Advanced Micro Devices Inc. (NASDAQ:AMD) is closing the gap with Nvidia Corp. (NASDAQ:NVDA) in AI inference at a pace that has surprised even researchers who previously criticized AMD’s software.

SemiAnalysis researchers said Wednesday they watched AMD’s existing GPUs catch and briefly overtake some Nvidia configurations as engineers improved the software running on the chips over just weeks.

AMD’s Software Gains Surprise Researchers

“Lots of people have been anti-AMD for a long time, especially the AMD software ecosystem, including us,” SemiAnalysis host Jordan Nanos said during Wednesday’s podcast.

The firm’s InferenceX platform tracks how efficiently competing GPUs run GLM 5.3, a Chinese AI model used for complex tasks including coding.

Nvidia initially led, but AMD’s performance improved as engineers optimized its ATOM inference engine, which helps AI models generate responses. AMD eventually overtook some Nvidia configurations before the latter began recovering ground.

Researcher Wenyao Gao explained that while a chip’s hardware remains fixed after launch, engineers can continually improve the software controlling how efficiently it serves users.

In September research, SemiAnalysis found AMD’s MI355X delivered roughly 9% lower estimated inference costs than Nvidia’s B200 when both systems had to begin responding within two seconds.

Nvidia’s GB200 and GB300 systems remained cheaper in other configurations, however, highlighting how performance depends on the workload and response-time requirements.

Same Chips, Up to 38% More Performance

AMD’s improvements extend beyond the GLM 5.3 tests.

In September, AMD reported that its MI355X chips could handle up to 38% more AI work than in an earlier round of MLPerf, an industry benchmark that measures chip performance.

The striking part is that the hardware hadn’t changed. AMD achieved the improvement entirely through software updates.

In another comparison, a system using 72 AMD GPUs outperformed an earlier setup with 94 of the same chips. For cloud providers, that could mean serving more customers without buying additional hardware.

AMD is also finding ways to improve its software faster.

Futurum Research reported that AMD engineers are using AI agents to help write and optimize code. The company has also accelerated updates to its ROCm software platform, releasing new versions roughly every six weeks instead of every three months.

Nvidia’s CUDA software ecosystem still gives it a major advantage, but AMD is showing how quickly that gap can narrow without a new generation of chips.

Despite AMD’s progress, prediction traders continue to favor Nvidia’s broader market dominance. On Polymarket, Nvidia has approximately 76% odds of finishing 2026 as the world’s most valuable company, with $8.2 million traded.

Nvidia continues to benefit from intense GPU demand, but AMD’s improving software could give cloud providers a more competitive alternative without requiring a new generation of chips.

Image: Shutterstock