An article from the Huxiu business outlet argues that Chinese large models will keep running even if top chips become unobtainable. Domestic chips already carry real workloads and is capable to support the running of large models. Zhipu operates a 100,000-card inference cluster that handles all live traffic for its GLM-5.3-Flash model, and its anonymous Ox-Alpha model has processed 62 trillion tokens on overseas platforms. Meituan trained LongCat-2.0 on domestic hardware. Efficiency lags, though: Huawei’s Ascend 910C reaches roughly 30-35 percent training utilization versus 40-50 percent for Nvidia H100 clusters. Since cost per token depends on chips, software, and model design, firms are optimizing the latter two.
Four firms are taking different paths. Zhipu is going asset-heavy, acquiring compiler firm Zhongke Jiajia and raising $5 billion in September. Management says 30 billion yuan ($4.2 billion) of compute, split 40 percent training and 60 percent inference, could generate up to 40 billion yuan ($5.6 billion) in annual revenue, a theoretical estimate. MiniMax, whose cloud costs exceeded 90 percent of cost of sales, is moving from renting to building domestic clusters that are not yet carrying live traffic. DeepSeek bets on model efficiency, yet raised V4-Pro prices in August due to chip limits; its 160,000-card cluster is reportedly not due until late 2027 or early 2028, and prices may fall once Huawei’s Ascend 950 ships in volume. Kimi uses open-source models and ecosystem partnerships to shorten the adaptation cycle for domestic chips; its models were quickly adapted by Huawei, Alibaba’s T-Head, Hygon, Moore Threads and Biren, but demand after K3’s launch forced it to pause new consumer sign-ups.
Adaptation solves “can run”; owned compute solves “can afford.” The author urges tracking capacity, efficiency, and cost per token, ranking Zhipu the surest short-term bet, DeepSeek the highest-upside medium-term one, and MiniMax and Kimi as unproven.
Source: Huxiu, September 25, 2026
https://www.huxiu.com/article/4893261.html