Beijing deepens AI self-sufficiency

The chip campaign shifts upstream

Beijing’s self-sufficiency drive in artificial intelligence is increasingly focused on a less visible but decisive layer of the technology stack: chip design. The current signal is not simply that China wants more domestic accelerators. It is that Chinese policy makers and companies are trying to use AI itself to compress the design cycle for semiconductors, narrowing the gap between ambition, engineering capacity and usable silicon [1].

The most direct example comes from Empyrean Technology, described as China’s top electronic design automation, or EDA, tool maker. Its chairman Liu Weiping said the company has been applying AI-optimised algorithms and agentic tools to simulation and layout work, including a circuit-layout task whose time was reduced from four weeks to one week [1]. That claim matters because EDA software is one of the least glamorous but most strategic parts of the semiconductor supply chain. If design teams cannot simulate, verify and lay out chips efficiently, capital spending on factories and data centres cannot fully translate into frontier AI capability.

Liu framed the shift as a move away from “humans operating tools” and toward “humans commanding agents,” and said Empyrean is building an agentic EDA platform able to work with partners’ agents [1]. The commercial model may also change: according to his remarks reported by SCMP, the industry could move from traditional software licensing toward token consumption for agentic EDA [1]. In other words, Beijing’s self-sufficiency effort is not only about replacing imported chips with domestic chips. It is also about replacing imported design leverage with local software, local workflows and AI-native engineering.

Policy is converging with infrastructure

The timing is important. On September 11, Premier Li Qiang chaired a State Council executive meeting that studied work on building a national computing-power network, calling computing power a basic support for AI development [2]. The meeting urged a more orderly layout of compute infrastructure, faster research and application of key technologies and equipment, and a multi-level networked computing system [2]. It also linked compute build-out to power and grid planning, including compute-power coordination, green electricity and storage-related projects [2].

That agenda fits the same logic as AI-assisted chip design. A domestic accelerator is only useful if it can be designed, manufactured, packaged, supplied with memory, placed inside servers, connected into clusters and supported by software. Beijing’s present campaign therefore looks less like a single “chip substitution” policy and more like an attempt to coordinate the whole chain from EDA to data-centre deployment.

This is where self-sufficiency becomes operational. The question is no longer whether Chinese companies can demonstrate a chip, a model or a benchmark. The question is whether domestic suppliers can scale enough reliable compute at a price Chinese AI developers can absorb. The answer depends on tools, wafers, packaging, memory, interconnects, power availability and procurement mandates all moving in the same direction.

Capital markets are being enlisted

The capital-market side of the drive was visible on September 11, when Shanghai Enflame Technology, a Tencent-backed AI chipmaker, surged about 200% in its Shanghai debut after raising 6.12 billion yuan, or about $912 million, in its initial public offering [3]. Reuters reported that Enflame’s debut added to a rush of Chinese listings aimed at supplying AI computing chips as Beijing pushes home-grown alternatives to US suppliers such as Nvidia [3].

This is more than a hot IPO story. Enflame’s listing shows how policy priorities, investor appetite and customer demand are reinforcing each other. The company sold 43.04 million new shares, equal to 10% of its enlarged share capital, on Shanghai’s STAR Market [3]. Its jump occurred even as broader Chinese equity indexes were weaker, a sign that domestic AI-chip scarcity is being priced as a strategic asset rather than just another technology theme [3].

For Beijing, this is useful. Chip design is expensive, slow and risky. Frontier accelerators require repeated architecture revisions, software compatibility work, board design, cluster-level validation and customer-specific optimisation. Public markets can move some of that burden from state balance sheets to investors, while still channelling money into firms aligned with the self-sufficiency agenda.

The demand signal is real, but uneven

Biren Technology offers another window into the same cycle. Tom’s Hardware reported on September 10 that Biren posted first-half 2026 revenue of $183.9 million, up 1,998% year on year from roughly $8.7 million, as sales of non-Nvidia AI processors rose in China under export-control pressure [4]. The article also noted that Biren’s gross margin reached 42.7%, while the company still lost $56.2 million as it invested in new accelerators, optically interconnected rack-scale solutions and software [4].

Those numbers show both the opportunity and the limit. Domestic suppliers are benefiting from reduced access to foreign chips, but they are not yet automatically equivalent substitutes. Tom’s Hardware cautioned that Biren’s expansion came from a very small base and that its shipments remain tiny compared with Nvidia’s previous China volumes [4]. It also said Biren must secure enough manufacturing capacity from SMIC or other suppliers if it wants to compete with larger Chinese AI accelerator vendors such as Huawei, Kunlunxin and Cambricon [4].

That caveat is central to the current story. Design capability is necessary, but not sufficient. Beijing can direct capital toward local designers and encourage AI-native EDA, yet the ability to scale frontier systems still depends on fabrication capacity, advanced packaging, high-bandwidth memory and software ecosystems. The dependency Beijing wants to patch is therefore not a single bug. It is a chain of bugs.

Why AI-assisted design matters strategically

AI-assisted EDA is attractive because it attacks several weaknesses at once. It can shorten design iteration, reduce the need for scarce expert labour and help local firms make better use of the manufacturing processes available to them. If Chinese designers cannot always access the most advanced production nodes, then improving architecture, layout efficiency and system-level optimisation becomes even more important.

This is why the Empyrean example is strategically significant. A four-week-to-one-week circuit-layout improvement, if repeatable across more design tasks, would not merely save time [1]. It would change the economics of experimentation. Faster design loops can support more architecture variants, quicker bug-fixing and tighter hardware-software co-design. For AI accelerators, where performance is shaped by memory movement, compiler support and cluster behaviour as much as raw transistor density, those gains may matter.

There is also a sovereignty dimension. EDA has historically been dominated by a small number of global vendors, and restrictions on advanced semiconductor tools have made design software a geopolitical pressure point. A credible domestic EDA ecosystem would reduce one of the highest-leverage vulnerabilities in China’s AI stack. Agentic EDA may not erase that dependence quickly, but it gives Beijing a way to push the bottleneck upstream and make domestic engineering capacity more productive.

The new benchmark: usable compute

The most realistic reading is that Beijing is not yet declaring victory over Nvidia or the broader US-led semiconductor supply chain. Instead, it is changing the metric of progress. Success is no longer only whether a Chinese accelerator matches a foreign chip on a public benchmark. It is whether Chinese AI developers can get enough usable compute, with acceptable software support and predictable costs, to train and deploy competitive models under external constraints.

That is why the recent signals belong together: Empyrean’s agentic EDA work points to upstream design leverage; the State Council’s computing-network agenda points to coordinated deployment; Enflame’s IPO shows capital mobilisation; and Biren’s growth shows domestic demand moving from theory into revenue [1][2][3][4]. Each piece is incomplete alone. Together, they show a self-sufficiency strategy becoming more system-level and less rhetorical.

The risk is fragmentation. China has multiple accelerator architectures, software stacks and cluster designs. If each domestic chipmaker builds its own island, developers may face migration costs similar to those they once faced when leaving CUDA-based systems. The prize, therefore, is not just “Chinese chips.” It is a Chinese compute ecosystem that can absorb policy support without trapping users in brittle, incompatible platforms.

For now, Beijing’s message is clear. In the silicon campaign behind AI, dependence is the bug it wants patched. The latest developments show the patch being written not only in fabs and data centres, but in design tools, IPO prospectuses and national compute plans.

Sources from the last 72 hours

1. [1]Beijing pushes AI-assisted chip design as part of self-sufficiency driveSep 13, 2026, 6:00 AM UTC
2. [2]李强主持召开国务院常务会议Sep 10, 2026, 4:00 PM UTC
3. [3]Tencent-backed Enflame triples in Shanghai debut as China AI chip bets surgeSep 11, 2026, 1:32 AM UTC
4. [4]China's AI accelerator supplier Biren posts 2,000% year-over-year revenue growth — US export controls benefit homegrown chips as Nvidia and AMD exit marketSep 10, 2026, 12:40 PM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.