Global GPU Computing News【20260426】

1. Cloud providers prioritize large customers, squeezing small and medium AI firms

The surge in AI applications, especially AI coding tools and agents, has led Microsoft, Amazon, and other cloud providers to prioritize GPU allocation for top-tier clients such as Anthropic and OpenAI. Small and medium startups are facing soaring prices, months-long queues, and are even required to purchase thousands of Blackwell chips under at least one-year contracts to secure capacity. Some are forced to buy their own NVIDIA GPUs and host them in colocation facilities to avoid uncertainty.

2. Global GPU rental prices surge, driving up computing leasing stocks

Due to supply-demand imbalance, spot rental prices for all NVIDIA GPU models have risen sharply. According to the Ornn Compute price index, Blackwell series chips now rent for $4.08 per hour, up 48% from two months ago. One-year lease contracts for H100 have also jumped from roughly $1.70 per GPU per hour in October 2025 to $2.35 per GPU per hour in March 2026, a nearly 40% increase. This has propelled China's A-share computing leasing sector, with multiple stocks hitting their daily upper limits.

3. Google Cloud and NVIDIA launch A5X bare-metal solution, scaling up to 960,000 GPUs

Google Cloud and NVIDIA announced a new A5X bare-metal solution based on NVIDIA’s latest Vera Rubin architecture and Google’s self-developed Virgo network. A single data center can connect up to ~80,000 GPUs, and by deploying across multiple regions, the total scale can reach 960,000 GPUs. The new system also significantly improves inference efficiency, reducing inference cost per token by up to 10x.

4. China's central authorities issue document requiring higher per-unit computing power efficiency

The General Office of the Communist Party of China Central Committee and the General Office of the State Council issued opinions on energy conservation and carbon reduction, explicitly calling for the transformation of computing infrastructure to continuously improve per-unit computing power efficiency, strictly enforce PUE and other energy efficiency standards, and support the development of green, low-carbon, intensive and circular computing facilities.

5. Shanghai deploys high-performance domestic AI computing clusters

The Shanghai municipal government released the implementation plan for the National Digital Economy Innovation and Development Pilot Zone (Shanghai), which requires upgrading and expanding AI computing infrastructure, deploying high-performance domestic AI computing clusters, promoting interconnectivity of computing resources, and accelerating pilot applications and productization of brain-computer interfaces, 6G, quantum computing, and other technologies.

6. Chinese inference GPU unicorn Sunrise raises over 1 billion RMB

On April 20, Sunrise, a Chinese full-stack self‑developed AI inference GPU company, announced the completion of a new funding round of over 1 billion RMB, with participation from multiple industrial strategic investors, local state-owned capital, and leading financial institutions, achieving a valuation exceeding 10 billion RMB and becoming the first unicorn in the pure inference GPU segment in China. Since its founding just over a year ago, Sunrise has completed seven funding rounds totaling about 4 billion RMB. This round will be used for mass production and delivery of the next-generation S3 inference GPU and R&D of subsequent S4/S5 chips.

7. NVIDIA RTX PRO 4500 server card officially goes on sale

The NVIDIA RTX PRO 4500 Blackwell server edition officially went on sale on April 24, with a starting price of approximately 29,400 RMB. It features 10,496 CUDA cores and 32GB GDDR7 ECC memory, with a power consumption of only 165W and a single-slot passive cooling design. This mid-range professional card based on the Blackwell architecture is now available for individual purchase by a wide range of enterprise users, no longer confined to OEM systems.

8. Bolt Graphics announces self‑developed Zeus GPU, taped out on 12nm

On April 23, Bolt Graphics announced that its self‑developed Zeus GPU architecture has completed tape-out, using TSMC’s 12nm FFC process after four years of development. The dual‑chip version, Zeus 2c26, delivers 5x the performance of a competing flagship GPU (575W) in path tracing tasks at just 250W. In electromagnetic simulation workloads, the quad‑chip solution achieves up to 300x the performance of a single high‑end GPU. Mass production is planned for the end of 2027.

9. Intel launches Xeon 600 workstation processors and Arc Pro B70/B65 GPUs

On April 23 in Beijing, Intel held a next‑generation AI workstation platform launch event, introducing the Xeon 600 workstation processors and Arc Pro B70 and B65 GPUs. Using a “dual‑chip synergy” approach, the platform provides a unified computing foundation ranging from lightweight applications to heavy AI workloads. The Xeon 600 delivers up to 61% higher multi‑threaded performance compared to the previous generation, while the Arc Pro B70 features 32GB of memory and up to 367 TOPS of AI compute.

10. MIIT: Guide orderly computing construction, promote computing‑power collaboration and space‑based computing development

At a State Council Information Office press conference on April 21, China’s Ministry of Industry and Information Technology (MIIT) said it will guide the orderly and demand‑driven construction of computing infrastructure, promote the coordinated deployment of green electricity and computing power, improve China’s computing platform to match supply with demand, and announced support for forward‑looking research on space‑based computing technology while promoting the orderly development of the space computing industry. The MIIT has also issued policy documents to empower small and medium enterprises with inclusive computing power services.