Global GPU Computing News【20260913】

JD Cloud plans to build a 100,000-card domestic GPU intelligent computing cluster.

On September 9, at the 2026 Global Technology Explorer Conference, JD Cloud announced that it would use Moore Threads’ full-function GPUs as the compute foundation to build a 100,000-card-level intelligent computing cluster, focusing on large-model training, inference, and embodied intelligence, and opening the compute power to the entire industry. This is the first time domestic GPUs have entered the core intelligent computing cluster of a leading AI cloud company at the 100,000-card level, marking that domestic intelligent computing is moving from “usable” to “large-scale commercial deployment.” The two sides will carry out full-stack collaboration from chips and cloud platforms to model training.

Nvidia restarts its Rubin CPX chip plan.

According to an industry survey, Nvidia has restarted the Rubin CPX chip designed specifically for long-context AI inference, with production expected to begin early next year. The restarted version has several specification changes: GPU compute is close to Rubin, and power consumption is also 2,300 watts, but memory has been upgraded from 128GB GDDR7 in the old version to 168GB HBM4. It uses an independent MGX ETL rack, and customers can configure 64 to 256 CPX units as needed. In operation, it is paired with Vera Rubin NVL72, with a recommended ratio of 1 to 1. CPX handles prefill and builds the KV cache.

Nvidia’s Vera Rubin enters full production and adopts a full liquid-cooling solution.

Nvidia’s next-generation Vera Rubin architecture GPU has entered full production and has officially been included in a full liquid-cooling reference design. The single-chip thermal design power is as high as 2,300 watts, and the power of a single rack far exceeds the limits of traditional air cooling. The cooling demand boundary has extended from GPU and CPU to the entire chain including switches and optical module cages. Liquid cooling is upgrading from a compute accessory to a rigid infrastructure necessity for AI data centers. Some data shows that the global AI server liquid-cooling system market is expected to surge from USD 8.9 billion in 2025 to more than USD 17 billion in 2026.

HBM shortages cause Huawei and Cambricon to significantly raise AI chip prices.

According to Reuters, export controls have caused high-bandwidth memory costs to soar. Huawei has raised the quoted price of its Ascend 950DT AI chip to more than RMB 250,000, with an increase of up to 50% within two months. Cambricon has repriced its next-generation chip “690,” raising it by 20% to 30% compared with two months ago. Smaller manufacturers such as MetaX and Iluvatar CoreX have also made similar price increases. The advanced HBM market is dominated by SK hynix, Samsung, and Micron, and Chinese companies are increasingly relying on the gray market for supply, with procurement prices often several times those of overseas buyers.

Samsung and Micron both expand HBM4 production.

Samsung Electronics plans to significantly increase HBM4 supply starting in the third quarter. As of last month, more than 50% of its total 4nm capacity had been allocated to HBM4 chips. Micron plans to increase equipment investment by the end of the year, targeting up to 60,000 additional wafers per month in capacity, raising HBM monthly output from 40,000 to 50,000 wafers last year to about 100,000 wafers, doubling the scale, with a focus on expanding 12-layer HBM4 products. The expansion by both manufacturers is intended to narrow the gap with SK hynix and seize the AI high-bandwidth memory market.

Amazon adds 2 million Nvidia GPUs.

Amazon announced the additional deployment of 2 million Nvidia GPUs in its AWS infrastructure, bringing its cumulative purchase commitment to more than 3 million units. This move reflects that AI compute demand is undergoing a structural shift from training to inference. The procurement logic of cloud service providers is being rewritten. Even in the face of enormous depreciation pressure, they still must maintain an overwhelming supply of compute power.

Microsoft plans to triple its data center capacity.

According to Bloomberg, because strong AI and cloud computing demand has strained existing compute capacity and even forced it to turn down some business, Microsoft plans to increase its data center capacity from about 12 gigawatts currently to more than 38 gigawatts by 2032. Of the current roughly 12 gigawatts of capacity, only about 2 gigawatts is specifically used for AI chips. It is expected that by 2032, AI-specific compute will account for about one-third of total capacity.

Nvidia invests USD 3.5 billion in MediaTek to jointly build a next-generation AI computing platform.

On August 31, Nvidia and MediaTek announced a deepening of their long-term strategic cooperation. Nvidia will spend USD 3.5 billion to subscribe to convertible bonds issued by MediaTek, and the two sides will jointly build a next-generation AI computing platform from the edge to the cloud. The cooperation focuses on AI infrastructure, local AI computing, and the automotive field. MediaTek will connect to Nvidia’s NVLink Fusion ecosystem, and the two sides will also jointly iterate RTX Spark and DGX-Spark PC chips. This investment may become Nvidia’s largest direct investment outside the United States.

The “four little dragons” of domestic GPUs gather in the capital market, with combined first-half revenue exceeding RMB 5.2 billion.

With Enflame Technology officially listed on the STAR Market, the four domestic computing chip companies—Moore Threads, MetaX, Biren Technology, and Enflame Technology—have all gathered in the capital market. The four companies’ combined first-half revenue exceeded RMB 5.2 billion. Biren Technology’s revenue grew nearly 20 times year-on-year, Moore Threads’ first-half revenue surpassed its entire revenue last year, and MetaX achieved its first half-year book profit. However, profit paths show significant divergence. MetaX still recorded a loss after deducting non-recurring items, and Iluvatar CoreX’s profit mainly came from fair value gains on equity. The industry as a whole presents a complex picture in which “revenue growth without profit growth” and “turning losses into profits” coexist.

The wave of compute price increases continues, with some cloud services rising more than 30%.

Since the beginning of this year, Tencent Cloud, Alibaba Cloud, Baidu AI Cloud, and other manufacturers have raised prices for compute products, with some services increasing by more than 30%. As of March this year, China’s average daily Token calls exceeded 140 trillion, more than 1,000 times higher than in early 2024. In terms of demand structure, inference data volume surpassed training data volume for the first time. In the future, the ratio of inference compute demand to training compute demand may reach 3 to 1 or even higher. The industry judges that this round of price increases is not a short-term fluctuation, but a landmark node in the AI industry’s shift from “subsidy expansion” to “commercial sustainability.”