine Wine” treatment for several older GPUs like Hopper and Blackwell, with incremental performance upgrades every quarter.
AMD’s MI300 and MI350 platforms are strong in hardware, but software optimization efforts still need improvement, especially in AI inference.
Morgan Stanley also highlighted that the total cost of ownership (TCO) of NVIDIA’s GB200 platform is around $800 million, matching AMD’s MI300X at $744 million. However, NVIDIA offers significantly higher AI inference performance, which accounts for 85% of the market in coming years.
NVIDIA and AMD are maintaining an annual cadence to stay competitive. In 2023, NVIDIA will launch Blackwell Ultra with a 50% uplift over GB200. Rubin is set for production in H1 2026, followed by Rubin Ultra and Feynman. AMD plans to release MI400 next year to challenge Rubin, with expected AI inference optimizations on MI400.
Sources: WallStreetCN, Jukanlosreve
