Global GPU Computing News【20260712】

1. Meta's "Selling Compute" Plan Triggers Panic, Then Rapidly Reverses

Early this week, Meta was reported to be planning a cloud business to sell its "excess" AI computing capacity, triggering sharp volatility in global capital markets. The Philadelphia Semiconductor Index plunged over 11% in two days, and China's A-share tech sector lost more than one trillion yuan in a single day. However, Meta quickly followed up with positive news: it plans to deploy 7GW of AI infrastructure this year and double that to 14GW next year; its in-house AI chip Iris is about to enter mass production; and it launched its most powerful model with a first‑time fee. Zuckerberg publicly denied any oversupply, saying the compute rental was offered because external bids were higher. Meta's stock rose nearly 15% over the week.

2. Intel Arc Pro B70 Outperforms NVIDIA Flagship in High‑Concurrency AI Inference

On July 9, test data released by GUNNIR showed that the Intel Arc Pro B70 32GB graphics card achieved 7.5%–8.6% higher throughput than the RTX 5090D and 34%–48% higher than the RTX 4090D in high‑concurrency DeepSeek‑R1 inference tests. The card sells for just 9,000–10,000 RMB, while the RTX 5090D costs over 20,000 RMB. Analysts pointed out that the RTX 4090D has only 24GB of VRAM, making it a clear bottleneck under high concurrency.

3. NVIDIA H200 Chips May See Limited Shipments to China

Sources indicate that the Chinese government plans to allow leading AI companies such as Alibaba, ByteDance, and DeepSeek to purchase limited quantities of NVIDIA H200 chips, with the total volume expected to be under 200,000 units – about half of what was requested. The H200 delivers 990 TFLOPS of FP16 compute, roughly twice the performance of the H100. Approved chips are restricted to AI model training, while inference tasks must prioritise domestic processors.

4. NVIDIA RTX 5090 SE Specifications Leaked

On July 10, tech media revealed that NVIDIA is planning to launch the RTX 5090 SE graphics card, featuring 14,080 CUDA cores, a 384‑bit memory bus, and 24GB of GDDR7 VRAM, with performance positioned between the RTX 5090 and RTX 5080.

5. AMD MI355X Inference Cost Half That of NVIDIA B200

When running the GLM 5.2 model, the AMD MI355X achieved a single‑node aggregate throughput of 2,626 tokens per second – about 80% of the performance of NVIDIA's flagship B200, but at less than half the deployment cost. This result was realised through software optimisation rather than specialised hardware.

6. Domestic GPU Maker MuXi MXC600 Enters Mass Shipment, Orders Booked Through 2027

MuXi's flagship MXC600 chip has entered large‑scale mass production and shipment, with some product categories already booked through 2027. The chip features a fully self‑developed architecture for AI training and inference, marking the first full domestic supply chain from design to software stack. MuXi has also launched its next‑generation C700 project, targeting performance close to the H100 level.

7. Huawei Atlas 950 SuperPoD to Debut, Aiming at NVIDIA

Huawei's Atlas 950 SuperPoD will make its public debut at the WAIC conference in Shanghai on July 17 and is scheduled for official release in Q4 2026. The supernode is powered by the昇腾950 chip, delivering 8 EFLOPS of FP8 compute and supporting interconnection of up to 8,192 NPUs. Huaxi Securities estimates its total compute power to be 6.7 times that of NVIDIA's NVL144.

8. China's First Fully Domestic 100,000‑Card AI Supercluster "Sugon 8000" Completed

On July 10, Sugon announced the completion of China's first fully domestic 100,000‑card AI supercluster – the Sugon 8000 (Dengfeng) – which has been simultaneously connected to the National Supercomputing Internet. This marks a leap from 10,000‑card to 100,000‑card scale in AI infrastructure. On the same day, the core node of the National Supercomputing Internet went live in Zhengzhou.

9. DeepSeek and Zhipu Reportedly Developing In‑House AI Chips

On July 7, Reuters cited sources saying that DeepSeek has quietly been working on an in‑house AI chip project for about a year, focusing on inference chips. Meanwhile, Zhipu is also reportedly seeking domestic foundry partners to advance its own chip development. Both companies aim to reduce reliance on external suppliers like NVIDIA.

10. GPU Rental Prices Continue to Rise; Kingsoft Cloud Raises Capital Expenditure

J.P. Morgan data shows that GPU rental prices in the non‑hyperscaler market continued to rise in June, with the average price for H100 increasing 3.7% month‑on‑month – the seventh consecutive monthly increase. AWS EC2 Capacity Blocks reservation prices rose about 20% from July. Kingsoft Cloud, driven by Xiaomi's billion‑yuan‑level compute procurement demand, raised its 2026 capital expenditure to 15 billion RMB.