Answers · Views and frameworks
Can Huawei's Ascend 950 carry China's AI compute demand, and what does DeepSeek training V5 on it mean?
FutureX's read: Ascend has moved from usable to being bet on by top model companies; the real constraint is not chip design but HBM supply and output.
Per TrendForce in early November, DeepSeek reportedly plans to train its next-generation V5 model on Ascend 950, targeting 2027, which would move domestic chips from inference only to training. On November 18-19 the first data center cluster built around Ascend 950DT to export tokens overseas began operating. Huawei had said Ascend supernode deployments exceed 1,000.
Sources: TrendForce (Nov 2026, reported); South China Morning Post (Nov 18, 2026), DigiTimes (Nov 19, 2026); Pandaily (Nov 2, 2026, over 1,000 supernodes).
The constraint is memory. Global HBM is booked by Samsung, SK hynix and Micron, and Chinese makers pay more for high-end HBM. Per a Reuters exclusive in September, Chinese AI chipmakers have raised prices as the HBM shortage bites. FutureX's view: what is scarcest in the Ascend ecosystem is software and toolchains that port models efficiently, plus memory-saving inference techniques; these are more certain bets than another chip design company.
Sources: Reuters (Sep 2026, Chinese AI chipmakers raise prices on HBM shortage); Yahoo Finance (Oct 28, 2026, SK hynix sees shortage until 2030).
Report · China's Domestic AI Chips Hit 52.3%; the Constraint Moves From H200 to HBM →Thesis · Open Source Resets the Cost Curve →
Related answers
FutureX core theses
All theses →Updated 2026-11-19. Facts and examples come from FutureX Capital's published research, public talks, and public reporting, graded per our site-wide standard (verified / reported / per company disclosure). Nothing here is investment advice or an offer.