Invest1 publisher2 min readPublished
Huawei answers Nvidia's per-chip lead by wiring 4,000 processors into one pod
The Ascend 960 is dated 2027 and the 970 is dated 2028, but the only large order on the record is DeepSeek's 160,000 chips of the current generation. That fleet is forty pods of silicon Huawei ships today.
The Investor · Invest desk
What happened
- Huawei chairman Wang Tao unveiled the Ascend 960 SuperPoD cluster at the company's annual summit in Shanghai on Thursday, with the system scheduled to reach the market in 2027 despite US restrictions.
- DeepSeek is building a data center in Inner Mongolia that will run 160,000 Ascend 950DT processors, the generation immediately below the newly announced 960.
- The roadmap runs past the 960 to an Ascend 970 in 2028, under a three-year pledge to double processor performance with every iteration.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint A 4,000-processor pod moves the purchase decision from the die to the rack, so a Chinese buyer comparing Huawei with Nvidia is comparing systems it can get against systems it cannot.
- exposure Anyone counting on 2027 delivery is relying on an inference: Bernstein treats the publication of the roadmap itself as the confidence signal on local foundry supply.
- decision Chinese labs buying training capacity now are buying the shipping 950 generation, which means the 960 has to win a second procurement decision in 2027 that no named customer has yet made.
- precedent By putting dated parts and a doubling cadence on the record, Huawei sets a public yardstick its 2028 silicon can be measured against.
Divide 160,000 by 4,000 and DeepSeek's order comes to forty pods [14]. The processors in it are Ascend 950DT, the generation Huawei sells now, one below the 960 that chairman Wang Tao put on stage in Shanghai on Thursday [7][1]. Cryptopolitan gives no price and no completion date for the Inner Mongolia site [16].
What Huawei is selling for 2027 is the wiring. An upgraded UnifiedBus interconnect and the Hi-ONE optical systems are what get 4,000 processors into a single SuperPoD [4], and near-packaged optics went into the architecture to handle speed and energy consumption together [3]. Huawei has already conceded the per-die comparison, acknowledging that its hardware falls short of Nvidia's on pure speed and power and saying it will compete on hardware volume, networking and state policy support [9]. Morgan Stanley analyst Charlie Chan said: "System-level competitiveness matters more than ever. The effective gap is narrowing through multi-die design, advanced packaging, rack-scale system architecture, optical networking, and software-hardware co-optimization." [12]
Bernstein's team, led by Qingyuan Lin, wrote that "Huawei's willingness to publicly articulate its AI roadmap represents a strong signal of confidence in the resilience of its future local foundry supply" [11]. Taken literally, the evidence for domestic wafer supply is the act of publishing a schedule. The schedule says the 960DT in the first quarter of 2027, the 960PR in the third [2], and the Ascend 970 in 2028 under a three-year pledge to double processor performance with every iteration [8].
Hold that pledge and the 970 lands at four times the 950 by 2028 [15]. If domestic foundry output caps how many 4,000-die pods can be built, the configuration stays a reference design and Chinese buyers who need capacity keep ordering 950s, because US export controls limit their access to Nvidia's most advanced processors anyway [13]. If Chan is right that system-level competitiveness is what counts, and Huawei's own line that networking offsets individual chip limits holds up in a training run [10], then per-die deficits stop deciding the purchase, and supervisory board chairman Guo Ping's stated goal of Ascend powering every AI model globally becomes a volume question [5]. Guo said: "We expect the company to achieve even better results, providing digital infrastructure solutions to global customers, offering diversified solutions for the development of artificial intelligence, and contributing to the rise of China's electronics industry." [6]
I would put more weight on the first branch, because the only large committed fleet on the record runs on silicon that already exists [7]. What would move me is a second commitment at 950-scale volume, placed against 960-generation pods, before the first-quarter 2027 date arrives [2].
What to watch
- Whether the 960DT actually ships in the first quarter of 2027, and to which named customer.
- Any commitment to 960-generation pods at DeepSeek-scale volume, which would move the roadmap from schedule to order book.
- Any disclosure of domestic wafer output, the input Bernstein is inferring.