AMD downgraded: Nvidia's AI chips still hold 'a significant performance advantage'

Investing.com -- Jefferies on Thursday downgraded Advanced Micro Devices (NASDAQ:AMD) to Hold, citing a widening performance gap with Nvidia's (NASDAQ:NVDA) AI chips and limited traction in the AI market.

The investment bank’s price target on AMD shares was also trimmed to $120 from $135.

Despite AMD's MI300x boasting higher advertised TeraFLOPs (TFLOPs) and memory bandwidth than Nvidia’s H200, Jefferies’ proprietary benchmarking suggests that the H200 "retains a significant performance advantage over the MI300x," with expectations that "the gap expands even further with Blackwell and Rubin."

The benchmarking analysis, conducted over the past two months with Trelis Research, compared real-world throughput of Nvidia’s H200 and AMD’s MI300x across open-source models such as Llama 405B, Llama 70B, and DeepSeek R1.

Results showed that Nvidia’s H200 consistently outperformed AMD’s MI300x, with throughput results beating AMD by "a wide margin (80-300%+)," demonstrating the strong impact of Nvidia’s more mature software stack.

Even though MI300x has higher reported compute power—2615 TFLOPs FP8 Dense versus H200’s 1979 TFLOPs FP8 Dense—and greater memory bandwidth, AMD’s hardware struggled in real-world inference workloads.

“A valid pushback would be that the AMD solution is not as well optimized for the models we tested, but we view this as exactly the point,” analysts led by Blayne Curtis said in a note.

“These results underscore the importance of NVDA’s mature software stack for managing GPU efficiencies and this gap will widen as system performance,” they added.

Additionally, Jefferies emphasized that Nvidia’s AI accelerator model maintains an advantage in total cost of ownership (TCO), with Blackwell GPUs expected to offer best-in-class TFLOPs per watt (TFLOPs/W) metrics.

While some competing chips outperform Nvidia on a GigaFLOPs/$ basis, "the more relevant stat for TCO is TFLOPs/W, where NVDA is pulling ahead with Blackwell."

Jefferies also pointed out that AMD faces increased competition from Intel (NASDAQ:INTC), which is expected to introduce more competitive chips starting in 2026.

"We see Intel’s new management as a catalyst for them to become more competitive, particularly in Client PC," the report noted.

Moreover, AMD is not expected to have a viable rack-level solution until the MI450, likely arriving in 2027, by which time Nvidia is expected to move to next-generation rack designs that could fit four times more GPUs.

AMD shares have already retreated substantially, losing more than 32% in the past six months. Nonetheless, analysts believe Street estimates remain too high “and we see more downside risk than upside potential near-term,” they said.