NVIDIA’s Blackwell GPUs have positioned themselves at the forefront of AI inference performance, leading to significantly higher profit margins for businesses leveraging this technology compared to their competitors.
NVIDIA’s Comprehensive AI Software and Optimizations: A Competitive Edge Over AMD
Recent analysis from Morgan Stanley Research provides a thorough comparison of profit margins and operational costs associated with AI inference workloads among various technological providers. The findings reveal that firms functioning as AI Inference “factories”are enjoying profit margins exceeding 50%, with NVIDIA emerging as the clear frontrunner.
The study evaluated a range of AI factories, specifically those requiring 100MW of power, incorporating server racks from several key industry players, including NVIDIA, Google, AMD, AWS, and Huawei. Among these, NVIDIA’s GB200 NVL72 “Blackwell”GPU platform stands out, achieving an impressive profit margin of 77.6% and an estimated profit of approximately $3.5 billion.
Google’s TPU v6e pod follows closely, with a profit margin of 74.9%, while AWS’s Trn2 Ultraserver secures the third position with a 62.5% profit margin. Other solutions are noted to have profit margins around 40-50%, but AMD has considerable ground to cover, as indicated by its performance metrics.
In stark contrast, AMD’s transition to its latest MI355X platform has resulted in a troubling negative profit margin of 28.2%.The earlier MI300X model fared even worse, with a staggering negative 64.0% profit margin in terms of AI inference performance. Morgan Stanley’s report also breaks down revenue generation per chip per hour, which shows NVIDIA’s GB200 achieving $7.5 per hour, followed by the HGX H200 at $3.7. In stark comparison, AMD’s MI355X generates only $1.7 per hour, while most other competitors range between $0.5 to $2.0, indicating NVIDIA’s dominance in this space.
The significant advantage NVIDIA holds in AI inference results primarily from its support for FP4 and ongoing enhancements to its CUDA AI stack. The company has effectively treated several of its earlier GPU models, including Hopper and even Blackwell, with what can be likened to fine wine treatment—incrementally boosting their performance each quarter.
While AMD’s MI300 and MI350 platforms are adept in terms of hardware capabilities, the company still faces challenges in optimizing software for AI inference, an area where improvements are critically needed.
Notably, Morgan Stanley also highlighted the Total Cost of Ownership (TCO) for AMD’s MI300X platforms reaching as high as $744 million, comparable to NVIDIA’s GB200 platform at approximately $800 million. This indicates that AMD’s cost structure may not be favorable in the competitive landscape. The newer MI355X server’s estimated TCO of $588 million aligns similarly with Huawei’s CloudMatrix 384, but the higher initial expenditure may deter potential users from choosing AMD, especially considering NVIDIA’s superior AI inference performance that is projected to dominate 85% of the AI market in the coming years.
As both NVIDIA and AMD strive to keep pace with each other, NVIDIA is set to launch its Blackwell Ultra GPU this year, promising a 50% performance uplift over the existing GB200 model. Following this, the forthcoming Rubin platform is slated for production in the first half of 2026, accompanied by Rubin Ultra and Feynman. Meanwhile, AMD plans to introduce the MI400 next year to contend with Rubin and is expected to implement several AI inference optimizations for its MI400 line, which will create an interesting dynamic competition in the AI segment.
News Sources: WallStreetCN, Jukanlosreve
Related Articles:
Intel Confirms Ongoing GPU Product Development and NVIDIA Partnership Enhances Product Roadmap
21:03
GeForce RTX 5070 Price Falls Below MSRP; Other RTX 50 Series Graphics Cards Also Available at MSRP
20:26
4 Key Reasons Ryzen 5 9600X Outsells Core i5 14600K Despite Having Fewer Cores and Slower Performance
18:50
Leave a Reply