Nvidia Server Benchmark Shows 10x Gain for Mixture‑of‑Expert Models; Huang Meets Trump on Export Rules

Nvidia released benchmark results showing its new AI server delivers up to 10x performance gains on mixture‑of‑expert models, including tests with Moonshoot AI’s Kimi K2 Thinking. The server is built around 72 high‑performance chips with fast interconnections in a single machine and is positioned toward the inference market ^[1].

The company’s benchmark release highlights growing competition in inference hardware: AMD and Cerebras are named competitors, and AMD is developing a similar multi‑chip server planned for 2026. The mixture‑of‑expert approach gained traction after DeepSeek’s efficient model launch in early 2025, per the report ^[1].

Separately, Nvidia CEO Jensen Huang met with President Trump to discuss chip export restrictions and AI regulation. The report notes the GAIN AI Act provision requiring U.S. customer priority will not be included in the defense bill, Huang warned that state‑by‑state AI rules could impede industry progress and threaten national security, and he downplayed GPU smuggling concerns by citing each chip’s weight and estimated price tag ^[2].

Nvidia’s new server benchmarks underscore a push to capture inference workloads amid rising hardware competition, while regulatory and export discussions — highlighted by Huang’s meeting with President Trump — could influence market access and deployment of such multi‑chip systems ^[1][2].