Huawei's next-gen Ascend NPUs could become China's best option

Huawei unveiled its Ascend 960DT AI accelerator, now targeted for Q1 2027—three quarters ahead of schedule—with up to 288GB of memory and 4 petaFLOPS of FP4 performance, doubling the performance and memory capacity of its Ascend 950-series. While the chip remains materially behind Nvidia Rubin and AMD MI455X, it offers Chinese developers a domestically available alternative to export-restricted leading U.S. GPUs and can scale to 4,096 chips for up to 16 exaFLOPS of FP4 compute. Huawei also outlined a 960PR inference-focused chip for Q3 2027 and longer-term Ascend roadmaps targeting 28 petaFLOPS FP4, 384GB memory and 38.4TB/s bandwidth by 2029.

Analysis

The investable implication is less a near-term revenue hit to NVDA than erosion of its China re-entry option value. Export controls have already capped the addressable premium-GPU market, but a credible domestic rack-scale alternative can make that demand structurally less recoverable even if policy loosens. Over 6-18 months, Chinese model builders that port workloads, tooling, and operations onto Ascend reduce switching back to CUDA; this is a platform-lock-in risk, not merely a chip-performance comparison.

The central unpriced risk is execution. Huawei's claimed cluster reliability and theoretical large-scale configurations must be validated in sustained commercial training runs, where compiler maturity, collective-communications efficiency, yields, memory supply, and serviceability matter more than peak specifications. A repeat of prior customer migration failures would preserve NVDA's software moat and could make the announced timetable a sentiment event rather than an earnings event; watch disclosed deployments by major Chinese cloud providers and model labs over the next 1-3 quarters.

AMD is the more ambiguous read-through. It lacks NVDA's installed software advantage but also has less China-specific scarcity premium embedded in its AI valuation; a domestic platform that commoditizes China inference could pressure both vendors' long-run China TAM while reinforcing the premium placed on NVDA's full-stack systems outside China. GOOG is marginally validated strategically: scale-out, heterogeneous accelerator architectures favor operators with proprietary silicon and vertically integrated workloads, though there is no direct near-term earnings transmission.

Consensus may overreact to headline compute metrics while underweighting the policy feedback loop. More capable domestic supply could make further US restrictions politically easier, reducing NVDA's China upside optionality, but it may simultaneously intensify Chinese demand for sanctioned-adjacent components and domestic networking infrastructure; neither effect is reliably monetizable through the listed tickers today.

Key Decisions for Investors

• Maintain NVDA core exposure but buy a 6-9 month, 10-15% out-of-the-money put spread sized to premium-at-risk only; the catalyst window is customer deployment evidence and any export-control tightening. Exit if NVDA demonstrates stable China-adjacent revenue despite restrictions and management raises data-center guidance without incremental gross-margin pressure.
• Do not initiate a standalone AMD short on this announcement. Reassess after AMD reports AI accelerator backlog and geographic mix: a material reduction in China-related demand commentary or lower-than-expected MI-series ramp would support a 3-6 month AMD underweight versus NVDA.
• Use a small long NVDA / short AMD relative-value position only on a broad semiconductor selloff, not immediately: NVDA's non-China software and systems moat should be more defensible if domestic Chinese alternatives fragment the accelerator market. Falsify if AMD closes the software/ecosystem gap through material hyperscaler wins or NVDA's gross margin falls on China-compliant product mix.
• Set an alert for independently verified Ascend production deployments at Chinese hyperscalers or leading model labs, accompanied by training-cost and uptime disclosures. Confirmation would justify increasing the NVDA hedge; roadmap claims without workload-level validation are not sufficient for a fundamental position change.