Computing Power "Inflation" and "Breakthrough" in Resonance

Core Judgment: Over the past week, the hottest signals in the computing power sector have come from two simultaneously accelerating curves—on one end, domestic computing power chips and supernode architectures are being deployed intensively; on the other end, global GPU rental prices and data center electricity costs continue to climb. Where these two curves intersect, a clear industrial logic is emerging: the "quantity" of computing power supply is exploding, but the "scarcity" of computing power has not faded—instead, it is concentrating upstream in a more structured manner. At the same time, the focus of computing power competition is shifting from single-chip performance to a comprehensive contest of system-level delivery capability, energy infrastructure support, and industry chain coordination.

Over the past week (September 17–24), the computing power industry has witnessed a rare "internal-external resonance." Domestically, giants like Huawei and Alibaba have successively released major developments, covering almost every key link in computing power infrastructure—from chips to supernodes to data center architecture. Internationally, signals such as rising GPU rental prices, chipmakers' full-capacity expectations, and large power equipment orders indicate that the global supply-demand balance for computing power is still tilting toward the supply side. Placing these two lines side by side reveals more clearly the deep structural changes underway in the computing power industry.

I. Domestic Computing Power: Intensive Fulfillment from "Single-Point Chips" to "System-Level Foundations"

From September 17 to 24, domestic computing power entered a rare intensive release window. Unlike previous rounds, the focus of this wave is no longer on benchmark scores or parameters of a single chip, but on systematic engineering centered on "how to deliver computing power at scale and with high efficiency."

Huawei's Direction: At Huawei Connect 2026 held on September 17, Rotating Chairman Wang Tao announced the AI infrastructure strategy for the intelligent era, with the core statement being "the core of Huawei's AI strategy is computing power," and the path revolving around "supernode + cluster" to advance system architecture innovation. A supernode refers to tightly coupling multiple AI chips through high-speed interconnect technology to form a logically "supercomputer," thereby improving overall computing power performance through cluster scale and interconnect efficiency when single-chip performance is constrained. Huawei views this path as a key breakthrough direction given restricted access to advanced process nodes.

In terms of specific progress, the Ascend 910C supernode has been deployed in over 1,000 sets, the Ascend 950 supernode has entered commercial-scale deployment, the Ascend 960 chip is progressing ahead of schedule with doubled performance, and Ascend 970 and 980 chips are planned for launch in 2028 and 2029 respectively. This means Huawei has provided a relatively clear chip evolution roadmap, with a tighter pace than previously expected by the outside world. CICC analysis suggests that 2026 is expected to be the first year of volume deployment for domestic supernodes, with the industry moving from proof-of-concept to batch delivery. On September 20, the DataPort and Huawei Ascend 910C supernode computing cluster was officially lit up, serving as a landing footnote to this judgment. From "over 1,000 sets deployed" to "commercial-scale deployment" to "batch delivery," domestic supernodes are crossing the critical threshold from technically feasible to engineering-usable.

Alibaba's Direction: At the Yunqi Conference opening on September 22, T-Head released the Zhenwu V900, currently China's most powerful self-developed AI chip in terms of computing performance, with computing power tripled compared to the Zhenwu M890, planned for mass production in Q1 2027, and a single AI cluster scalable to 500,000 cards. The 500,000-card cluster scale is a highly symbolic figure—it means domestic chips are not only catching up in single-card performance but also setting higher requirements for cluster interconnect, scheduling, and stability. If this goal can be achieved, it will significantly narrow the scale gap between domestic computing clusters and leading international clusters.

Even more striking is the infrastructure-level data: based on the CUBE 5.0 modular architecture, AIDC data center delivery time has been compressed from the traditional 6 to 12 months to 100 days, and computing power density per unit area has reached 5 to 10 times that of the previous generation. The significant shortening of data center delivery cycles directly relates to the responsiveness of computing power supply. In an era where AI model iterations are measured in weeks, the speed of computing power delivery is itself a competitive advantage. Alibaba CEO Wu Yongming offered a highly symbolic judgment—"the total amount of machine thinking in the future will reach more than 1,000 times that of humans," while today that ratio is less than 3%—and announced a goal for Alibaba Cloud's global data center scale to exceed 20GW by 2032. The 20GW scale target means Alibaba Cloud is elevating the competition dimension of computing power infrastructure from "number of chips" to "energy capacity."

Both paths point to the same trend: the competitive focus of domestic computing power is shifting from "whether chips exist" to "whether large-scale, high-efficiency computing clusters can be delivered systematically." Supernodes, modular data centers, and full-stack collaboration have become the keywords of this stage. Looking further, the breakthrough path of domestic computing power has a distinct "systems engineering" characteristic: when single-card process technology temporarily lags, high-speed interconnect, cluster scheduling, modular delivery, and energy infrastructure are used to compensate for single-point gaps. The advantages and challenges of this path are both obvious—the advantage is that it bypasses some process limitations and can form usable computing power relatively quickly; the challenge is that system complexity is higher, and the demands on software ecosystems, operations capabilities, and industry chain coordination are also greater.

From an industry chain perspective, the intensive deployment of domestic supernodes and large AIDCs will directly drive demand in segments such as servers, liquid cooling, optical modules, power supplies, cabinets, and temperature control. Liquid cooling and high-speed optical interconnect, in particular, are almost "rigid requirements" of supernode architecture. With the commercial-scale deployment of the Ascend 950 supernode and the promotion of Alibaba's CUBE 5.0 architecture, related industry chains are expected to usher in a period of relatively certain volume growth from the second half of 2026 to 2027.

II. Global Computing Power Supply and Demand: Price Increases Are Not Noise, They Are Structural Signals

Almost in sync with the intensive domestic computing power releases, overseas computing power markets have transmitted more tangible signals.

On September 17, AI cloud service provider Nebius announced that starting October 1, it would raise prices for on-demand computing resources, with the largest increases for NVIDIA's latest-generation GPUs: H100 up about 17%, H200 up 20%, B200 up nearly 19%, and B300 up about 21%. Price increases were not limited to GPUs—AMD EPYC CPU rates were raised 25%, and Genoa memory rates were raised about 41%. GPU rental price increases are not uncommon, but the simultaneous significant increases in CPU and memory rates indicate that the tension in computing power supply has spread from accelerator cards to the entire server hardware system. The rise in memory rates is particularly noteworthy, as supply bottlenecks for HBM and DDR5 are becoming hidden constraints on AI server delivery.

On the same day, Jensen Huang said at an AI summit in Scotland: "I expect the number of chips NVIDIA sells next year will be twice this year's," and explicitly pointed out that the current bottleneck is not demand but production capacity. This statement conveys two key pieces of information: First, demand remains strong—NVIDIA's expectation of doubling shipments next year means global AI computing power construction is far from reaching its peak. Second, supply-side capacity constraints still exist—chip manufacturing, advanced packaging, HBM supply, and other links are all operating at full capacity. By explicitly pointing to capacity rather than demand as the bottleneck, Huang is effectively telling the market that the core contradiction of the computing power shortage lies on the supply side, and supply-side expansion takes time.

If we broaden our perspective further, the "rush to buy" AI infrastructure has spread from chips to electricity. On September 16, backup generator manufacturer Generac announced a long-term supply agreement with Amazon, with cumulative procurement amounts of up to $8 billion, and its stock price surged more than 30% intraday after the announcement. Backup generators were originally ancillary equipment for data centers, but now they have become strategic materials for cloud providers to lock in long-term supply. The logic behind this is that data center electricity demand is exceeding the grid's carrying capacity, and self-provided power has become a necessary condition for ensuring stable computing power operations. Putting cloud provider price increases, chipmaker full capacity, and large power equipment orders together, the conclusion is quite clear: the supply-demand balance in the computing power market is still tilting toward the supply side, industry chain profits are concentrating upstream, and the computing power cost pressure on downstream AI application companies is rising in tandem.

Another noteworthy development: On September 23, according to reports, Anthropic is in preliminary negotiations with data center developers to lease up to 1 gigawatt of computing capacity, attempting to bypass cloud service providers and directly lock in infrastructure. A 1-gigawatt computing capacity roughly corresponds to the output of a large nuclear power unit. Negotiations at this scale indicate that leading AI companies' thirst for computing power has exceeded the "renting cloud" layer and begun to penetrate deeper into physical infrastructure. Bypassing cloud service providers to work directly with data center developers can reduce long-term computing power costs on the one hand and obtain higher customization and supply certainty on the other. This may signal that the relationship between AI companies and infrastructure providers will become more direct and tighter in the future.

From a broader macro perspective, the tension in global computing power supply and demand is not a short-term phenomenon. AI model parameter scales are still growing, inference demand is rising rapidly with application deployment, and capacity expansion in chip manufacturing, advanced packaging, HBM, and power supply all require time and capital expenditure. Supply-demand mismatches will be difficult to fully eliminate in the medium term, which means computing power prices will remain elevated for the foreseeable future, and the ability to secure computing power resources will become an important component of AI companies' core competitiveness.

III. Why These Two Lines Must Be Viewed Together

Looking at domestic chip releases alone, one easily falls into a "parameter race" narrative; looking at overseas price increases alone, one easily remains within a "cyclical fluctuation" framework. But placing the two side by side reveals a more complete picture.

First, the "structural" nature of computing power demand has not weakened—only the main body of demand is migrating. Domestically, after Zhipu GLM-5.3-Flash optimized inference efficiency on domestic computing clusters, its single-day call volume reached 4.36 trillion tokens on September 22, up 73% from the previous statistical period, ranking first in model call volume on the OpenRouter platform, and its end-2026 ARR guidance was raised from $2.4 billion to $3 billion. CITIC Securities' judgment is that AI training, inference, and infrastructure demand at home and abroad remains in a high-prosperity state. This shows that while overseas leading model companies are beginning to discuss "slowing down," domestic models are accelerating on the inference and application sides to absorb computing power demand. The explosion of inference demand places different requirements on computing power: compared to training, inference places more emphasis on throughput efficiency per unit of computing power, latency, and cost—exactly the areas where supernodes and specialized inference chips can leverage their advantages. The opportunity for domestic computing power on the inference side may be even greater than on the training side.

Second, the breakthrough path of domestic computing power has a unique "systemic compensation" characteristic. The strategic significance of Huawei's "supernode + cluster" path was summarized by Wang Lei, director of the Chinese Association for Artificial Intelligence, as "using system-level innovation to compensate for single-point gaps." Alibaba's CUBE 5.0 modular architecture compressing delivery cycles to 100 days is likewise seeking efficiency increments at the engineering and system levels beyond chip process technology. Under the background of restricted access to advanced global process nodes, domestic computing power is using systems engineering capabilities across four dimensions—"cluster scale, interconnect efficiency, delivery speed, and energy density"—to partially offset the gap in single-card performance. The feasibility of this path has been preliminarily verified in supernode deployment and AIDC delivery, but its long-term competitiveness still depends on the maturity of the software ecosystem and the coordination efficiency of the industry chain. If domestic computing power can establish a reputation of "sufficient, easy to use, and affordable" in inference scenarios, it may be able to occupy a dominant position in the massive domestic AI application market.

Third, pricing power in the computing power rental market is shifting upstream—this is both pressure and a window for the domestic computing power industry chain. Rising overseas GPU rental prices directly push up the computing power acquisition costs of domestic AI companies. But at the same time, the scaled delivery of domestic supernodes and self-developed chips provides a relatively cost-controllable alternative path for the domestic market. The Ministry of Industry and Information Technology recently deployed work to expand and upgrade computing power infrastructure, and analysis suggests that domestic computing power is entering a "dual-driver window of industrial prosperity + capital catalysis," with this policy rhythm forming a配合 with the industrial rhythm. Policy-level expansion and upgrading includes both increasing computing power supply and optimizing computing power scheduling and reducing usage costs. If domestic computing power can find a balance between performance and cost, it is expected to accelerate penetration with policy support.

Fourth, energy is becoming a new variable in computing power competition. Whether it is Alibaba's 20GW data center target, Amazon's $8 billion purchase of backup generators, or Anthropic's negotiation for 1 gigawatt of computing capacity, all point to the same fact: the endgame of computing power is electricity. Chips can be stacked, clusters can be expanded, but power supply and cooling capacity are hard physical constraints. In the future, the siting of computing power centers will depend more on energy resources, and regions rich in nuclear, hydropower, wind, and solar resources may become new computing power highlands. For domestic computing power, energy infrastructure support capability may become a competitive dimension as important as chip performance. Whoever can lock in low-cost, stable power supply earlier may gain an advantage in computing power costs.

Conclusion

The density of computing power news over the past week has been extremely high, but perhaps only one sentence is truly worth remembering: computing power is transforming from a "technical metric" into an "infrastructure engineering project." Whether it is Huawei's supernode clusters, Alibaba's modular data centers, or Nebius's GPU price increases and Anthropic's direct leasing negotiations, all point to the same fact—AI's demand for computing power has not slowed; only the way that demand is expressed has changed. It has shifted from "how many cards to buy" to "whether large-scale, high-efficiency, affordable computing power delivery can be obtained stably."

In this transformation, the opportunity for domestic computing power does not lie in catching up with the parameters of a particular chip, but in using the certainty of systems engineering to answer the market's question about the certainty of computing power supply. Supernodes, modular data centers, inference optimization, and energy infrastructure—these seemingly scattered links actually constitute a complete computing power delivery system. Whoever can form a closed loop in this system may be able to seize the initiative in the next stage of computing power competition.

For practitioners, the key nodes to watch in the coming year include: the actual delivery scale of the Ascend 950 supernode, the mass production progress of the Zhenwu V900, whether the cost advantage of domestic computing power in inference scenarios can be realized, and whether power infrastructure policies can keep up with the pace of computing power expansion. For investors, the value of the computing power industry chain is spreading from a single chip link to system integration, liquid cooling, optical interconnect, power supplies, and other supporting links, and structural opportunities may be more prominent than overall opportunities.

The resonance of computing power "inflation" and "breakthrough" will continue, and the true watershed will appear in those enterprises that can translate system capabilities into stable delivery.