Beijing’s push to bulk up domestic AI chip capacity is no longer just policymaker rhetoric. Mainland financial dailies have spent the week flagging accelerated build-outs tied to Huawei and provincial IC clusters, even as U.S. export controls keep tightening. The Financial Times reports China aims to triple AI chip output by 2026. Local coverage fills in the machinery: more packaging, more accelerators at mature nodes, and more state-directed capex to make it stick.
Shanghai Securities News and Securities Times have front-paged the same theme for days: accelerate domestic substitution in core chips. One headline captured the mood, urging “加快关键芯片国产化替代” — speed up domestic replacement of key chips. The policy intent is familiar. Since 2023, Beijing’s “东数西算” plan knit data center builds to industrial policy. The Ministry of Industry and Information Technology echoes that cadence, calling to “着力补链强链，推动先进封装等环节突破” — focus on strengthening supply chains and pushing breakthroughs in advanced packaging. China’s telecoms academy went further in a recent note: “算力供给是AI发展的关键基础设施” — computing power supply is foundational infrastructure for AI. Taken together, the signals explain why Huawei-linked fabs and OSATs are pulling forward equipment deliveries and labor hiring in the Yangtze River Delta and Chengdu-Chongqing corridors.
Onshore China semiconductor names outperformed into the close, led by OSATs, EDA software and power management. STAR Market turnover picked up, a sign that locals are rotating back into policy-favored tech after weeks of defensiveness. Hong Kong’s tech complex was mixed: AI server assemblers and optics caught a bid, but offshore-listed chip IP and smartphone names lagged. In Taiwan, reactions were muted at large-cap foundries, with investors focused on export orders and HBM constraints rather than China’s onshore volume. Korea’s memory leaders saw two-way trade as the market weighed the near-term lift from Chinese AI server demand against medium-term localization risks. Japan’s semiconductor equipment bellwethers firmed as traders leaned into a familiar read-through: policy-led capacity adds mean more vacuum tools, wafer handling, and test gear orders before year-end.
Tripling AI chip output by 2026 does not mean a sudden leap to Nvidia-class performance. The local build is centered on 14nm and above accelerators, networking ASICs, and advanced packaging to stitch smaller dies together. If the goal is inference at scale for domestic platforms and state workloads, volume and resilience matter more than bleeding-edge TOPS. That is why OSATs like JCET and Tongfu Micro and equipment makers such as AMEC and NAURA are structurally advantaged: advanced packaging, substrate supply, and thermal solutions are the near-term bottlenecks, not just the wafer node. The FT notes Huawei is lining up dedicated AI chip fabs through 2026. Local press consistently adds the “举国体制” element — whole-of-nation mobilization — to explain financing velocity and land allocation. For investors, the capacity math points to a spending curve that ramps OSAT, EDA, inspection and power components first, with foundry node upgrades pacing behind.
Huawei’s playbook is unchanged since the U.S. blacklist era. “备胎计划” — the spare tire plan — prioritized redundancy in chips, software, and supply chain nodes. That network is now spreading into AI accelerators and server platforms. Mainland coverage highlights how Huawei’s Ascend stack is being tied to domestic servers, storage, and networking to create captive demand for accelerators built on mature nodes. EDA champion Empyrean Technology has been showcased in provincial roadshows as the designated homegrown toolchain for specific flows. OSATs are being nudged to qualify new packaging lines compatible with domestic materials. A Securities Times line made the intent plain: “算力自主可控是产业发展底座” — self-sufficiency in computing power is the base layer for industrial development. Read in that light, plans to add multiple fabs by 2026 are less about catching Nvidia’s H100 and more about locking in a secure baseline for China’s AI economy.
There is a real gap. Local developers complain, “生态不成熟，迁移成本高” — the ecosystem is not mature and migration costs are high. Reports around Chinese startup DeepSeek’s delay on an Ascend-based model underscore the friction. CUDA lock-in is not just about speed; it is about the breadth of libraries, tools, and engineers trained on them. Huawei’s CANN and MindSpore stacks have made strides, but porting large models and custom ops still adds time and risk. The power and memory footprint is also different. If you cannot buy H100-class parts, you substitute dollars and engineering hours: more boards, more racks, more power, heavier cooling. That is why Chinese financial media have spotlighted power equipment and liquid cooling suppliers alongside chip names. The scaling strategy is pragmatic but expensive — and it channels capital into a wider domestic industrial set beyond the chip itself.
Washington’s successive rounds of export controls slammed the door on top-end Nvidia parts and restricted certain tools. China’s response has been to push the envelope on DUV, multipatterning, and chiplet-style assembly while expanding inspection and test to keep yields tolerable. Advanced packaging becomes a profit pool and a pressure point. CoWoS-class capability is still scarce; homegrown alternatives will carry lower density at first, but volume can close some gaps. Memory is the other choke. Without access to advanced HBM at scale, bandwidth becomes the limiter. The market’s quiet read is that China will field larger inference fleets built around domestic accelerators, older HBM or GDDR pairing, and heavy networking, enough for sovereign workloads and mainstream consumer AI, if not frontier training. The Japanese trade press has framed it simply: “規制が厳しくなるほど、国内供給網の自立化は加速する” — the tighter the controls, the faster the domestic supply chain tries to stand on its own.
If output triples, procurement dollars in China will tilt to domestic players across the stack. OSATs should see sustained utilization and better pricing power in select lines. Equipment makers tied to etch, deposition, cleaning, metrology, and test gain backlog visibility into 2026. Server assemblers benefit as state and large internet platforms diversify away from constrained imports; optics and power management ride alongside. Foundries capture volume on mature nodes with higher mix toward AI-centric parts, but pricing remains a swing factor. This is not a one-quarter pop; it is a multi-year, state-backed capex cycle with periodic pauses for sanctions risk and financing checks. Local banks and policy funds will underwrite it. As MIIT puts it, “加快发展新质生产力” — accelerate the development of new productive forces — and AI compute fits squarely under that banner.
Execution risk remains meaningful. Domestic alternatives must hit reliability and TCO targets in production, not demos. Ecosystem stickiness is hard to buy; it has to be built with tools, docs, and engineers. Power and cooling constraints are local bottlenecks. Sanctions escalation can reach packaging, substrates, or specific materials. And geopolitics can skew procurement. But there is also upside risk. If Chinese accelerators clear the bar for mainstream inference and state workloads, the addressable market is large and front-loaded. Local exchanges will keep rewarding policy-aligned names in chips, equipment, and infrastructure even through earnings volatility. Global suppliers in Japan and Europe exposed to consumables, metrology, and test can benefit from the build-out without tripping controls.
The Anglophone debate is still fixated on whether China can replicate Nvidia. That misses where cash flows will accrue over the next 24 months. The investable story is not a single killer chip; it is a state-orchestrated scale-up in advanced packaging, test, power, cooling, EDA, and AI server hardware that raises baseline compute self-sufficiency. Local media’s drumbeat — self-reliance in computing power, domestic substitution, whole-of-nation mobilization — is not propaganda filler; it is a procurement plan. If output does triple by 2026, the surprise will be how much revenue lands in picks-and-shovels segments and how little English-language coverage tracks those line items until after the orders are booked.
