Global GPU Computing News【20260308】

Nvidia Plans to Launch Dedicated Inference Chips: It has been revealed that Nvidia is designing a new system focused on "inference" computing, aimed at more efficiently supporting response queries for AI models. The system is expected to be unveiled at this month's GTC conference and will integrate technology from the startup Groq (specifically its LPU), with OpenAI projected to become its largest customer.

Domestic GPU Manufacturers Report Strong Performance: Moore Threads and MXC both released their 2025 performance reports, showing revenue more than doubling. Moore Threads' revenue broke through 1.505 billion RMB, a year-on-year increase of 243.37%; MXC's revenue reached 1.644 billion RMB, a year-on-year increase of 121.26%. Losses also significantly narrowed, indicating the accelerating commercialization of domestic AI chips.

GTC Conference Preview Focuses on Computing Infrastructure Upgrades: As Nvidia's GTC conference approaches, market focus has shifted from individual chips to infrastructure areas like power architecture, optical communication, and liquid cooling. To support new-generation GPUs with power consumption reaching thousands of watts, technologies like 800V high-voltage power supplies, CPO switches, and full-liquid cooling solutions are expected to be key highlights of the conference.

Rental Prices for High-End GPUs Soar: Driven by surging AI inference demand and tight supply, rental prices for Nvidia's high-end GPUs have entered an upward cycle since the beginning of 2026. By the end of February, the monthly rental price for the H200 model had risen to 60,000-66,000 RMB, a month-on-month increase of 25%-30%. H100 prices also increased by 15%-20%, with order backlogs extending into 2027.

Token Consumption Grows Exponentially: The transformation of AI applications from "Q&A tools" to "productivity tools" is driving a surge in token consumption. Data from MiniMax shows its model's average daily token consumption has increased over sixfold compared to last December, with coding scenarios growing more than tenfold. This sustained demand for inference computing power presents an opportunity for value reassessment in the chip industry.

MXC Issues Q1 Performance Forecast, Losses Narrow: MXC disclosed its performance forecast for the first quarter of 2026, projecting revenue between 400 million and 600 million RMB, a year-on-year increase of 24.84% to 87.26%. The net loss attributable to parent company narrowed significantly by 21.93% to 60.97% year-on-year. This positive news boosted its stock price by 5.66% in a single day, reflecting strong market expectations for an inflection point in the profitability of domestic GPUs.

Paratera Plans to Invest Over 80 Million RMB in Computing Equipment Procurement: To meet business development needs, Paratera announced its intention to procure GPU computing servers and storage equipment from suppliers, with a total estimated amount not exceeding 80.66 million RMB. The procurement amount for GPU computing servers is capped at 27.36 million RMB. The relevant proposal has been reviewed and approved by the board of directors.

Chinese Models' Token Calls Surpass the US for the First Time: Data from OpenRouter, a global API aggregation platform for AI models, showed that in the week from February 9th to 15th, Chinese models recorded 4.12 trillion token calls, surpassing the US models' 2.94 trillion for the first time. This data reflects the explosive growth in domestic AI application demand and consequently drives up the need for underlying domestic computing power.

Domestic GPU Performance Approaches International Mainstream Levels: Domestic chips have made breakthroughs in computing density. The single-card FP8 computing power of Moore Threads' S5000 has exceeded 1000 TFLOPS, with training accuracy differing by less than 1% from Nvidia's H100. Meanwhile, Huawei is expected to release the Ascend 950PR in the first quarter of this year, also achieving 1 PFLOPS (FP8) of computing power. Domestic chips are moving from "technologically viable" to "deployable at scale."

Domestic GPUs Accelerate Ecosystem Adaptation and Commercial Deployment: Facing the barrier of Nvidia's CUDA ecosystem, domestic manufacturers are forging different paths. MXC is attracting developers through its open-source MXMACA strategy, with its C500/C550 series gradually adapting to multiple domestic large models like Tencent's Hunyuan and Zhipu's GLM-5. Moore Threads adheres to its self-developed MUSA architecture, cultivating a native ecosystem while maintaining compatibility.