Key Takeaways
- China's DeepSeek AI model has upended assumptions about the cost of building powerful large language models (LLMs), opening the door to innovation by the country's internet companies.
- We see potential for substantial upside business risk among the Chinese internet firms that push deeper into AI.
- DeepSeek's breakthrough will put powerful LLM models in the hands of many, including Chinese firms without access to leading-edge AI chips.
China's DeepSeek LLM is upending cost assumptions in AI. The launch changed what was an arms race, with entities spending billions on advanced NVIDIA Corp. chips, to a more even competition. S&P Global Ratings believes the rollout will enable Chinese internet firms to rapidly integrate powerful, cheap AI models. This will be a boon for the many Chinese firms without access to leading-edge chips.
While the actual development costs of DeepSeek's LLM remain unclear, most observers agree that it introduced a significantly more cost-efficient open-source approach.
DeepSeek models are now readily available across multiple platforms including Alibaba Cloud, Huawei Cloud, Microsoft Corp.'s Azure, and Amazon Web Services. The inclusion points to rapid adaptation in the industry to DeepSeek's advances.
DeepSeek's breakthrough also suggests that leading LLMs could be developed using existing AI infrastructure in China. Although this may not have immediate ratings implications, they nevertheless indicate positive business risk for the country's technology industry.
Domestic Chips For Domestic AI
DeepSeek's new approach could substantially cut computational requirements for training or inference. This may enable Chinese firms to build compelling LLMs on less-advanced chips such as Nvidia's H20 chips or domestically produced semiconductor. The latter may drive demand for competitive products produced by China's chipmakers. According to press reports, Huawei Technologies Co. Ltd. is mass-producing AI chips that could theoretically rival Nvidia's H100 semiconductor in performance.
Competition is a great motivator. Chinese internet firms face both opportunities and risks from these developments.
DeepSeek's open-source model could empower China's technology startups or larger tech firms to leverage their LLMs to create superior search engines, e-commerce platforms, social media applications or games, or to generally improve users' online experience. This could pose considerable disruption risks to established Chinese tech giants.
Their options to respond may be more limited. Unlike U.S. companies, such as Microsoft, which has invested in OpenAI, Chinese internet firms may encounter significant government resistance to the control of, or acquisition of, disruptive startups. This flows from the government's stringent enforcement of the Anti-Monopoly Law (see "China's Internet Regulation: Fewer Surprises, Not Zero Surprises," May 23, 2023.)
Such constraints would compel the internet giants to deepen their own AI capabilities, including Alibaba Group Holding Ltd. and Tencent Holdings Ltd. Alibaba released its own Qwen 2.5-Max LLM model soon after DeepSeek and the LLM purportedly surpassed the DeepSeek-V3 LLM, and many Western LLMs too.
We believe Chinese internet companies will continue to heavily invest in AI. Their expenditures increased significantly starting in 2020 with the release of Nvidia's A100 chip. The launch of the powerful chip coincided with the emergence of LLMs and led many firms to investing in AI.
After a slowdown in 2022 and 2023 as U.S. chip restrictions took effect, spending spiked again in 2024 after Nvidia released the H20 chip. This was designed specifically for the Chinese market. We forecast Alibaba's and Tencent's capital expenditure in 2025 and 2026 to match spending in 2024 barring any further tightening of existing chip restrictions.
Chart 1
China's Tech Push And U.S. Restrictions
DeepSeek has shown that China can drive innovation in the industry, and that the gap between Chinese entities and AI leaders in the U.S. may not be as large as we had believed. It also affirms China's potential for breakthroughs, particularly as the country marshals its considerable resources in an all-out tech push.
The Chinese government's tech efforts involve mobilizing the country's financial and human capital, backed by the public and private sectors. U.S. restrictions on advanced technology exports to the country have only made this mobilization more urgent (see "China's Chip 'Moon Shot'--The Response To Restrictions," Nov. 2, 2023.)
The U.S. restrictions have had unintended consequences: they compelled Chinese firms to do more with less, which is driving innovation. Chinese firms can leverage the country's large pools of science and technology talent.
By 2025, China will produce almost three times as many math and science PhDs annually as the U.S., with nearly 80,000 graduates, versus about 21,000 for the U.S. If the estimated 20,000 international students in the U.S. were included in the latter count, China's total would still be double that of the U.S.
Chart 2
What May Come Next
Achieving China's AI goals may be more feasible than achieving its goals for advanced chip production. The semiconductor industry depends on a complex global supply chain that is challenging for any single country to replicate.
In contrast, LLMs can be developed independently by startups with access to the necessary funding and AI infrastructure. DeepSeek showed that funding may be very manageable for these new entities. On the second criterion, China's infrastructure may already be sufficient to compete globally in AI.
The U.S. may tighten existing chip restrictions if it perceives that current measures are insufficient to impede China's progress in AI. Reports indicate that the Trump administration is already contemplating stricter controls on Nvidia's H20 chip exports.
Access, or ability to produce, advanced chips may remain the crucial hurdle. Jevons Paradox suggests that increased efficiency in technology can lead to greater resource consumption. By this view, more efficient LLMs could enable the development of even more sophisticated AI models requiring more powerful AI chips, not fewer.
Firms' AI costs will continue to stack up. How much they will spend will depend on whether internet companies can monetize these models, justifying their hefty investments. The technology is changing rapidly but the essential economics of running a business remains the same.
Editor: Jasper Moiseiwitsch
Related Research
- China's Chip 'Moon Shot'--The Response To Restrictions, Nov. 2, 2023
- China's Internet Regulation: Fewer Surprises, Not Zero Surprises, May 23, 2023
This report does not constitute a rating action.
| Primary Credit Analyst: | Clifford Waits Kurz, CFA, Hong Kong + 852 2533 3534; clifford.kurz@spglobal.com |
| China Country Lead, Corporates: | Charles Chang, Hong Kong (852) 2533-3543; charles.chang@spglobal.com |
No content (including ratings, credit-related analyses and data, valuations, model, software, or other application or output therefrom) or any part thereof (Content) may be modified, reverse engineered, reproduced, or distributed in any form by any means, or stored in a database or retrieval system, without the prior written permission of Standard & Poor’s Financial Services LLC or its affiliates (collectively, S&P). The Content shall not be used for any unlawful or unauthorized purposes. S&P and any third-party providers, as well as their directors, officers, shareholders, employees, or agents (collectively S&P Parties) do not guarantee the accuracy, completeness, timeliness, or availability of the Content. S&P Parties are not responsible for any errors or omissions (negligent or otherwise), regardless of the cause, for the results obtained from the use of the Content, or for the security or maintenance of any data input by the user. The Content is provided on an “as is” basis. S&P PARTIES DISCLAIM ANY AND ALL EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, ANY WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE OR USE, FREEDOM FROM BUGS, SOFTWARE ERRORS OR DEFECTS, THAT THE CONTENT’S FUNCTIONING WILL BE UNINTERRUPTED, OR THAT THE CONTENT WILL OPERATE WITH ANY SOFTWARE OR HARDWARE CONFIGURATION. In no event shall S&P Parties be liable to any party for any direct, indirect, incidental, exemplary, compensatory, punitive, special or consequential damages, costs, expenses, legal fees, or losses (including, without limitation, lost income or lost profits and opportunity costs or losses caused by negligence) in connection with any use of the Content even if advised of the possibility of such damages.
Credit-related and other analyses, including ratings, and statements in the Content are statements of opinion as of the date they are expressed and not statements of fact. S&P’s opinions, analyses, and rating acknowledgment decisions (described below) are not recommendations to purchase, hold, or sell any securities or to make any investment decisions, and do not address the suitability of any security. S&P assumes no obligation to update the Content following publication in any form or format. The Content should not be relied on and is not a substitute for the skill, judgment, and experience of the user, its management, employees, advisors, and/or clients when making investment and other business decisions. S&P does not act as a fiduciary or an investment advisor except where registered as such. While S&P has obtained information from sources it believes to be reliable, S&P does not perform an audit and undertakes no duty of due diligence or independent verification of any information it receives. Rating-related publications may be published for a variety of reasons that are not necessarily dependent on action by rating committees, including, but not limited to, the publication of a periodic update on a credit rating and related analyses.
To the extent that regulatory authorities allow a rating agency to acknowledge in one jurisdiction a rating issued in another jurisdiction for certain regulatory purposes, S&P reserves the right to assign, withdraw, or suspend such acknowledgement at any time and in its sole discretion. S&P Parties disclaim any duty whatsoever arising out of the assignment, withdrawal, or suspension of an acknowledgment as well as any liability for any damage alleged to have been suffered on account thereof.
S&P keeps certain activities of its business units separate from each other in order to preserve the independence and objectivity of their respective activities. As a result, certain business units of S&P may have information that is not available to other S&P business units. S&P has established policies and procedures to maintain the confidentiality of certain nonpublic information received in connection with each analytical process.
S&P may receive compensation for its ratings and certain analyses, normally from issuers or underwriters of securities or from obligors. S&P reserves the right to disseminate its opinions and analyses. S&P's public ratings and analyses are made available on its Web sites, www.spglobal.com/ratings (free of charge), and www.ratingsdirect.com (subscription), and may be distributed through other means, including via S&P publications and third-party redistributors. Additional information about our ratings fees is available at www.spglobal.com/usratingsfees.