The "Sold-Out" Era of Computing Power: Demand Exceeds Supply by Tenfold, Domestic Chip Orders Backlogged for Three Years

The past week has witnessed an extraordinary "super week" in the computing power sector. From across‑the‑board sell‑outs of domestic AI chips to the official launch of a four‑trillion‑yuan investment plan for computing networks, from breakthroughs in the national integrated computing network to the routine in‑orbit operation of space‑based computing services, the entire industry chain is undergoing unprecedented supply‑demand imbalances, ecosystem restructuring, and strategic shifts. This is a wave that underpins the digital economy, and a long‑term race that tests industrial foresight and patience.

I. Domestic AI Chips Fully Sold Out – How to Close the Million‑Unit Gap?

"Domestic chips are already sold out, with delivery schedules possibly extending to one year from now." This remark by a representative of China Power Engineering Consulting Group at a recent industry forum quickly went viral among tech professionals, revealing the real temperature of today's computing power market and the severity of supply tightness. According to multiple authoritative media reports, delivery lead times for domestic computing chips have generally stretched to more than a year, while total domestic demand is now more than ten times actual supply, and the gap continues to widen.

Even more striking is the shortage in the high‑end segment. At the Yuanjing星河 Base Smart Computing Park in Ulanqab, Inner Mongolia – which hosts the world's largest single‑building data centre – a senior executive stated bluntly that "orders on hand are already booked for three years ahead." This is not an isolated case. It is understood that many leading domestic computing enterprises have extended their delivery backlogs to two or three years, making high‑end AI chips and intelligent servers de facto "hard currency" and pushing the industry from the "product validation" phase firmly into "mass delivery" expansion.

Performance figures tell the story best. In the last week of August, interim reports from leading domestic AI chip makers painted an impressive picture. Biren Technology reported nearly 20‑fold year‑on‑year revenue growth, while Cambricon, Moore Threads, and other front‑runners all recorded substantial increases, with some posting their first quarterly gross profit turnarounds. According to industry estimates, total demand for domestic AI chips in 2026 will reach about 4 million units, while actual delivery capacity stands at only 3 million, leaving a clear million‑unit production gap that cannot be fully bridged in the short term – placing urgent pressure on upstream segments such as wafer fabrication and advanced packaging.

At the same time, domestic chips are gaining wider recognition for their performance and stability. Li Xinrong, Chairman of the Strategic Technology Committee at Biren Technology, confidently stated in a recent keynote: "Domestic chips have moved far beyond mere 'usability' – they have achieved a leap to 'good usability' in terms of performance and stability." From training clusters in large internet firms to inference deployments in finance, energy, and government sectors, domestic computing chips are steadily replacing imported solutions across diverse scenarios, building an increasingly complete independent technology stack and toolchain ecosystem.

II. Four‑Trillion‑Yuan Computing Network Investment Launched, "Token Factories" Accelerate

Soaring demand has directly spurred rapid investment in infrastructure. According to national plans, the "15th Five‑Year Plan" period will see an additional direct investment of 4 trillion yuan in computing network construction – a figure that underpins a vast blueprint from national hub nodes to city‑level edge facilities. On August 22, the Hohhot Green Computing Summit was held, where 13 key projects with a total investment of 136.1 billion yuan were signed, covering chip testing, equipment manufacturing, intelligent computing centres, and "token factories" – a microcosm of the computing infrastructure boom.

"Token factories" – regarded as the "power plants of the AI era" – are springing up across the country. In the Beijing Economic and Technological Development Zone, a token factory equipped with domestic large models and domestic chips entered mass production in August, processing 500 billion tokens per day – equivalent to the computing consumption of millions of concurrent users. In the Huangpu Comprehensive Bonded Zone of Guangzhou, a 950‑million‑yuan intelligent computing complex was recently approved, set to serve high‑consumption services such as text‑to‑video generation, code writing, and AI painting, while also promoting Chinese tokens to overseas markets. Sichuan's first "Token Factory" also went live in Ya'an, targeting a monthly capacity of 10 trillion tokens within the year, injecting strong momentum into the AI industry in southwestern China.

Meanwhile, the national integrated computing network has reached a critical milestone. According to the National Data Administration, more than 20 computing scheduling entities have now officially joined the national integrated computing network monitoring and scheduling pilot platform, significantly enhancing cross‑regional coordination. This means that enterprises in Beijing can, in real time, tap into idle computing resources in western hubs such as Guizhou and Inner Mongolia – moving from concept to practical implementation. By the end of July 2026, the country's total AI computing capacity reached 2.45 million PFLOPS (FP16), with the eight national hub nodes and three computing‑power‑electricity synergy zones accounting for over 85% of the total, signalling an initial formation of a concentrated and green development pattern.

Green and low‑carbon considerations are also central to infrastructure planning. In western regions rich in wind and solar resources, more intelligent computing centres are being co‑located with renewable energy projects, adopting integrated source‑grid‑load‑storage models to convert green electricity into computing power for transmission to the east. This "computing‑electricity synergy" not only reduces operating costs but also fundamentally alleviates concerns about energy consumption, contributing to China's carbon peaking and neutrality goals.

III. Industry Events Pack the Calendar, as Large Models and Computing Platforms Forge a "Two‑Way Street"

The last week of August saw a flurry of major industry conferences across the country, pushing computing power to the top of the agenda. From August 28 to 30, the 2026 China International Big Data Industry Expo was held in Guiyang under the theme "Tokens – A New Path to Unlocking the Value of Data Elements," attracting over 300 exhibitors covering the entire chain from chip design and scheduling to model training and data annotation. At the expo, the Ascend 950 supernode computing cabinet made its debut, achieving high‑speed interconnection of more than 1,000 cards through the latest chips, providing a new computing foundation for large‑model training.

Concurrently, the 2026 AGIC Shenzhen International General Artificial Intelligence Industry Expo opened in Shenzhen, with an exhibition area of 81,000 square metres and 1,012 participating companies. The bustling halls were filled with buzzwords like "computing power," "tokens," and "domestic chips," as the latest products from every segment of the domestic computing chain – from AI servers and liquid cooling systems to high‑speed interconnect chips – showcased the vibrant vitality of the domestic ecosystem.

The large‑model arena was equally lively, with new releases and open‑source announcements coming thick and fast. On August 26, Alibaba released and simultaneously open‑sourced its latest Qwen model, Qwen3.8‑Flash, attracting wide attention from the developer community for its smaller parameter count and higher inference efficiency. On the same day, Zhipu AI launched and open‑sourced GLM‑5.3‑Flash, the first natively multimodal model in the GLM‑5 series, whose training was supported by domestic chip clusters such as Hygon DCU – marking a full‑stack training accomplishment on domestic platforms. Moonshot AI released the Kimi K3 large model and open‑sourced it, with a total parameter count of 2.8 trillion; within less than a day, multiple domestic computing platforms announced successful adaptation, demonstrating an astonishing "China speed."

A strong consensus is emerging across the industry: large models and computing platforms are not in a one‑way consumption relationship, but rather a virtuous cycle of mutual reinforcement. As Professor Chen Wenguang of Tsinghua University noted during the Guiyang expo: "Between models and computing power, there is a complex interplay. Computing provides efficiency, models provide capability – neither can do without the other." The mass deployment of large models is forcing continuous improvements in interconnect bandwidth, memory capacity, and computational precision, while advances in computing platforms, in turn, enable further scaling of model parameters. This "two‑way street" is accelerating China's AI industry, forming a complete closed loop from underlying chips to top‑level applications.

IV. Policy Support Intensifies and Capital Enthusiasm Runs High

The past week also brought a steady stream of policy tailwinds. On August 31, the Ministry of Industry and Information Technology issued the "Notice on Launching a Special Action to Cultivate AI Application Service Providers," explicitly supporting providers to strengthen their connections with national computing network hub nodes and leverage policy tools such as "computing vouchers" to effectively reduce the cost of computing for small and medium‑sized enterprises. Earlier, the Central Cyberspace Affairs Commission had issued the "Action Plan for Promoting High‑Quality Development of Cyberspace Enterprises (2026‑2030)," which guides new computing facilities to co‑locate with renewable energy generation and provides key support in land, electricity, and water usage.

Local governments are also taking active measures. Beijing, Shanghai, Shenzhen, and other cities have introduced computing subsidy policies, offering "computing vouchers" that directly subsidise the computing costs of SMEs and university research teams, with annual subsidies ranging from hundreds of thousands to millions of yuan per enterprise. These policies have effectively lowered the barrier to AI entrepreneurship, enabling more innovators to access the large‑scale computing resources that were once only affordable to tech giants, and unleashing AI innovation across society.

Capital markets have maintained strong interest and active participation in the computing sector. On the afternoon of August 31, computing chip stocks collectively rebounded, with Wantong Development hitting the daily limit, while Cambricon and Xinpu shares led the gains. Computing leasing concepts also rallied, with Yitian Smart achieving a 20% surge. Cloud computing ETFs saw net capital inflows of more than 33 million yuan over the past five trading days, signalling long‑term institutional confidence in the sector. At the primary market level, several computing chip start‑ups completed new funding rounds in the past month, with investors including industrial capital, state‑owned funds, and VC firms, mainly earmarked for next‑generation chip tape‑out and mass production – the entire赛道 is undeniably "red‑hot."

However, the capital influx also carries certain risks. Some industry observers point out that there is a degree of redundant construction and froth in the computing field, with some projects lacking clear business models and customer validation, relying solely on the "computing power" narrative to secure financing. Maintaining rationality amid the frenzy and focusing on quality over sheer expansion will be a shared challenge for all stakeholders in the coming period.

V. From "Usable" to "Good" – The Long‑Haul Race in a Global Competitive Landscape

Looking back at the week's bustling computing scene, we may be standing at the threshold of a qualitative transformation. "Usable" is merely the starting point; "good" is the goal. In previous years, domestic computing chips lagged notably in toolchain maturity, framework compatibility, and distributed training stability. But through sustained R&D investment and application refinement, that gap is now closing at a visible pace.

The true value of the entire domestic computing chain is not simply "cheaper" chips and servers, but a "more controllable" full‑stack system – cost‑controllable, data sovereignty‑controllable, and supply chain security‑controllable. These three pillars form a genuine moat for the domestic computing industry. In an era of heightened geopolitical uncertainty and mounting restrictions on advanced processes and equipment, possessing complete independent capabilities from design to manufacturing to packaging and testing is especially valuable.

In the international arena, China's computing industry is competing head‑to‑head with the world's most advanced players. The US "Stargate" project plans to invest $500 billion to build the world's largest AI computing infrastructure, while Europe, Japan, and South Korea have also unveiled their own computing investment plans – the global computing race has entered full swing. China enjoys unique advantages, including the world's largest application market, the richest industrial scenarios, and the most complete manufacturing system. At the same time, however, gaps remain in advanced process technology, high‑end IP, and EDA tools – areas that require sustained effort to close.

From the daily call volume of 500 trillion tokens, to the tenfold supply‑demand gap, to the 4‑trillion‑yuan investment blueprint and the million‑unit chip shortfall – these numbers paint not only a thriving industrial landscape, but also a long‑haul race that tests planning wisdom, technological accumulation, and industrial patience. The domestic computing industry is transitioning from policy‑driven and capital‑driven growth to innovation‑driven and market‑driven development – a path that is bound to be challenging, but one that must be travelled.

And in the broader realm of imagination, the physical boundaries of computing are being redefined. On August 30, the world's first space‑based computing cloud to provide external experimental services, led by Beijing University of Posts and Telecommunications, officially entered routine in‑orbit service. This means that computing power now extends beyond terrestrial data centres to the distant reaches of space, ushering in a new era of integrated terrestrial‑space intelligent computing. Key technologies – from radiation‑hardened chip design in extreme space environments, to inter‑satellite laser communication networking, to on‑orbit resource scheduling – are laying the groundwork for future deep‑space exploration and satellite internet.

This week's computing power frenzy may be just the beginning. When computing becomes as ubiquitous and on‑demand as electricity, and when "tokens" become a key metric for gauging the vitality of the digital economy, every one of our lives will be profoundly reshaped by this wave. From voice assistants to autonomous driving, from weather forecasting to drug discovery – computing power is, in its own way, lifting the foundations of a smarter, more efficient, and more promising future.