NVIDIA's GB200 Leads AI Inference Profitability at 78%, AMD Faces Losses in Key Sector

A recent report from Morgan Stanley highlights the burgeoning profitability of the global AI inference market, with standardized "AI inference factories" generally achieving average profit margins exceeding 50%. The analysis, which utilizes a detailed financial model, positions NVIDIA's GB200 chip at the forefront, boasting an impressive profit margin of approximately 78%. Google and Huawei's AI chips were also identified as "guaranteed profit businesses," underscoring the lucrative nature of this specialized segment of the artificial intelligence industry.

The comprehensive assessment by Morgan Stanley, as reported by Jukan on social media, indicates a significant divergence in performance among major chip manufacturers. According to the tweet, "NVIDIA’s GB200 ranked first with an astonishing profit margin of about 78%. Google and Huawei’s chips were also assessed as 'guaranteed profit businesses.'" This robust profitability for NVIDIA is attributed to its superior computing, memory, and networking performance, alongside the strong ecosystem of its CUDA software, which continues to dominate the market.

Conversely, AMD, which had garnered high market anticipation, was found to be suffering significant losses in the AI platform's inference sector. Morgan Stanley's model calculated negative profit margins for AMD's MI300X and MI355X platforms, at -28.2% and -64.0% respectively. This unexpected outcome stems from a severe imbalance between high total cost of ownership (TCO) and lower token output efficiency in inference tasks, as detailed in the report.

Industry analysts have weighed in on AMD's performance, noting that Morgan Stanley's model assumed lower utilization and efficiency for AMD's accelerators compared to NVIDIA's and Google's, partly due to a less mature software ecosystem (ROCm). Despite these findings, some market observers suggest that future software updates, such as ROCm7 expected in Q3 2025, could significantly improve AMD's inference performance and address some of the current profitability challenges.

The Morgan Stanley report's "100MW AI Factory Model" quantifies investment returns by using 100 megawatts of power consumption as a baseline, calculating comprehensive TCO including infrastructure and hardware costs, and linking revenue to token output based on mainstream API pricing. This model underscores that while AI inference is highly profitable overall, the efficiency and ecosystem maturity of the chips play a critical role in determining individual profitability. The findings highlight the intense competition and strategic importance of hardware and software integration in the rapidly evolving AI compute landscape.