Global GPU Computing News【20260614】

1. NVIDIA RTX Spark processor officially unveiled, entering the PC market

At Computex 2026 in Taipei, NVIDIA and MediaTek jointly launched the new ARM-based PC processor RTX Spark. Built on TSMC’s 3nm process, the chip integrates a 20-core Grace CPU and a Blackwell GPU with 6,144 CUDA cores, delivering 1 PetaFLOP of AI compute power and supporting up to 128GB of unified memory, capable of running 120-billion-parameter LLMs locally. The first wave of thin-and-light laptops featuring this chip will be launched this fall by Lenovo, Dell, HP, Microsoft, and others, marking NVIDIA’s official entry into the PC processor market.

2. NVIDIA Vera Rubin AI chips enter mass production, adopted by multiple leading customers

NVIDIA announced that its next-generation AI superchip platform Vera Rubin has entered full mass production, with OpenAI, Anthropic, and SpaceX as the first customers for the Vera chip. The Vera CPU is NVIDIA’s first general-purpose processor purpose-built for AI agents, delivering 1.8x the performance of traditional x86 processors in AI agent tasks. Hyperscalers including ByteDance, CoreWeave, and Oracle Cloud have also planned to adopt Vera to upgrade their AI factories.

3. Intel launches Crescent Island data center GPU with up to 480GB of memory

As NVIDIA moves into the PC market, Intel is countering in the AI data center space. At Computex 2026, Intel officially unveiled the Crescent Island GPU platform for AI data centers. Based on the Xe3P architecture, it features up to 480GB of LPDDR5x memory, air‑cooling design (350W TDP), and focuses on energy efficiency in AI inference. It supports multiple data types from FP4 to FP64, with a planned release in the second half of 2026.

4. U.S. tightens AI chip export controls, bans shipments to overseas subsidiaries of Chinese entities

On May 31 local time, the U.S. Bureau of Industry and Security issued new guidance clarifying that export license requirements for advanced computing chips apply to entities headquartered in China (including Macau) and their overseas subsidiaries. This prohibits Chinese-backed companies from using overseas branches to bypass restrictions on advanced AI chips such as NVIDIA Blackwell and AMD MI350x. NVIDIA’s Rubin processor is also covered, closing a previous regulatory loophole.

5. SpaceX and Google sign $30 billion computing agreement

Bloomberg reported that SpaceX and Google have signed a computing agreement worth approximately $30 billion. Google will pay about $920 million per month from October 2026 through June 2029, obtaining the right to use 110,000 NVIDIA GPU chips along with CPUs, memory, and other components. Based on H200 capacity, this represents over 100 megawatts of computing power. Google sees the agreement as a transitional arrangement to meet surging AI service demand.

6. China plans to invest ¥2 trillion to build nationwide data center network

Reports indicate that the Chinese government plans to invest about ¥2 trillion (RMB) over the next five years to build a nationwide data center network, aiming to boost AI development and surpass the U.S. in key technology areas. Drafted by the NDRC and other agencies, the plan envisions an interconnected computing hub network, requiring local suppliers such as Huawei to provide at least 80% of AI chips. State-owned enterprises China Mobile and China Telecom will be responsible for operations. The goal is to connect scattered data facilities into a unified network by 2028.

7. Wuxi domestic AI computing center Phase I goes live, powered by Moore Threads

On June 11, Phase I of the Wuxi (Huishan) domestic AI computing center was officially lit up, with computing power provided by Moore Threads. It is an important node of the southern Jiangsu computing cluster of the “East Data West Computing” Yangtze River Delta hub. Moore Threads founder Zhang Jianzhong also proposed three fully domestic “AI factories”, building a full‑chain solution from computing infrastructure to embodied AI applications.

8. Computing power rental market sees extreme divergence in supply and demand

H200 (141GB HBM3e) remains in short supply, with OpenAI, Meta, and Google queuing for purchases. Lead times for new orders have stretched to 12–18 months, with some pushed to Q2 2027. In the first half of the year, monthly rental prices per H200 card rose from about 55,000 RMB to 85,000–90,000 RMB, with some cloud vendors increasing prices by 30%–55%. At the same time, the B200 suffers from inter‑generational price inversion due to supply constraints, while monthly rental costs for a single B300 server exceed 120,000 RMB. In contrast, the “computing supermarket” under Zhejiang Computing Technology said that under token‑based pricing for large models, the current cost is just a few RMB per million tokens and is still falling.

9. Kingsoft Cloud raises prices for AI computing products by 15%–50%

On June 12, Kingsoft Cloud announced price adjustments for certain products due to surging global AI computing demand and rising hardware costs. AI‑computing‑related products will see price increases of approximately 15%–50%, while file storage products will rise 30%–50%. The adjustments take effect on July 12, 2026, and existing orders will not be affected during their current billing cycles.

10. Chinese GPU company Muxi shares plans Hong Kong IPO

On June 12, Muxi (Moshare) held a board meeting approving a proposal to issue H shares and list on the main board of the Hong Kong Stock Exchange. The company plans to issue no more than 5% of its total share capital, with proceeds to be used for next‑generation general‑purpose GPU R&D and commercialization, MXMACA software ecosystem development, industrial chain investment, and more. Muxi was listed on the STAR Market last December and currently has a market value of about 278 billion RMB, becoming another domestic GPU company pursuing an “A+H” dual listing platform.