In the second half of 2026, China’s push to localize computing power is set to enter a new phase.

Ascend 950PR is Huawei’s next-generation AI processor for large-model inference and integrated training-inference workloads. Its launch comes at an industry inflection point: uncertainty around external computing power supply continues to rise, China’s large-model race is shifting from parameter counts to real-world deployment, and demand for inference computing power is beginning to exceed demand for training.

At this point in the cycle, a domestically made inference chip that can be supplied reliably, iterated continuously and kept cost-controlled matters far beyond the chip itself. If the discussion stays fixed on chip specifications, it quickly narrows into a familiar set of questions: How far is it from NVIDIA? Can it replace NVIDIA? Which company benefits most directly?

But real industrial change usually does not happen around one parameter. It happens in system architecture. Has China’s domestic AI chip industry truly reached a turning point?

What makes Ascend 950PR more worth watching is that it could push domestic computing power from “usable” to “good enough,” and from competition around a single chip or server to competition at the rack and cluster level.

That is also why its investment opportunity cannot be reduced to “Huawei concept stocks.” The real story sits in four ledgers: performance, capacity, systems and commercialization.

The performance ledger explains why it is useful; the capacity ledger shows where the bottlenecks are; the systems ledger shows where supernodes will redirect value; and the commercialization ledger asks who can actually make money.

Taken together, these four ledgers form the full logic of the domestic computing power supply chain behind Ascend 950PR.

Performance: Not Chasing the Strongest Single Card, but Solving Inference Costs

Over the past two years, training has been the hottest part of the AI industry. The race to build larger models has essentially been a computing power arms race. But once models are trained, inference is what determines commercialization.

Training is like building a factory: capital-intensive, long-cycle and highly concentrated. Inference is the factory’s daily production after it opens: the more calls it handles, the more computing power it consumes, and the more visible the cost pressure becomes. What companies really care about is not only whether a model can run, but the cost, latency and stability of each call.

That is why Ascend 950PR needs to be understood in the context of inference.

For the Chinese market, the most practical question is not whether China must own the world’s strongest training chip. It is whether, amid rising external uncertainty, it can find a computing power solution that is sustainable, available, portable and deployable at scale. Internet giants, telecom operators, financial institutions, government agencies and enterprise customers will not pay simply because a chip is domestic. In the end, they will ask whether the system runs reliably, whether models can migrate, and whether the economics work.

So the value of Ascend 950PR should not be judged only by peak single-card performance. It should be judged by its real-world usability in scenarios such as large-model inference, recommendation systems, high-concurrency AI services and enterprise knowledge bases.

If domestic computing power can replace only part of training demand, the market space is limited. But if it can enter the main battlefield of inference, the logic changes completely. Inference computing power is a continuous-consumption demand: the more users there are, the more frequent the calls and the deeper the applications, the more stable the demand for computing power becomes.

This is also the key step for domestic computing power to move from “project construction” to “infrastructure operation.”

Capacity: The Hardest Bottleneck Is Upstream, Not in Complete Servers

When looking at domestic computing power, the market often first focuses on server OEMs, because orders are the most visible and revenue recognition is the fastest. But the real bottlenecks in the supply chain are often further upstream.

For a high-end AI chip to move from design to mass production, it needs coordinated support across front-end manufacturing, advanced packaging, package substrates, PCBs, high-speed materials, memory and cooling. If any link fails, delivery schedules are affected.

Wafer manufacturing is the foundation. SMIC’s position in China’s domestic advanced process node ecosystem needs little explanation. For high-end AI chips, the stability, yield and capacity of front-end manufacturing are the prerequisites for the entire supply chain to deliver. Without sustainable manufacturing capability, downstream server orders and AI data center plans are just paper prosperity.

Advanced packaging is also becoming more important quickly. AI chips increasingly depend on high-density packaging, multi-chiplet integration and more complex interconnect structures. The roles played by JCET and TFME are no longer merely “back-end processing” in the traditional sense; they are now key to whether chip performance can be released. Future gains in domestic computing power will depend not only on design, but on coordination across packaging, substrates, materials and process technologies.

Further downstream, package substrates, high-end PCBs and high-speed materials will become very real bottlenecks. AI servers require far more from board layer counts, signal integrity, low-loss materials and yield than ordinary servers do. Companies such as Shennan Circuits, Wus Printed Circuit, Victory Giant Technology, Fastprint and Shengyi Technology are repeatedly discussed not because their names sound close to “AI,” but because high-end computing power systems raise the threshold for PCBs and materials.

The core judgment here is this: in the domestic computing power supply chain, not every supporting supplier has pricing power. The companies that matter are those positioned around high-end processes, yield, certification and delivery stability.

Low-end capacity is not scarce. What is scarce is supply capability that can enter high-end AI servers and future supernode systems. That is why the first type of opportunity brought by Ascend 950PR is not simply finding “who supplies Huawei,” but identifying “who is irreplaceable.”

Systems: Supernodes Will Reprice the Value Inside the Rack

If the performance ledger asks whether the chip is usable, and the capacity ledger asks whether the chip can be built, the systems ledger asks whether the chip can actually become effective computing power.

AI computing power competition stopped being a single-card contest long ago. One of the biggest bottlenecks in the large-model era is data movement between chips. The larger the model and the more parameters it has, the higher the cost of communication and memory access. In many cases, chips are not short of compute; they are waiting for data, communication and scheduling.

That is why “supernodes” matter.

A supernode can be understood as a rack-level supercomputing unit that organizes dozens or even hundreds of AI chips through high-speed interconnects, backplanes, liquid cooling, power supply and system software. It is not simply stacking chips together. It tries to make many chips work together like one larger chip.

This changes how value is distributed across the supply chain.

In the past, when the market looked at AI servers, it focused more on chips, complete systems and optical modules. But if computing power architecture shifts from standalone servers to supernodes, the carrying, connection, power supply and cooling systems inside the rack will be repriced.

First are rack-level PCBs and high-speed materials. Supernodes require more complex backplanes, midplanes or high-end PCB systems to support high-density chips and high-speed signal transmission. The requirements are not just about adding more layers; they demand comprehensive upgrades in materials, processes, thermal management, power integrity and signal integrity. Opportunities for companies such as Shennan Circuits, Wus Printed Circuit, Victory Giant Technology, Fastprint and Shengyi Technology should not be viewed only through the traditional PCB cycle, but against the broader rise in AI system complexity.

Second is high-speed interconnect. Cross-rack and cross-data-center communications rely on optical modules, and companies such as Zhongji Innolight, Eoptolink, Accelink and Huagong Tech have long benefited from upgrades in high-speed optical communications. But inside racks, and in short-distance board-to-board and high-density connection scenarios, high-speed copper cables, backplane connectors and high-density I/O connectors will become increasingly important. Connector companies such as Huafeng Technology are worth watching because the denser computing power becomes, the less connectors look like low-value accessories; they become critical components that determine system efficiency.

Third are liquid cooling and power supply. The direct result of supernodes is a sharp rise in power density. Traditional air cooling will struggle more with high-density AI racks, and liquid cooling will shift from optional to mandatory. The thermal management segment occupied by Envicool and Shenling Environmental should not be understood simply as “cooling equipment,” but as part of the capability to deliver high-density computing power. The real value in the future will not come from selling cold plates, CDUs or pumps and valves alone, but from providing integrated thermal management solutions from the chip level and board level to the rack and data center side.

The same is true for power supply systems. High power density will drive upgrades in power modules, power distribution architecture and energy efficiency management. The more concentrated computing power becomes, the more important systems engineering capability becomes.

So the essence of a supernode is not installing a few more chips. It is pushing AI computing power from a “server product” into “rack-level infrastructure.” Whoever can solve carrying capacity, connection, cooling, power supply and reliability may capture higher value in this round of domestic computing power repricing.

This is also the most easily underestimated part of Ascend 950PR: the market is watching the chip, but the real incremental value may be hidden inside the rack.

Commercialization: Revenue Elasticity and Profit Elasticity Are Not the Same

Investing in domestic computing power ultimately comes back to the commercialization ledger. The companies that recognize revenue first are not necessarily the ones that make the most money. The companies closest to the orders are not necessarily the ones with the deepest moats.

Server OEMs and system integrators will be the easiest direction for the market to understand. Companies such as Huakun Zhenyu under Gaoxin Development, Talkweb, Digital China and iSoftStone naturally attract attention because they are tied to domestic computing power servers, AI data center construction and Huawei’s ecosystem. For a chip to become computing power, it must enter servers; AI data center construction first shows up as server procurement. That logic holds.

But the problem with the complete-server segment is also very real: revenue elasticity can be high, while profit elasticity may not keep pace. If a company wins a large order but mainly handles integration, delivery and financing, its gross margin may not be high. A beautiful order book does not necessarily translate into beautiful shareholder returns.

So when assessing server OEMs, investors cannot look only at order size. They need to ask whether these companies can move from hardware sales to “hardware plus software plus operations and maintenance plus industry solutions.” If a company is merely an assembler, it captures the revenue dividend of the construction cycle. Only companies with system design, scheduling optimization, industry delivery and ongoing operations capabilities have a chance at longer-cycle valuation rerating.

The software ecosystem is another, slower but deeper line.

NVIDIA’s real strength is not just its GPUs, but the CUDA ecosystem. Developer tools, operator libraries, model adaptation, debugging environments and developer habits create a very high migration barrier. This has also been the biggest historical difficulty for domestic computing power.

The Ascend ecosystem is built around foundational software stacks such as CANN, MindSpore and openEuler. If it can keep lowering model migration costs and build stable projects across government and enterprise, telecom operator, financial and internet scenarios, it will create opportunities for a group of software services and industry application companies.

Companies such as iSoftStone, HopeRun Software, Talkweb, Digital China and iFlytek should not be viewed simply as “software outsourcing” or “AI concept” names. Their real value depends on whether they can turn one-off project delivery into replicable industry solutions and recurring operations revenue. Otherwise, no matter how hot the ecosystem becomes, the revenue may remain project-based and struggle to support higher valuations.

Computing power operations and industry applications will determine how far this construction cycle can go.

Once computing power is built, someone has to use it. If AI data centers run by telecom operators, cloud service providers, government and enterprise customers, and industry large-model applications cannot continue contributing call volume, even hot upstream equipment demand may turn into inventory. A healthy domestic computing power supply chain should form a closed loop from chips, servers and system integration to model migration, industry applications and computing power operations.

This is also the point investors should be most alert to when judging opportunities: not every company in the “Ascend chain” will be rerated.

Low-margin server integration may post fast revenue growth, but profit quality can be average. Pure concept-driven software adaptation may generate many announcements, but lack replicability. Low-end supporting components may be lifted by market sentiment, but may not enter the high-end AI system supply chain.

Companies with real value ultimately need to answer three questions: Are they part of the core supply chain? Can incremental revenue be verified in financial statements? Can gross margins and cash flow improve?

The Best Opportunities Are Not Where the Noise Is Loudest

The significance of Ascend 950PR is not that it proves domestic AI chips have fully pulled ahead, nor that NVIDIA’s advantages will disappear overnight.

Its real importance lies in the possibility that domestic computing power is moving from “point replacement” to “system competition.”

Behind this is a collective climb across the entire supply chain: manufacturing capability represented by SMIC; advanced packaging represented by JCET and TFME; PCBs and materials represented by Shennan Circuits, Wus Printed Circuit, Victory Giant Technology, Fastprint and Shengyi Technology; high-speed interconnect represented by Huagong Tech, Zhongji Innolight, Eoptolink, Accelink and Huafeng Technology; liquid cooling thermal management represented by Envicool and Shenling Environmental; complete servers and system integration represented by Gaoxin Development, Huakun Zhenyu, Talkweb, Digital China and iSoftStone; and the software ecosystem and industry applications represented by companies such as HopeRun Software and iFlytek.

But company names are only the entry point, not the conclusion.

What investors really need to watch is not who gets placed into a concept basket, but who solves industrial bottlenecks, who wins real orders, who can turn orders into profits, and who can turn one-off construction into recurring revenue.

What Ascend 950PR may open is not a simple chip rally, but a long-cycle opportunity in which domestic AI infrastructure moves from “usable” to “good enough,” and then toward “scaled use.”

In this cycle, the thing most worth tracking is not just the chip itself, but the entire system that carries it, connects it, cools it, drives it and ultimately turns it into sellable computing power. When computing power becomes systems engineering, where will the opportunity flow?

The real investment value in domestic computing power is not where the market is loudest, but where the bottlenecks are.