Executive Summary

  • Huawei is aggressively ramping up its Ascend AI chip production, utilizing “die banks” and continued TSMC production involvement despite US restrictions. High Bandwidth Memory (HBM) remains the primary bottleneck, as China struggles with CoWoS packaging capacity and the Foreign Direct Product Rule’s impact on supply.

Strategic Deep-Dive

Source: SemiAnalysis

Date: 2026-04-23

URL: https://semianalysis.com/2025/09/08/huawei-ascend-production-ramp/

The global landscape of artificial intelligence is increasingly defined by the control of “compute.” As the adage goes, “He who controls the spice, controls the universe”—in the modern era, compute is that spice, and its control dictates the production of AI tokens and the resulting economic benefits. The United States technology community is all-in on compute as the next fundamental platform. Simultaneously, Huawei is leading China’s counter-offensive by ramping up the production of its Ascend series of AI accelerators, aiming to secure a “seat at the table” in the global AI race.

The ramp-up of Huawei Ascend production is characterized by complex geopolitical maneuvering and supply chain ingenuity. Central to this effort is the use of “die banks.” This strategy involves manufacturing and storing massive quantities of semi-finished silicon dies. By doing so, Huawei creates a buffer against potential tightening of the “Foreign Direct Product Rule” or sudden equipment failures in domestic foundries.

These die banks allow Huawei to continue shipping accelerators even if their access to advanced lithography is temporarily severed. Furthermore, despite heavy restrictions, there are indications of continued production involvement through complex channels associated with TSMC. This persistence underscores the reality that advanced semiconductor manufacturing is nearly impossible to fully decouple from global leaders.

However, the primary hurdle facing Huawei is not just logic die production, but the integration of High Bandwidth Memory (HBM). HBM has emerged as the definitive bottleneck in the AI hardware supply chain. Modern accelerators require massive data throughput to keep their processing units fed; without HBM, the raw FLOPS of the Ascend 910 series cannot be effectively utilized.

The challenge is two-fold: first, the production of HBM itself is dominated by SK Hynix, Samsung, and Micron, all of whom are subject to US export controls. Second, the advanced packaging process known as CoWoS (Chip on Wafer on Substrate) is required to bond HBM to the logic die. Huawei’s domestic partners currently lack the scale and yield rates of TSMC’s CoWoS lines, creating a severe supply ceiling.

The role of compute in modern geopolitics cannot be overstated. Without a domestic supply of accelerators, a nation cannot train the next generation of sovereign AI models. Huawei’s struggle to secure reliable HBM sources at scale represents the front line of the US-China compute race.

If Huawei can successfully bridge the HBM and packaging gap through domestic innovation or alternative supply routes, it will significantly undermine the efficacy of current US sanctions. For now, the “compute race” is a race for stability, and Huawei’s ability to manage its die banks and HBM bottlenecks will determine whether China can remain a top-tier player in the production of AI tokens and the broader digital economy.