AI Trends

The US tightens controls on overseas remote computing power, and demand for domestic computing power is expected to accelerate

East MoneyUpdated Nov 18 16:03

AI summary

The US BIS has extended advanced AI chip controls to remote computing power services of overseas data centers. Providing more than 12 million TPP of computing power to enterprises in China and other places requires case-by-case license applications. Morgan Stanley estimates that this threshold is roughly equivalent to 700 to 800 H100s. NVIDIA's revenue in the third quarter of fiscal year 2027 was US$115.8 billion, a year-on-year increase of 103%, and data center business revenue was US$107.6 billion, confirming the prosperity of AI infrastructure demand. CITIC Securities believes that demand for domestic computing power is accelerating, and the industrial chain is entering a period of performance realization.

Body · AI translation

Main text EM_StockImg_Start

EM_StockImg_End

Major A-share indices rose collectively today. As of the close, the Shanghai Composite Index was up 0.09%, the Shenzhen Component Index up 0.20%, the ChiNext Index up 0.85%, and the STAR 50 Index up 2.19%. Trading volume across the Shanghai, Shenzhen and Beijing markets fell short of 1.4 trillion yuan, shrinking by nearly 100 billion yuan from yesterday. Across industry sectors, decliners outnumbered gainers, with communication services, electronic components, and shipping ports leading the gains, while rare earths, film and cinema lines, medical aesthetics, precious metals, gaming, aviation equipment, medical services, general retail, and media led the declines. At the individual stock level, more than 1,300 stocks rose, with over 40 hitting the daily limit-up.

The United States has brought overseas remote computing power access under licensing controls, drawing renewed attention to demand for domestic computing power. On November 16 local time, the U.S. Department of Commerce's Bureau of Industry and Security (BIS) issued an interim final rule extending advanced AI chip controls from physical exports to remote computing power services at overseas data centers. If a data center provides more than 12 million TPP of computing power to enterprises whose headquarters or ultimate parent companies are located in China (including Macao, China) and other such places, the chip exporter with knowledge must first apply for a license, subject to case-by-case review. Morgan Stanley estimates that this threshold roughly equates to 700 to 800 H100 chips; existing services that wish to continue must also obtain retroactive licenses within 90 days after the rule takes effect.

In addition, the latest financial reports from leading overseas computing power companies further validate demand for AI infrastructure. On November 17 local time, Nvidia announced third-quarter results for fiscal 2027, with revenue reaching $115.8 billion, up 103% year on year; of that, data center business revenue was $107.6 billion, up 110% year on year, continuing to be the company's main growth engine. At the same time, Nvidia disclosed on its earnings call that rental prices for some of its GPUs continue to rise: multi-year lease contracts renewed for Hopper-architecture GPUs reached as much as twice the previous price, and rental prices for B300 GPUs have risen more than 60% since launch. The company also expects revenue to grow about 75% year on year in fiscal 2028, further reinforcing market expectations for AI computing power demand.

On the domestic policy front, the People's Government of Guangdong Province issued the "15th Five-Year Plan for Building a Digital Guangdong," which states that a high-quality computing power supply system will be built. It calls for coordinating the construction of computing power hubs in the Guangdong-Hong Kong-Macao Greater Bay Area, optimizing the layout of computing facilities, and developing computing power in light of actual conditions. At the industry level, domestic large models have been updated intensively recently. On the evening of November 17, Kling AI's new-generation image creation model Kling Image 4.0 officially launched. That same evening, SenseTime officially released the SenseNovaLoop-9B model.

A CITIC Securities research report said that demand for domestic computing power is being released at an accelerating pace, and driven by continued increases in capital expenditure, the industrial chain is entering a period of earnings realization. Driven by the twin engines of capital expenditure from "telecom operators + cloud vendors," construction of domestic intelligent computing centers and super node clusters is comprehensively accelerating. Guolian Minsheng Securities noted that as overseas restrictions on the AI industry continue to intensify, the value of the domestic AI chain is becoming more prominent.

CITIC Securities: Demand for domestic computing power is being released at an accelerating pace

Demand for domestic computing power is being released at an accelerating pace, and driven by continued increases in capital expenditure, the industrial chain is entering a period of earnings realization. Driven by the twin engines of capital expenditure from "telecom operators + cloud vendors," construction of domestic intelligent computing centers and super node clusters is comprehensively accelerating; demand for domestic computing power is trending upward at the margin, and the earnings realization window for each positioned segment is opening in turn, with focus on core domestic computing power segments such as IDC, connectors, and switches.

Overseas restrictions on the AI industry continue to intensify, highlighting the value of the domestic AI chain; at the same time, Ascend super nodes are being deployed at scale and domestic large models are iterating intensively, with model-computing coordination of "domestic models" paired with "domestic chips" extending from inference adaptation to the training side, making domestic computing power investment opportunities worth watching.

Kaiyuan Securities: The key to computing power growth is extending toward system-level interconnect efficiency

AI generation lowers the threshold for content creation, and applications such as short dramas are expanding toward mass production, which is expected to generate more computing power demand. At the same time, as the parameter scale of AI models continues to expand, the key to computing power growth is extending toward system-level interconnect efficiency, with new GPU interconnect demand emerging within super nodes, and the weight of switching networks in computing clusters expected to rise. Application expansion and cluster architecture upgrades are bringing new demand increments to the AI industrial chain.

CITIC Construction Investment: The main investment theme in the AI industry will shift from spending expectations to ROI realization

At present, construction of domestic intelligent computing centers is accelerating, the penetration rate of domestic AI chips on the training and inference sides continues to rise, AI applications are moving from proof of concept to implementation in vertical scenarios, model ARR and token growth are exponential, computing power service prices are rising, and the AI industry maintains high prosperity. Looking ahead to 2027, the main investment theme in the AI industry will shift from spending expectations to ROI realization, and the core industry variables are the pace of domestic substitution and the validation of application commercialization.

Guosheng Securities: The long-term substitution logic of the upstream domestic computing power and infrastructure chain continues to strengthen

English text is being prepared.

English text is being prepared.

(This article does not constitute any investment advice; investors who act on it bear all consequences themselves. Markets carry risks, and investment requires caution.)

(Source: East Money Research Center) Main text middle Embedded advertisement

Source

East Money