This report does not constitute a rating action.
A new large language model (LLM) developed by Chinese startup DeepSeek promises democratization of AI access, competition to incumbents, and potential shifts in investment . Innovative engineering techniques enabled the developers to use relatively cheap hardware, improve compute and energy efficiency, and offer performance comparable to existing LLMs, suggesting that AI remains prone to disruption.
What's Happening
DeepSeek grabbed global attention on January 27 when it unveiled its open-source R1 LLM, a rival to OpenAI’s o1 model in terms of performance, cost, efficiency, and energy consumption. A resultant stock selloff, notably in U.S. tech, sent global equity markets down about 1.4%, or $1.7 trillion--based on the January 25 to January 28 change in ECAUWRLD, which tracks the global market capitalization of listed companies.
Why It Matters
DeepSeek’s model could prove a material advancement in generative AI (genAI). Its performance is comparable to the most advanced LLMs, but uses a fraction of their compute power and is supported by lower-cost and simpler graphics processing units (GPUs). S&P Global Ratings considers that DeepSeeks' launch will have widespread implications, including with regard to:
- Competition. It is increasingly apparent that no nation will have a monopoly on AI leadership and that there is room for diversification across genAI. That will put pressure on AI services' prices, particularly as new models are quickly emerging (see chart). Nations will continue to maneuver for AI supremacy, including through export restrictions, so we expect geopolitical dynamics to remain fluid.
- Democratization of innovation. R1's open-source status provides new access to advanced-reasoning LLMs that could reduce big-tech dependency and costs. This might particularly benefit small- and medium-sized enterprises, startups, academia, developers, educators, the public sector, and nations with limited financial resources. And it could foster collaborative AI development. From a credit standpoint, it could prove a positive factor for those that benefit from implementing AI solutions that differentiate.
- Engineering advancement. DeepSeek's engineering differs from established counterparts in three significant ways. First, its model was trained using less-advanced GPUs that proved cheaper and more computationally efficient (a choice somewhat dictated by U.S. export restrictions). Second, its model delivers operating efficiencies by partly adapting computing capacity to users’ requests--activating parameters only when needed (a concept known as Mixture-of-Experts, or MoE) rather than always activating all of its parameters. Third, facilitated by its MoE architecture, it deploys intermediate logical deductions to enhance efficiency, a concept known as Chain-of Thought reasoning.
- Market valuation and volatility. The U.S. tech stock selloff was a reassessment of incumbent, large AI companies' valuations. It also proved the power of AI technological advancements to disrupt markets (and the potential for malicious actors to amplify negative effects in the cyber space) and cause volatility that threatens financial stability and amplifies systemic risks.
What Comes Next
We expect some corporations, investors, and nations will re-assess genAI investments and expensive partnerships with established big tech companies , potentially shifting budgets to more affordable and open-source options instead of higher-cost and proprietary models.
DeepSeek’s example of economy over computing power, and its open-source nature, offers researchers technology to develop sustainable AI models. These could be optimized for edge computing (on devices like phones and laptops) and reduce reliance on energy-intensive cloud technology.
Policymakers may need to harmonize regulation of open-source AI models to combat the risk of miss-use and bolster data security and privacy. DeepSeek's story also highlights the importance of data sovereignty (control and ownership) and could prompt regulators outside China to implement laws requiring data to be processed and stored in the jurisdiction where it is collected.
Related Research
- Seek And Deploy: How DeepSeek Is Reshaping U.S. Tech Industry's Competitive Dynamics, Feb. 5, 2025
- Your Three Minutes In AI: Financial Systems Will Face New Systemic Risks, Oct. 4, 2024.
| Primary Contacts: | Miriam Fernandez, CFA, Madrid 34917887232; Miriam.Fernandez@spglobal.com |
| Martin J Whitworth, London 44-2071766745; martin.whitworth@spglobal.com | |
| Xu Han, New York 1-212-438-1491; xu.han@spglobal.com | |
| Secondary Contact: | Sudeep K Kesh, New York 1-212-438-7982; sudeep.kesh@spglobal.com |
| Research Contributor: | Sudheesh R Meesala, CRISIL Global Analytical Center, an S&P Global Ratings affiliate, Hyderabad ; |
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.