Domestic AI Chip Makers See "Sold-Out" Orders, Mega Smart Computing Projects Approved: AI Computing Demand Is Soaring

On August 27, 2026, a news report from China Central Television’s financial channel sent shockwaves through the AI industry: "Domestic AI chips are sold out, with delivery schedules possibly stretching a year ahead." At the same time, a smart computing complex with a total investment of 9.5 billion yuan was recently approved in the Guangzhou Huangpu Comprehensive Bonded Zone. The Inner Mongolia computing hub just signed 13 key projects worth 136.1 billion yuan, and the Ministry of Industry and Information Technology disclosed that China’s intelligent computing capacity has reached 2,185 EFLOPS. From chips to data centers, from fundamental hardware to upper‑layer applications, China’s domestic AI computing industry chain is moving decisively from "product validation" to "mass delivery" – and the explosive growth in demand has far exceeded everyone’s expectations.

Token Calls Exceed 500 Trillion per Day; Computing Demand Is 10 Times Supply

Why is computing power so scarce? The answer lies in the key metric "tokens." Tokens are the smallest units of text processed by large language models – one Chinese character may correspond to one or several tokens – and every model inference, every dialogue generation consumes massive amounts of token‑based computation. As of June this year, China’s daily average token call volume had surpassed 500 trillion – a figure that was still below 100 trillion just a year ago. China’s large models are now firmly in the global first tier. From text‑to‑image, text‑to‑video, to code writing and scientific research assistance, fierce competition among models and the rapid rollout of AI applications have directly driven exponential growth in computing demand. Take video generation alone: generating a 10‑second 1080p HD video may consume tens of thousands of tokens’ worth of computing resources, and domestic short‑video platforms now generate millions of AI‑produced clips every day.

Liu Miao, founder of Guangzhou Turing New Smart Computing, stated bluntly: "Demand may be ten times or more than supply." Global smart computing resources are already in tight supply, and many regions in China are accelerating the construction of domestic token factories and smart computing complexes. In Changshu, Jiangsu, a smart computing center saw its peak electricity load grow over 70% year‑on‑year in August, with the computing industry contributing nearly half of the city’s total electricity load growth – computing power is replacing air conditioning and industrial manufacturing as the new "biggest consumer of electricity." The pace of power infrastructure expansion can no longer keep up with the speed of GPU server deployment. In some areas, substation renovation timelines have been cut in half, yet new smart computing centers still face the awkward situation of "power available but racks empty, racks ready but chips missing."

Domestic Chips "Sold Out" with Year‑Long Lead Times; Earnings Explosion Is Stunning

In this computing hunger, the fate of domestic chips has drawn the most attention. Zhang Chen, deputy general manager of the Market Development Division at China Power Engineering Consulting Group, made it clear: "Domestic chips are already sold out, and the lead time could be a year or more." This judgment is not only based on frontline industry feedback but is also strongly supported by semi‑annual reports from listed companies. In the first half of 2026, Hygon Information achieved revenue of 9.099 billion yuan, up 66.52% year‑on‑year, with its ShenSuan series GPUs being deployed in multiple smart computing centers. Cambricon posted revenue of 5.996 billion yuan, a 108.13% increase, with its SiYuan series AI chips continuing to expand their share in bids from internet and telecom operators. A‑share digital chip design companies collectively recorded net profits attributable to shareholders of 39.496 billion yuan in the first half of the year, a staggering 405.64% year‑on‑year growth. The earnings firepower of domestic AI chip makers is remarkable, and more encouraging still, leading companies have backlogs far exceeding their first‑half revenues, promising an even steeper growth curve in the second half.

At the same time, the integration of domestic models with domestic chips is accelerating – not only a commercial choice but also a core strategic imperative for supply‑chain security. Zhipu’s newly released GLM‑5.3‑Flash, for the first time, attempted to provide services on large‑scale traffic using domestic chip clusters, with all internal test traffic carried by domestic chips, meeting commercial standards for inference latency and throughput. MiniMax is also advancing the adaptation of its M3 and H3 models to domestic chips and has indicated that large‑scale domestic computing clusters will go online soon. In addition, domestic chips such as Huawei’s Ascend 910 series, Enflame’s Suisi series, and Biren’s BR series have achieved large‑scale substitution in key industries including finance, government, and energy. The old stereotype that "domestic chips can only train, not infer" is being broken; increasingly, production environments are adopting hybrid architectures – training on Nvidia, inference fully shifted to domestic chips – to reduce costs and mitigate supply‑chain risks.

Mega Smart Computing Projects Approved in Quick Succession; Full‑Chain Acceleration from Chips to Data Centers

With demand gushing, infrastructure investment is ramping up in tandem. In the Guangzhou Huangpu Comprehensive Bonded Zone, a smart computing complex with a total investment of 9.5 billion yuan was recently approved. The project plans to deploy 3,000 high‑power racks, using liquid cooling and waste‑heat recovery technologies. Once completed, it will support AI applications such as text‑to‑video and code writing, and leverage the bonded zone’s advantages to promote Chinese token exports, providing low‑latency computing services to AI companies in Southeast Asia, the Middle East, and beyond. At the national level, broader layouts are already underway. On August 22, at the "2026 Green Computing (AI) Conference," the Inner Mongolia computing hub signed 13 key projects with a total investment of 136.1 billion yuan, involving major players like Volcano Engine, Cambricon, and GDS Holdings, with a long‑term plan to reach 2 million P (PetaFLOPS) of computing capacity by 2030. According to the latest data from the Ministry of Industry and Information Technology, as of the end of June this year, China’s intelligent computing scale had reached 2,185 EFLOPS, with over 70 major computing corridors built around national hub nodes, forming a preliminary national computing backbone network.

In addition to large centralized computing centers, edge computing is also moving downstream. Major domestic cloud providers have begun deploying localized, small‑scale domestic computing nodes in industrial parks, ports, mines, and other scenarios, extending large‑model capabilities to the front lines of production. This "center + edge" distributed computing architecture is effectively alleviating computing congestion in core cities while reducing latency costs for cross‑regional data transmission. According to incomplete statistics, in the first half of 2026 alone, more than 40 new smart computing center projects broke ground nationwide, with total investment approaching 200 billion yuan – and over 60% of these projects explicitly require domestic chips to account for no less than 50% of total capacity.

Global Resonance: Nvidia Beats Expectations, Anthropic Signs $45 Billion Deal

The surge in domestic computing power is not an isolated phenomenon. The performance of overseas computing giants equally underscores the robust global demand for AI compute. Nvidia reported second‑quarter revenue for fiscal 2027 of $96.2 billion, up 106% year‑on‑year, with data‑center revenue reaching $89 billion, a 117% increase, while maintaining gross margins above 78%. The company unusually gave a guidance a full year ahead, projecting revenue growth of about 70% in fiscal 2028, and announced it would accelerate mass production of its next‑generation Rubin‑architecture GPUs. Even more stunning, on the very day of August 27, AI company Anthropic agreed to spend $45 billion to lease computing capacity from Nscale’s data center in West Virginia, USA, over a six‑year term, corresponding to about 460 megawatts of power capacity – the largest single computing lease deal in history, with an average annual rental of $7.5 billion, nearly several times Anthropic’s current annual revenue. On the overseas computing‑rental front, CoreWeave posted a 112% year‑on‑year revenue increase in the second quarter, with a backlog of orders reaching $104.2 billion, and its GPU cloud service prices have risen over 30% in the past year while still in short supply.

Global computing supply chains are also facing new uncertainties. Due to capacity constraints in high‑end GPU manufacturing, HBM (High‑Bandwidth Memory) and advanced packaging have become new bottlenecks, with TSMC’s CoWoS capacity booked through the end of 2027. This structural tightness on the supply side, however, has opened a valuable window for domestic chips – by combining mature process nodes with advanced packaging and architectural innovation, Chinese manufacturers are seizing the mid‑ and low‑end inference market with cost‑performance advantages, and gradually penetrating the training side as well.

Compute Equals Revenue, but the Supply Gap Is Still Widening; the Next Three Years Remain a Golden Window

Nvidia CEO Jensen Huang commented in the company’s quarterly report: "AI has reached an inflection point. Compute is revenue." This phrase perfectly encapsulates the essence of today’s industry – computing power is no longer just a technical metric, but a strategic resource that directly translates into commercial revenue. For cloud providers, every rack of GPU servers deployed can generate hundreds of thousands of yuan in monthly revenue; for large‑model startups, sufficient compute means faster iteration and lower trial‑and‑error costs. In vertical industries such as finance, healthcare, education, and manufacturing, computing power has even become a core determinant of a company’s competitiveness.

However, the supply‑demand imbalance continues to worsen. According to data from the China Academy of Information and Communications Technology, domestic AI computing demand grew 417% year‑on‑year in the first quarter of 2026, while effective supply grew only 128% – the absolute gap is widening rather than narrowing. CITIC Securities points out that the core pain point for domestic computing lies on the supply side; demand across the upstream chip, board, and system‑level chain continues to surge, but capacity release takes time, especially with bottlenecks still existing in advanced packaging and testing. It is expected that by 2027, as multiple domestic wafer fabs and packaging plants complete their expansion projects, the supply‑demand tightness may gradually ease. This also means that the next three years will remain a golden window for domestic computing development – whether chip designers, system integrators, or computing‑service providers, all stand to achieve qualitative leaps in this wave.

From sold‑out domestic chips to hundred‑billion‑yuan smart computing investments, from 500 trillion daily token calls to year‑long chip lead times, from the Guangzhou bonded‑zone smart complex to the green computing clusters on the Inner Mongolian grasslands – the computing frenzy of August 2026 reveals a clear truth: the AI industry has fully transitioned from technological exploration to large‑scale commercial deployment, and computing power is the scarcest "hard currency" in this transformation. When every line of code, every video frame, and every dialogue is voraciously consuming computing resources, whoever can first unblock the entire chain – from chips to power to algorithms – will secure an unassailable high ground in the smart economy of the next decade.