Huawei's chip comeback tests U.S. controls

Business news

The Financial Times said Wednesday that Huawei has pushed its chip comeback from phones into AI systems, forcing U.S. officials to test whether export controls can slow a company that has rebuilt much of its hardware chain inside China.

U.S. officials tried to choke Huawei's access to advanced chips in 2019 by putting the company and affiliates on the Commerce Department's Entity List. The company began building backup parts through HiSilicon, its chip design arm, and leaned on Chinese manufacturers after U.S. suppliers lost access to one of their largest telecom and handset customers.

Huawei's answer has become clearer. The company pairs domestic chip design with advanced packaging, optical links, and large clusters that compensate for weaker single-chip performance. That strategy matters in AI because developers buy working systems, not bare benchmarks. If Huawei can offer enough compute for training and inference, Chinese cloud firms and state customers can reduce purchases from Nvidia even if each Huawei accelerator trails Nvidia's best silicon.

Huawei's 2024 annual report shows the business rebound. Revenue reached 862.1 billion yuan, a 22% increase from 2023. Net profit fell to 62.6 billion yuan as the company spent more on research and development, but consumer revenue climbed 38% to 339.0 billion yuan. The auto technology unit reached 26.4 billion yuan, a sharp gain from a small base.

Research spending gives the comeback its financial spine. Huawei spent $22.1 billion on research and development in 2024, equal to 22.4% of revenue. That budget lets the company fund chip design, operating systems, cloud software, and automotive systems without depending on a single product cycle.

TechInsights put the consumer rebound in motion in 2023 when its teardown linked Huawei's Mate 60 Pro to a Kirin 9000S processor from SMIC's 7-nanometer process. Engineers and policymakers treated the phone as evidence that China could push older lithography tools further than Washington expected. Chinese buyers gave Huawei a patriotic premium in handsets, and that demand helped restore volume in a business sanctions had damaged.

Market context

Commerce officials built the broader controls in stages. The Bureau of Industry and Security added advanced computing chips, semiconductor tools, and supercomputer end uses to its China controls in October 2022. U.S. officials later tightened the rules as Nvidia and other vendors designed lower-power China versions.

Those rules gave Huawei a market opening. Nvidia engineers designed parts such as the A800, H800, and H20 to preserve China revenue under U.S. limits. Each revision left Chinese buyers with less certainty about future access. Huawei sales teams could offer a different pitch: local chips, local service, and fewer U.S. licensing shocks.

Huawei engineers chose system scale as their main workaround. Huawei researchers described CloudMatrix384, a production AI data center architecture that links 384 Ascend 910C NPUs with 192 Kunpeng CPUs. The design uses high-bandwidth links and software tuned for large language model serving. That approach lets Huawei pool many weaker chips into a system that can handle memory movement and token routing for large models.

Analysts have pointed to a cost. CloudMatrix-style systems use more chips and more power than Nvidia rack systems. They need more space, more cooling, and tighter operations. China can absorb some of that burden because state-backed cloud and telecom customers can fund power-hungry clusters when U.S. controls limit access to Nvidia gear.

Nvidia holds the stronger software ecosystem. Developers have spent more than a decade building around CUDA, and major AI frameworks gained Nvidia support first. Huawei counters with CANN, MindSpore, and cloud integration, but developers must port code, tune kernels, and handle gaps in tooling. That software burden slows adoption outside China and gives Nvidia protection in markets where buyers can obtain its chips.

Huawei uses scale to hide part of that gap. A cloud provider buying a full Huawei stack can get chips, servers, networking, software, and service from one vendor. That package appeals to Chinese customers that value procurement certainty as much as raw performance.

Implications

Huawei changes the export-control debate from chip denial to system competition. U.S. officials can slow China's access to the best tools, but Huawei shows that a constrained company can redirect spending into design workarounds, packaging, and cluster engineering. Chinese ministries and state customers can answer tighter U.S. rules by routing more orders to domestic vendors.

Procurement teams reward the same behavior Washington wants to deter. A bank, cloud operator, or government lab that fears future Nvidia restrictions gains a reason to test Ascend systems. Huawei gets more field data, more developer feedback, and more revenue for the next chip cycle.

Washington faces a market risk. American chipmakers lose China revenue as controls tighten, and Chinese rivals gain guaranteed demand at home. Chinese cloud providers learn to live with less efficient hardware because procurement policy and U.S. limits point them toward domestic systems. That learning curve can narrow the software gap over time.

Investors should treat Huawei's comeback as a margin story and a capacity story. Huawei can sell premium AI systems into a protected domestic market, but the company needs enough advanced memory, packaging capacity, and SMIC output to meet demand. Low yield and tight high-bandwidth memory access cap volume. SMIC has made progress without EUV lithography, but TSMC and Samsung keep a process lead at the front edge.

Nvidia can charge for performance per watt, mature software, and broad developer support in markets with legal access. Huawei can charge for availability in China and for political fit in markets that face U.S. licensing friction. Buyers in Southeast Asia, the Middle East, Africa, and Latin America will compare those offers as AI infrastructure spending spreads beyond U.S. hyperscalers.

Beijing will measure success in deployments rather than benchmark wins. If Huawei ships enough Ascend clusters to banks, telecom groups, model labs, and state clouds, China gains a domestic base for AI compute even with lower efficiency. U.S. officials would retain a choke point on the most advanced tools, but buyers would fund the Chinese alternative they hoped to contain.