Global GPU Computing News【20260906】

1. NVIDIA beats earnings guidance, AI computing demand remains strong

NVIDIA reported fiscal 2027 Q2 results with data center revenue reaching $89 billion, up 117% year‑over‑year. Q3 revenue guidance implies 10%–14.5% sequential growth, with a midpoint of $108 billion exceeding market expectations. The company projects fiscal 2028 revenue growth of about 70%, significantly higher than the prior consensus of roughly 45%. The Vera Rubin chip has entered full production and begun shipping, expected to contribute about 20% of data center revenue in Q3.

2. Anthropic signs $35 billion cloud computing deal with Lambda

AI company Anthropic signed a $35 billion cloud computing agreement with GPU cloud provider Lambda, with the underlying data center located in Texas, USA. This large order has sparked a new wave of AI server demand, bringing opportunities to supply chain players such as Hon Hai, Wistron, and Quanta.

3. Nscale and Figure expand strategic partnership to over $6 billion

AI computing firm Nscale and Figure signed a multi‑year strategic partnership expanded to over $6 billion, representing the potential to deploy up to 100,000 NVIDIA Vera Rubin GPUs.

4. NVIDIA, Equinix, and Together AI launch enterprise AI inference solution

NVIDIA, together with global data center colocation leader Equinix and AI inference platform Together AI, launched a distributed AI inference service called “Equinix Inference Exchange.” This marks NVIDIA’s expansion of its computing ecosystem from training to inference, with inference revenue now exceeding training and accounting for over 55% of total.

5. GPT‑6 Astra released, trained on over 100,000 GPUs

On September 4 at 00:00 UTC+8, OpenAI released its new flagship model GPT‑6 Astra, which was pre‑trained using over 100,000 GPUs at the Stargate data center in Texas. On the same day, China’s A‑share AI hardware sector tumbled, with Inspur hitting the limit‑down and three large‑cap hardware companies losing a combined over RMB 29 billion in market value.

6. All four major domestic GPU “dragons” complete capital market listings

On September 2, Enflame Technology started its online and offline subscription. With this, the four domestic GPU chip companies – Moore Threads, MetaX, Biren, and Enflame – have all completed their capital market debuts, marking the initial technical validation and capitalization of China’s high‑end GPU industry.

7. DeepSeek plans to purchase 160,000 Huawei Ascend 950DT chips

According to Bloomberg, DeepSeek plans to deploy at least 160,000 units of Huawei’s next‑generation Ascend 950DT chips at a large data center in Inner Mongolia for model inference. If completed as planned, it would become one of the largest known domestic AI chip clusters. The chip is reportedly equipped with 144 GB of memory and 4 TB/s memory bandwidth, with an expected launch in Q4 2026.

8. AMD launches RX 9060 XT graphics card and Threadripper Halo Station workstation

On September 2, AMD released the Radeon RX 9060 XT graphics card, featuring 32 RDNA 4 compute units and delivering 821 TOPS of AI compute. Additionally, at IFA 2026, AMD showcased the Threadripper Halo Station AI workstation, equipped with a 96‑core Zen5 processor and dual Instinct MI350P accelerators (totaling 288 GB HBM3E), positioned as a desktop‑class workstation for trillion‑parameter MoE model inference.

9. Seven‑ministry directive requires domestic chip procurement to be no less than 75% in new AI data centers

The “2026–2028 Hard Technology Synergy Development Plan” jointly issued by seven ministries including the Ministry of Industry and Information Technology mandates that, effective September 2026, state‑owned enterprises, government agencies, and central state‑owned enterprises must ensure that domestic hardware procurement accounts for no less than 75% of total AI computing infrastructure in new builds. Procurement of foreign high‑end GPUs is subject to a quota system, and in some regions, import permits require purchasing a corresponding proportion of domestic chips.

10. U.S. reportedly drafting new rules to block China’s overseas AI compute access

According to media reports, the Trump administration is drafting a new export control rule on artificial intelligence that would restrict Chinese companies from remotely accessing advanced AI servers deployed in third‑country locations such as Thailand and Singapore. The U.S. Department of Commerce could solicit industry feedback as early as September. The proposed rule aims to extend export controls from the physical transfer of chips to the right to use computing power.