1. Global GPU computing shortage worsens, H100 rental prices surge 40% in six months
Data from industry research firm SemiAnalysis shows that the one-year rental price of an NVIDIA H100 GPU has soared from US$1.70 per GPU per hour in October 2025 to US$2.35 in March 2026, an increase of nearly 40%. On-demand computing power for all types of GPUs is completely sold out, with the market noting that finding GPU computing power in early 2026 is "like trying to book a ticket on the last flight out."
2. Blackwell series in short supply, new cluster deployment lead times extended to mid-2026
NVIDIA's new Blackwell series GPUs are also facing supply shortages, with delivery lead times extended to June-July 2026, and all capacity for August-September already pre-booked. Notably, the more energy-efficient new chips have not driven down prices of the older H100; demand for Hopper architecture GPUs continues to strengthen.
3. China's AI server market landscape reverses, NVIDIA's share drops to 55%
Market data released on April 3, 2026, shows that Chinese chipmakers now hold 41% of the domestic AI server market, while NVIDIA's share in the Chinese market has fallen to 55% (about 2.2 million GPUs), a sharp decline from the claimed 95% before sanctions. Huawei is the leading Chinese chipmaker, having delivered approximately 812,000 AI chips (nearly 20% market share).
4. Domestic GPU makers release 2025 financial results, inference becomes core revenue driver
Four domestic GPU makers – Moore Threads, M-Thread, Biren Technology, and Iluvatar CoreX – all saw significant revenue growth in 2025, but remained collectively loss-making. M-Thread had the highest revenue at RMB 1.644 billion, while Moore Threads posted RMB 1.506 billion. Inference became the core revenue driver – M-Thread's inference GPU board sales surged 866.02% year-on-year, and Iluvatar CoreX's inference product revenue grew 238.2% year-on-year.
5. MIIT launches inclusive computing power initiative, "Computing Bank/Market" elevated to national strategy
On April 2, 2026, the Ministry of Industry and Information Technology (MIIT) issued the "Special Action for Inclusive Computing Power to Empower SME Development," elevating the concepts of Computing Bank and Computing Market from ideas to national strategy. This marks a new phase for China's computing industry: resource pooling, cross-region scheduling, and inclusive trading. Computing power is no longer a "heavy-asset purchase" but can be stored, accessed, and paid for on demand like water and electricity.
6. First university-scale GPGPU thousand-cluster to land at Tongji University
A GPGPU thousand-card computing cluster, jointly built by Hygon and Tongji University, will go online in May. It will target high-performance computing scenarios such as AI4S (AI for Science) and AI4E (AI for Engineering). Hygon has proposed a solution for university AI computing infrastructure featuring "elastic computing power, hardware-level security, and true compatibility."
7. NVIDIA sets new MLPerf inference benchmarks, software ecosystem remains core moat
NVIDIA set multiple records in the latest MLPerf inference benchmarks, emphasizing that the company has moved beyond being just a GPU chip vendor to a platform strategy driven by software innovations including CUDA, NVLink interconnect, and the Dynamo inference framework. For fiscal year 2026, NVIDIA reported full-year revenue of US$215.9 billion, up 6.5% year-on-year, with data center revenue reaching US$193.7 billion.
8. South Korea's FuriosaAI launches second-gen AI inference chip Renegade, focusing on high energy efficiency
South Korean AI chip company FuriosaAI launched its second-generation AI inference chip, Renegade (RNGD). Built on TSMC's 5nm process, the chip has a TDP of about 180 watts. It can reduce total data center cost of ownership by about 40% compared to NVIDIA's RTX Pro 6000, while handling 7.4 times more concurrent users. The company has mass-produced about 4,000 chips this year, and Samsung SDS plans to launch an NPU subscription service based on the chip in July.
9. Parallel Tech plans to purchase over RMB 148 million in computing infrastructure
On April 3, Parallel Tech announced plans to purchase up to RMB 102 million in network equipment from Tsinghua Unisplendour Software and up to RMB 46.24 million in GPU servers from Beijing Youwei Xintong. This is a typical case of a domestic computing service provider accelerating computing infrastructure deployment amid tight GPU supply.
10. Decentralized computing triggers institutional reaction on Nasdaq for the first time
Nasdaq-listed Axe Compute (AGPU) held its annual earnings call, becoming the first public company to disclose complete financial data to institutional investors based on decentralized GPU infrastructure, with its underlying computing power coming from Aethir's decentralized GPU network. Following the earnings release, AGPU's stock rose as much as 152% in a single day, with daily trading volume exceeding US$300 million, marking the first time decentralized computing has triggered an institutional-level reaction on Nasdaq.