Global GPU Computing News【20260823】

China approves small‑batch shipments of NVIDIA H200 chips. On August 19, the Financial Times reported that ByteDance and Tencent recently received about 10,000 H200 processors each, with several other Chinese tech companies expected to obtain similar supplies. The U.S. has authorised each Chinese firm to purchase up to 100,000 H200 chips, but actual releases are only about 10% of that quota. The news triggered a sharp drop in China’s A‑share chip sector.

Alibaba’s T‑Head accelerates commercialisation of the Zhenwu M890 GPU. In its Q1 FY2027 earnings report, Alibaba disclosed that the Zhenwu M890 uses a self‑developed architecture, integrates 144 GB of HBM, offers 800 GB/s interconnect bandwidth, and delivers a three‑fold performance improvement over its predecessor. Supernode instances based on the M890 are now officially sold on Alibaba Cloud, serving over 650 external customers as of early August.

Domestic GPU vendors leverage cluster scale to overcome process constraints. Huawei, Moore Threads, and other local chipmakers are pooling multiple GPUs into large‑scale networks to unlock overall computing power. At the 2026 World Robot Conference, Moore Threads’ founder Zhang Jianzhong stressed the importance of autonomous computing infrastructure for embodied AI, noting that domestic AI clusters are evolving from thousand‑card to ten‑thousand‑card and even hundred‑thousand‑card scales.

Sugon showcases its supernode at the Anhui Computing Power Conference. On August 21, Sugon demonstrated its cable‑less box‑type supernode scaleX40, which integrates 40 GPUs per node, delivers over 28 PFLOPS of total computing power, and uses orthogonal cable‑less interconnection. The design shortens deployment time from months to hours.

Shanghai unveils new policies for intelligent computing. The city’s “15th Five‑Year Plan” for software and information services calls for breakthroughs in ultra‑large AI cluster networking technologies, strengthens domestic supply in core areas such as GPU/NPU, QPU, CPO, and HBM, and promotes deep integration of indigenous chips with large AI models.

NVIDIA’s Vera Rubin racks go live at OpenAI. OpenAI’s infrastructure lead announced that the first Vera Rubin NVL72 racks are operational and running the company’s training stack, supporting next‑generation frontier AI pre‑training. Each rack sells for up to $8.8 million, significantly higher than the ~$5 million for the previous GB series, and the platform is now deployed in over 350 factory nodes across 30 countries.

**NVIDIA guarantees up to $105 billion for OpenAI’s Ohio data centre.** In an SEC filing, NVIDIA disclosed it could provide up to $105 billion in credit support and residual value guarantees for OpenAI’s 8‑gigawatt AI data centre campus in Ohio, with an initial capacity of 4.25 GW. The facility, developed by SB Energy, is expected to enter operation in 2028.

**NVIDIA and six Wall Street firms plan a $500+ billion computing financing platform.** NVIDIA signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilise over $500 billion in third‑party capital for AI infrastructure. CEO Jensen Huang noted that GPUs have become income‑generating productive assets. More details are expected in NVIDIA’s August 26 earnings report.

NVIDIA acts as a matchmaker for AI computing in the Nordic region. NVIDIA is connecting GPU‑holding enterprises with Nordic data centre operators that have suitable deployment conditions. Leveraging abundant power, land, and naturally cool climates, the Nordics are becoming a hotspot for global AI data centres. Pure DC announced a €1.5 billion investment in a Finnish data centre campus.

Anthropic hires a Google TPU veteran to boost its own AI chip efforts. Anthropic has brought on Amir Salek, a co‑founder of Google’s TPU project, to lay the groundwork for its in‑house AI semiconductor business. This move signals a shift from sole reliance on NVIDIA GPUs toward a heterogeneous computing mix of NVIDIA, AMD, and custom AI ASICs.