Can China construct a globally competitive end-to-end AI and robotics stack despite advanced-semiconductor restrictions?

Research note · 26 September 2026

Chinese manufacturers supplied 195,000 industrial robots installed in China in 2025. That is the starting point for assessing whether semiconductor restrictions can prevent a competitive Chinese AI-and-robotics stack: the country already has a substantial base of products and factory customers. [1]

Our view is yes: by 2028–2030, China can construct a globally competitive commercial stack, with a persistent frontier-compute disadvantage and incomplete technological self-sufficiency. We expect its strongest results in application deployment, industrial automation and selected robot products. Comprehensive frontier leadership is outside our base case. By competitive, we mean suppliers delivering usable systems across compute, models and physical deployment at acceptable customer economics.

The organizing idea is the difference between buying useful work and advancing the frontier. A factory needs a machine to perform a task at an acceptable lifetime cost. A frontier laboratory needs the resources to explore beyond today’s best models. Restrictions affect both, but their computing requirements differ. China’s commercial capability can consequently advance faster than complete independence—and faster than every supplier’s investment returns.

China’s factory base provides a route to useful robotics

The existing market gives Chinese suppliers somewhere to improve their products. China installed 354,000 industrial robots in 2025, up 20% from 295,000 a year earlier. Its operating stock had already exceeded two million at the end of 2024. That creates demand for controls, motion components, integration and service, as well as opportunities to learn from deployment. [1][2]

Domestic suppliers’ share nevertheless slipped from 57% to 55% while their installations increased. Our positive view rests on the scale of the customer and supplier base, rather than an assumption that localization rises every year. Foreign suppliers remain substantial competitors inside China. [1]

The distinction between established automation and general autonomy matters. A welding or handling cell working with defined fixtures can use conventional motion control and established sensing. A machine manipulating unfamiliar objects needs harder perception, dexterity, adaptation and safety validation. China’s manufacturing base helps with both, but the second problem requires considerably more development. [3]

We believe that supplier network and installed customer base are more defensible advantages than humanoid demonstrations. We give the proposed factory-data advantage less weight: useful learning requires customer permission, suitable observations and improvements that transfer between sites. Industrial scale supplies opportunities to collect such evidence. [3]

That explains how the stack can find customers. The harder question is whether China can supply enough reliable compute to keep their applications competitive.

Memory and system reliability limit the pace of independence

An accelerator needs working logic, high-bandwidth memory and a qualified package connecting them. More logic dies produce little additional capacity when the accompanying memory is scarce. Washington’s December 2024 controls reached HBM, semiconductor equipment and specified software tools, extending the restriction through the production chain. [4][5]

Our leading supply constraint is the combination of domestic HBM availability and advanced-node good-die supply. The memory evidence is more specific: Epoch’s central 2026 production model assigns domestic HBM to 240,000 of Huawei’s projected 1.5 million accelerator units, or 16%. A further 750,000 use less advanced, domestic-sourceable non-HBM memory. Different memory configurations serve different workloads. [4]

Epoch’s same-year comparison puts new Huawei production at 0.88 million H100-equivalents of specified processing capacity, against 23 million for Nvidia’s worldwide production. Dividing those projections gives a 26-fold gap between the two vendors. That is a formidable scaling disadvantage, although a Chinese laboratory can train a competitive model with a fraction of a global vendor’s output. We therefore give the supply comparison substantial weight without translating it mechanically into a model-quality gap. [4]

There are already substantial systems to build on. Huawei executive Eric Xu reported more than 300 Atlas 900 A3 SuperPoDs deployed to over 20 customers by September 2025. Alibaba reported more than 100,000 Zhenwu processors on its public cloud in May 2026, followed in August by more than 650 external customers adopting its newer processor through cloud services. These management disclosures show the scale of deployment and adoption. [6][7][8]

The work between installation and reliable service can be expensive. A July 2026 field study operated demanding DeepSeek-family workloads on sixteen Ascend devices. Making them serviceable required twelve source-level patches, disabling some throughput features to preserve numerical correctness and provisions for recurring faults. The episode shows why a compiler, runtime and operating team belong in the economics of the stack alongside the processor. [9]

Power consumption alone cannot settle those economics. SemiAnalysis estimated roughly 2.5 times the energy per theoretical BF16 operation for its CloudMatrix comparison with Nvidia’s GB200 NVL72. We test that engineering baseline across equipment-price, electricity-price and useful-output assumptions; the resulting task-cost range is approximately 0.8–3.4 times the comparator. The favorable case combines cheaper equipment and power with strong useful output. [10]

Our judgment is workload-dependent competitiveness. Lower local costs can compensate for some hardware disadvantages. Poor useful output can consume those savings. We expect applications with tractable computing requirements to clear the commercial hurdle more readily than continuous frontier training leadership.

Access to imported hardware adds another variable. Nvidia’s FY2026 filing recorded a US$4.5 billion charge following H20 licensing requirements and approximately US$60 million of revenue under subsequently granted licenses. It also described limited H200 licensing alongside uncertainty over Chinese import admission. The sequence shows how a commercially viable supply route can be interrupted and partially reopened. [11]

Software gives available compute a larger market

China’s software progress improves the application case. DeepSeek’s V3 report describes architectural and training efficiencies, including mixture-of-experts design and FP8 training. Its official run consumed approximately 2.79 million Nvidia H800 GPU-hours; an assumed US$2 per hour produces the widely quoted US$5.6 million charge. The report explicitly excludes “prior research and ablation experiments on architectures, algorithms, or data.” The engineering achievement is more quality from the compute used for that run. [12]

We think such efficiency helps Chinese suppliers meet more customer requirements with available hardware. Publicly shared techniques can also benefit competitors, so we are less persuaded that software efficiency alone delivers sustained frontier leadership. Reuters’ April 2026 report that Huawei chips participated in some DeepSeek-V4 training is a meaningful step toward deeper integration. [12][13]

Distribution is another tangible advantage. Hugging Face counted approximately 151,000 Qwen derivative repositories in its summer 2026 study. On OpenRouter, DeepSeek’s share of the gateway’s input-plus-output token volume rose from roughly 9% in January to 18% in June. Developer reuse and growing traffic show useful software finding an audience through different channels. [14][15]

We remain skeptical of universal Chinese inference-cost leadership because customers pay for completed tasks. In NIST’s September 2025 evaluation of specified model versions, GPT-5-mini delivered similar performance to DeepSeek-V3.1 at 35% lower average end-to-end expense across the tested benchmarks, using listed API prices. The comparison illustrates why task success and the work required to achieve it belong beside token prices. [16]

Useful software and distribution explain demand. The investment question is how much of the resulting revenue survives the cost of supplying it.

Cloud growth is producing segment earnings

Alibaba provides the clearest financial evidence of commercialization. Its AI Cloud and Compute Services segment generated RMB48.4 billion of revenue in the June 2026 quarter, approximately 45% above the prior year. Adjusted segment EBITA rose to RMB5.63 billion, taking the margin from 7.2% to 11.6%, calculated from reported revenue and EBITA. Chief executive Eddie Wu attributed the quarter to “improving commercialization of our full-stack AI capabilities.” [17]

Identified AI-related product revenue was RMB12.4 billion, equivalent to 25.6% of segment revenue. Its reported triple-digit growth bounds the previous year’s sales at RMB6.19 billion or less. Against the segment’s RMB15.0 billion revenue increase, the implied AI-product contribution is approximately 41–82%, assuming consistent classifications. That is a substantial contribution to growth. [17][18]

The allocation of profit remains consequential. Assigning the remaining cloud business an assumed adjusted EBITA margin of 10–20% leaves the AI products with a residual quarterly result ranging from a RMB1.6 billion loss to a RMB2.0 billion profit. This sensitivity subtracts assumed non-AI earnings from the reported segment total. Our view is that demand is commercially meaningful while AI-specific returns still require stronger evidence. [17][18]

The financing burden is already visible at group level: Alibaba reported RMB67.7 billion of capital expenditure and negative RMB44.7 billion of free cash flow in the quarter. These figures cover the whole company. A growing cloud profit pool and heavy group investment can coexist while the returns on new capacity develop. [17]

The wider market is uneven. Baidu’s AI Cloud Infra revenue reached RMB7.3 billion in the June quarter, up from RMB4.9 billion a year earlier but down from RMB8.8 billion in the preceding quarter. Tencent described cloud growth in the low twenties percent, supported by AI demand, international expansion and general cloud uptake. We would resist turning these different businesses and trajectories into a single claim of uniformly accelerating AI profits. [19][20]

Humanoids must earn back the cost of useful work

For robots, the customer’s return is more physical. An integrator must deliver an accepted system: the machine, tooling, safety provisions, programming, commissioning and support. Unitree’s advertised US$13,500 G1 provides a concrete hardware starting point. The factory’s budget must also cover deployment and operation. [21][3]

We favor established automation because a repeatable task offers a clearer route to dependable savings. Our fixed-cell scenario produces a 1.6–4.5-year simple payback, using an assumed RMB300,000 installed cost and genuinely avoided labor. The wage anchor is the statistical office’s approximately RMB114,000 annual average for urban non-private manufacturing employees; an assumed employer-cost uplift produces a RMB136,000–170,000 range. Site wages and the work actually displaced determine where a customer falls. [22]

The humanoid hurdle is sensitive to much more than its purchase price. Across our installed-cost, service, supervision and wage assumptions, annual break-even ranges from approximately 0.8 to 2.8 effective worker-years. At the base case’s RMB350,000 installed cost, the requirement is approximately 1.4 worker-years, or 2,800 labor-equivalent hours. By comparison, the fixed cell requires about 0.90 worker-year at a RMB150,000 loaded wage. Annualized capital recovery, service and supervision set those thresholds. [22]

Range assumptions: humanoid installed cost RMB250,000–600,000, annual service RMB30,000–100,000, supervision 0.15–0.75 worker-year and loaded wages RMB136,000–170,000, with the four-year life and 8% capital cost above.

Our base humanoid’s operating assumptions yield 975 labor-equivalent hours and a RMB29,000 annual shortfall before capital costs. The favorable scenario reaches 2,720 equivalent hours and pays back in about 1.4 years because acquisition and support costs are lower and avoided wages higher. We are skeptical of general-purpose labor replacement because its economics require dependable output with little human intervention; high-utilization, narrow-task sites have a more credible path. [22][21]

Those customer economics ultimately decide whether manufacturers win repeat orders.

The strongest businesses already have customers and cash

Our investment preference is for qualified products, repeat customers and cash conversion. The company accounts show why this is more selective than simply owning exposure to robot volume.

Established integrators face the same discipline. Estun’s FY2025 attributable profit of RMB149 million on RMB4.89 billion revenue gives a 3.1% net margin. Its RMB505 million operating cash inflow exceeded specified long-term-asset purchases of RMB262 million. Siasun, by contrast, reported RMB1.45 billion of first-half 2026 revenue and RMB433 million of operating cash outflow. Deploying systems can leave the supplier financing the customer’s project. [27][28]

Profitable diversified suppliers offer a different exposure. Inovance earned an approximately 11.2% FY2025 group net margin on RMB45.1 billion of revenue. Dividing reported category revenue by group sales puts its combined robotics-and-digital-energy exposure at about 4%, and Midea’s robotics-and-automation exposure at 6.4% in the first half of 2026. KUKA’s FY2025 EBIT margin was 1.5%. Established controls and customer relationships appeal to us; the current size and profitability of the robot contribution still matter. [29][30][31]

Semiconductor substitution also requires capital. SMIC generated US$3.01 billion of Q2 2026 revenue and spent US$1.84 billion on capital expenditure, a capex-to-sales ratio of approximately 61%. Gross margin was 25.3%, with management guiding to 26–28% in the following quarter. Local demand creates an opportunity, but the foundry must fund capacity and improve yields to capture it. [32]

Qualification, renewals and productive hours will decide the next stage

Global competition has a procurement gate as well as a technical one. Unitree reported approximately RMB732 million of overseas main-business revenue in 2025. The July 2026 FCC action generally bars new authorizations for covered foreign-produced advanced mobile robotic devices, with existing authorizations and specified exceptions shaping continuing access. Repeat overseas orders will matter as much as technical availability. [23][33]

The strongest challenge to our positive view is dependence on foreign memory and imported compute while domestic suppliers ramp. Tighter enforcement could interrupt that transition. The strongest challenge to our frontier skepticism is the combination of concentrated compute, efficient models and improving domestic systems. Reported participation in DeepSeek training and actual Ascend inference give that case weight, especially for selected workloads. [4][13][9]

We would become more constructive on autonomy with independently documented domestic HBM and qualified-package deliveries, followed by repeated full competitive training runs on identified domestic hardware. One milestone we would watch is annual domestic-HBM-equipped accelerator shipments above one million—more than four times Epoch’s modeled 2026 level—with product capability and sourcing specified. [4]

For broad economic parity, our proposed test is matched-model production across several operators: comparable quality and latency, sustained reliability, and all-in task costs within approximately 20% of the available global alternative for at least 90 days. Persistent premiums above twice the alternative after successive product improvements would weaken our commercial base case.

For investors, the remaining tests are separately disclosed AI-product earnings after depreciation, paid-customer renewal and sustained cash generation. In robotics, we want independently logged useful work at unrelated sites; our base humanoid economics require roughly 2,800 labor-equivalent hours annually. Evidence that factory data improve performance across sites would strengthen the case; inaccessible or nontransferable data would weaken it. [22]

These milestones would favor suppliers that turn qualification into repeat orders and platforms that turn usage into earnings. The next decisive demonstration is a repeat order from a customer whose first system has paid for itself.