DeepSeek's 160,000-chip Ascend order outgrows Huawei's largest advertised cluster

Somebody at a Chinese AI company has to fill in the 2027 inference line of a capacity plan this quarter, and DeepSeek's Inner Mongolia order is the first public number large enough to plan against. Huawei says its SuperPod structure will soon link as many as 100,000 Ascend chips into a single cluster, with faster data transmission over its UnifiedBus protocol ^[8]. At least 160,000 chips is 1.6 times that ceiling ^[9], so the Inner Mongolia site needs more than one maximum-size cluster. Wang did not say how many chips a single supernode holds ^[6]. (The Straits Times names the rotating chairman presenting the parts as Wang Tao; Digital Trends, citing Reuters, calls him David Wang ^[1][2].)

Huawei raised pricing on the Ascend 950DT by 60 percent, citing tight component supply in communications with customers ^[4]. That says more about availability than the roadmap does. Nvidia's most advanced Blackwell processors are not directly available in China ^[10], and the top Nvidia silicon that does get in arrives through smuggling, in modest quantities ^[11]. The Straits Times did not report a unit price. Morgan Stanley puts the Chinese AI compute market at as much as 646 billion yuan by 2030, and a production capacity constraint is what stops Huawei from serving the demand it already has ^[12].

What Huawei is selling and what its largest domestic customer is doing are two different things. Frontier Chinese labs including DeepSeek still largely train on Nvidia chips and then operate the finished models on domestic semiconductors ^[7], and the 950DT is the memory-heavy inference member of the family ^[23]. Guo Ping, chairman of Huawei's supervisory board, told new hires at an internal meeting in September that the goal is for Ascend to run every single AI model in China and beyond ^[15]. "We are closing the gap by innovating chip architectures," Guo said ^[13]. "Apple was not a leader in making processors at the beginning, but it delivered exceptional user experience through architecture redesigns and software optimisation," he said ^[14].

"System-level competitiveness matters more than ever," Morgan Stanley analyst Charlie Chan wrote in a research note ^[16]. Chan wrote that "the effective gap is narrowing through multi-die design, advanced packaging, rack-scale system architecture, optical networking and software-hardware co-optimisation" ^[17]. Huawei's own count of its developer base is thinner: 5,270 monthly active developers on its AI chip ecosystem ^[5]. Huawei says it has delivered over 1,000 supernodes to more than 370 customers, about 2.7 each ^[6][24].

The Ascend 950 chips are the first Huawei accelerators powered by memory the company designs in-house, and Huawei has provided little production detail ^[18]. Huawei and other Chinese AI chipmakers are cut off from the latest AI memory products from SK Hynix, Samsung and Micron under US export controls ^[19].

For most teams it comes down to whether the workload trains or serves, and whether the unit economics can absorb a 60 percent step in part cost with no second supplier. For serving inside China, Ascend is a procurement path now, and the cost of using it is a single vendor whose memory line is internal and whose parts are rationed by capacity. For training at frontier scale the record is not there yet, and the Shanghai summit left open whether the 960 is powerful enough for frontier training ^[22]. Huawei says each new generation doubles the computing power of its predecessor. On the company's own numbers that puts the 2028 Ascend 970 at four times the 950 ^[20][21].