Global GPU Computing News【20260301】

Chip giant AMD and social media giant Meta announced a significant five-year cooperation agreement. Meta plans to deploy AMD Instinct GPUs and EPYC CPUs with computing power capacity of up to 6 gigawatts (GW) for its AI data center construction. The deal is estimated to be worth hundreds of billions of dollars or more, aiming to diversify computing power sources and reduce dependence on Nvidia.

Nvidia released its Q4 2025 earnings, significantly exceeding market expectations with total revenue reaching $68.13 billion, a 73% year-over-year increase. Despite the impressive performance, concerns over the sustainability of future capital expenditures led to the company's stock price plunging 5.5% the day after the earnings release, wiping out nearly $260 billion in market value overnight.

Amidst the stock price drop, Nvidia revealed internal structural details of its next-generation Vera Rubin computing system, which will integrate 72 Rubin GPUs and 36 Vera CPUs. Market attention is focused on the potentially revolutionary new chip, codenamed "Feynman," which may be unveiled at the upcoming GTC 2026 conference on March 16th.

Driven by sustained explosive demand for AI, the supply-demand tension in the memory chip market is intensifying. SK hynix stated it cannot meet all customer demand, with HBM (High Bandwidth Memory) sold out for the entire year, and expects memory prices to continue rising. Samsung's new-generation HBM4 is reportedly priced 20% to 30% higher than previous generations.

The market's pursuit of AI chips is shifting from "training" to "inference," sparking interest in specialized chips like LPUs (Language Processing Units) and SRAM (Static Random-Access Memory) technology. Rumors suggest Nvidia will showcase related technologies at the GTC conference, which also drives demand expectations for upstream materials like high-layer count PCBs.

Google released its new-generation model, Gemini 3.1 Pro, featuring comprehensive improvements in reasoning capabilities and multimodal performance. The continuous upgrading of models and the deepening application of AI agents are persistently driving strong demand for global underlying computing power hardware.

A landmark event occurred: Global AI model invocation data for February 2026 showed that the weekly invocation volume of Chinese AI models surpassed that of the United States for the first time. Among the top five models globally by invocation volume, Chinese models claimed four spots, demonstrating a collective rise.

Against the backdrop of surging AI model invocations in China, the market logic for computing power has undergone a dramatic shift. The A-share market reacted enthusiastically, with sectors like computing power leasing, cloud computing, and data centers experiencing a surge of limit-up trades. The stark contrast between Nvidia's sharp decline and the A-share rally formed an "ice and fire" scenario, prompting the market to reassess the distribution of computing power value.

Domestic GPU manufacturers have entered a critical period of value realization. Moore Threads and MXC (Muxi Holdings) successively released performance forecasts, showing revenue growth of 243.37% and 121.26% respectively for 2025, with losses significantly narrowing. This indicates that domestic general-purpose GPUs, leveraging technological strength and cost advantages, are gradually gaining market recognition, benefiting from the exponential growth in domestic computing power demand.

The outlook for domestic computing power demand is widely viewed as promising. Analysts point out that Chinese AI models, utilizing efficient MoE (Mixture of Experts) architectures and significant cost advantages (such as lower electricity costs), are weakening the market's linear dependence on Nvidia's high-end GPUs. JPMorgan Chase predicts that China's token consumption will grow at a compound annual growth rate of 330% from 2025 to 2030, which will directly generate immense demand for domestic computing power infrastructure.