For half a century, Bridgewater has focused on building a deep understanding of global markets and economies to deliver insights for the most sophisticated institutional investors. In this newsletter, our co-CIOs share key themes from this research. Sign up and get access to our latest edition.
In both our investment engine and in AIA Labs, our in-house group focused on using AI to trade markets, we have been closely tracking developments across the AI space and their implications for economies and assets. This includes following what has been happening at DeepSeek—a leading Chinese AI lab. This week, the release of DeepSeek’s latest model—DeepSeek-R1—led to significant turbulence in tech stocks, including Nvidia’s share price falling 17% in one day.
Today, I want to share with you our thoughts on DeepSeek and the big questions it raises. To give some context, DeepSeek-R1 is a new reasoning model. These types of models take more time to respond to tasks and can engage in multiple steps of reasoning—a key shortcoming of traditional chatbot LLMs.
The current leading publicly available reasoning model is OpenAI’s o1 (OpenAI announced but has not yet released o3), which has achieved notable breakthroughs like solving challenging math problems. DeepSeek-R1 rivals OpenAI’s o1 model—equaling it on several math and reasoning benchmarks—and, importantly, does so at far lower cost.
This is raising significant questions about AI from here:
- Does DeepSeek’s success suggest advances in AI models can be quickly matched by competitors—effectively making good AI models a commodity?
- If AI models become commoditized, which companies will be the winners and which companies will be the losers? And will leading AI labs be incentivized to pursue significant new investment to improve their models? If that investment were to decline, it would be a notable negative pressure on growth.
- What does this mean about the prospects for further AI breakthroughs from here?
- And how much should we even trust the claims being stated about what DeepSeek has achieved?
Below, AIA Labs Chief Scientist Jas Sekhon and I share our early thoughts on these questions and what they mean for investors.
Greg Jensen
Co-Chief Investment Officer, Bridgewater Associates
What DeepSeek has done is important and impressive.
Rather quickly, they have assembled a top-five AI lab. They appear to have made significant scientific breakthroughs around model architecture, efficient reward functions, and software optimizations to radically improve reasoning per unit of compute. They are producing results only months behind the frontier models at a fraction of the cost. They are open-sourcing what they are doing and providing the tool at low cost—in practical usage tests, at one-twentieth of the cost of OpenAI’s model. They have, at least for now, leapfrogged Meta for the leading open-source LLM.
The DeepSeek results are a threat to the leading AI labs, as it is clear that close-to-state-of-the-art models can quickly be commodified.
This will make it more and more difficult for the frontier labs like OpenAI and Anthropic to monetize their existing IP. It likely will lead them to be much more discreet in how they expose their IP in the future as well. But there is still a ton to do between where they are today and their actual goal of artificial general intelligence (AGI). There will likely continue to be an incredible amount of capital willing to chase the AGI goal, since it will be so transformational and quite possibly a “winner takes all” scenario.
DeepSeek results show AI progress and efficiency is accelerating. This is good news for much of the ecosystem—and bullish for new AI investment.
Demand for compute isn’t slowing down as a result and will likely accelerate. The existential business threats that the likes of Microsoft and Google face have not eased and likely have just grown. They will invest whatever is necessary to ensure they are the leaders. This is good for hyperscalers because their complements (LLMs) got cheaper, and demand for reasoning is only going up. The existence of DeepSeek is even better for application builders because they can use an open-source model that is near the frontier. Meta’s open-source model was about two generations behind the leading models.
In the short run, there may be a correction in many companies’ stock prices, as this news clearly poses risks to many of the companies that have surged the most. We’ve seen some of this already in the recent market action. Reconsideration of how secure the incumbents are makes sense. As we describe below, many investors have expressed concerns about Nvidia, because DeepSeek’s success may encourage companies to invest more in achieving efficiency gains by optimizing how AI software interacts with the hardware. A significant selling point of Nvidia chips is its proprietary software, so concerns have grown that companies may invest in alternative software that may undermine Nvidia’s moat. However, it’s important to separate changes in market pricing from changes in fundamentals. Software development takes time, and against the grand scheme of accelerating compute demand, it’s unlikely to dent Nvidia’s revenue in the near term.
DeepSeek’s success with its open-source model means progress outside the frontier research labs can go faster, as more of the research can happen in public.
Researchers around the world are energized. AI research has already accelerated in public. This will cause AI developments to speed up because people outside of the frontier labs (such as OpenAI, Anthropic, and Google) can help develop the next-generation models. Moreover, even people in the frontier labs are excited because many of them previously did not know how reasoning models worked.
In addition, more efficiency in reasoning means people will buy more reasoning—we are not yet at the point of diminishing returns on the demand curve for reasoning. For instance, a lot of the demand for AI is not for LLMs but for other uses of generative AI, such as robotics, self-driving, chip design, and biology. LLMs are often an input for these broader applications. With better LLMs, the compute bottleneck moves elsewhere, and demand for these kinds of applications is unlocked.
In terms of how much DeepSeek spent to build its model, many of the headlines and discourse in the financial community do not even accurately reflect DeepSeek’s claims.
For example, some of the discourse we see is misreading DeepSeek’s own claim about its costs. DeepSeek has said it cost $6 million for the final training run. That may be right given the size of the model and current compute costs in China, which are low. This does not include the costs of data acquisition, prior research, experiments on model architectures, algorithms, and the salaries of the people involved. All-in cost estimates start at $100 million at least. The final model run is the cheapest part. You do that after you know what works, and all the data, infrastructure, and modeling choices have been made. That said—the $6 million figure is still a big improvement. However, such efficiency improvements are to be expected because of advances in both AI software and hardware over time. For example, Claude 3.5 Sonnet, which was released 15 months later than the original GPT-4, is about 10x cheaper per call than the original GPT-4 even though it is a far better model. The cost of DeepSeek’s model is in line with what industry insiders expected given its performance and time of release. The only surprise is that the model came out of China. One can also see the significant efficiency improvements in Google’s recently released Gemini Flash 2 model.
There is plausible reason to believe DeepSeek’s claims themselves may not be entirely accurate because they cannot admit to having access to Nvidia’s H100 AI chip (currently under US export controls), which many believe that they do.
Still, it appears DeepSeek made great progress on efficiency in using chips.
The software stacks of all the chip makers, including Nvidia’s, are poorly optimized and often buggy. Nvidia’s software is better than that of competing firms and is one of the key reasons for their current moat. The labs write their own kernels to access the hardware and do not rely on Nvidia’s CUDA software stack directly. DeepSeek shows that the labs could do more here. Since DeepSeek released their previous model in December, people at AI labs everywhere have been looking into how to achieve the efficiency gains that DeepSeek has been able to. They have revealed publicly many of their efficiency innovations so others can build on their findings.
If more investment goes into improving the software layer on top of chip hardware, this may weaken Nvidia’s hold. For example, Amazon’s Trainium2 chip is good at the hardware level, but evaluations in December showed that the software remains about a year away from being competitive with Nvidia’s. Significant work has already been going into improving the software for Trainium2 chips, but the DeepSeek results may push people to go faster and seek out greater efficiency gains. A similar story may play out with AMD’s latest neural processing unit, etc.
DeepSeek appears to be overoptimized, meaning its success may be overstated.
As more research is done, the initial reporting on public benchmarks seems to show DeepSeek is more fit for the tests than the best models. Benchmarks are appearing that show that the DeepSeek-R1 model’s performance may deteriorate more than that of competing models if one changes the reasoning questions in benchmarks a little. This is expected, given that there are at best only three labs that don’t simply optimize against the benchmarks. There is also evidence, although it is not conclusive, that DeepSeek trained on the output of OpenAI’s o1 model. But the main point remains—the open-source models are behind but not by that much.
To sum it all up, DeepSeek’s progress is big news, but not bad news, for most of the AI ecosystem.
It accelerates the path to useful agentic AI. It has marginally shrunk our view of the time to what we call the “Barnes & Noble moment,” meaning the moment when a serious competitor to a non-tech leader disrupts the business as Amazon did in the late ’90s to Barnes & Noble. That is the moment when AI adoption becomes as existential to everyone as it is today for Google and Microsoft. It is then that we expect the true bubble to manifest.
This research paper is prepared by and is the property of Bridgewater Associates, LP and is circulated for informational and educational purposes only. There is no consideration given to the specific investment needs, objectives, or tolerances of any of the recipients. Additionally, Bridgewater's actual investment positions may, and often will, vary from its conclusions discussed herein based on any number of factors, such as client investment restrictions, portfolio rebalancing and transactions costs, among others. Recipients should consult their own advisors, including tax advisors, before making any investment decision. This material is for informational and educational purposes only and is not an offer to sell or the solicitation of an offer to buy the securities or other instruments mentioned. Any such offering will be made pursuant to a definitive offering memorandum. This material does not constitute a personal recommendation or take into account the particular investment objectives, financial situations, or needs of individual investors which are necessary considerations before making any investment decision. Investors should consider whether any advice or recommendation in this research is suitable for their particular circumstances and, where appropriate, seek professional advice, including legal, tax, accounting, investment, or other advice. No discussion with respect to specific companies should be considered a recommendation to purchase or sell any particular investment. The companies discussed should not be taken to represent holdings in any Bridgewater strategy. It should not be assumed that any of the companies discussed were or will be profitable, or that recommendations made in the future will be profitable. Greg Jensen is an investor in Anthropic in a personal capacity. Any mention of specific companies is for general informational purposes only and should not be considered investment advice.
The information provided herein is not intended to provide a sufficient basis on which to make an investment decision and investment decisions should not be based on illustrative information that has inherent limitations. Bridgewater makes no representation that any account will or is likely to achieve returns similar to those shown. The price and value of the investments referred to in this research and the income therefrom may fluctuate. Every investment involves risk and in volatile or uncertain market conditions, significant variations in the value or return on that investment may occur. Investments in hedge funds are complex, speculative and carry a high degree of risk, including the risk of a complete loss of an investor’s entire investment. Past performance is not a guide to future performance, future returns are not guaranteed, and a complete loss of original capital may occur. Certain transactions, including those involving leverage, futures, options, and other derivatives, give rise to substantial risk and are not suitable for all investors. Fluctuations in exchange rates could have material adverse effects on the value or price of, or income derived from, certain investments.
Bridgewater research utilizes data and information from public, private, and internal sources, including data from actual Bridgewater trades. Sources include Absolute Strategy Research, BCA, Bloomberg Finance L.P., Bond Radar, Candeal, CEIC Data Company Ltd., Ceras Analytics, China Bull Research, Clarus Financial Technology, CLS Processing Solutions, Conference Board of Canada, Consensus Economics Inc., DTCC Data Repository, Ecoanalitica, Empirical Research Partners, Energy Aspects Corp, Entis (Axioma Qontigo Simcorp), Enverus, EPFR Global, Eurasia Group, Evercore ISI, FactSet Research Systems, Fastmarkets Global Limited, The Financial Times Limited, Finaeon, Inc., FINRA, GaveKal Research Ltd., GlobalSource Partners, Gordon Haskett Research Associates, Harvard Business Review, Haver Analytics, Inc., Institutional Shareholder Services (ISS), Insync Analytics, The Investment Funds Institute of Canada, ICE Derived Data (UK), Investment Company Institute, International Institute of Finance, JP Morgan, JSTA Advisors, LSEG Data and Analytics, MarketAxess, Medley Global Advisors, Metals Focus Ltd, MSCI, Inc., National Bureau of Economic Research, Neudata, Organisation for Economic Cooperation and Development, Pensions & Investments Research Center, Pitchbook, Political Alpha, Renaissance Capital Research, Rhodium Group, RP Data, Rubinson Research, Rystad Energy, S&P Global Market Intelligence, Sentix GmbH, SGH Macro, Shanghai Metals Market, Smart Insider Ltd., Sustainalytics, Swaps Monitor, Tradeweb, United Nations, US Department of Commerce, Visible Alpha, Wells Bay, Wind Financial Information LLC, With Intelligence, Wood Mackenzie Limited, World Bureau of Metal Statistics, World Economic Forum, YieldBook. While we consider information from external sources to be reliable, we do not assume responsibility for its accuracy. Data leveraged from third-party providers, related to financial and non-financial characteristics, may not be accurate or complete. The data and factors that Bridgewater considers within its investment process may change over time.
This information is not directed at or intended for distribution to or use by any person or entity located in any jurisdiction where such distribution, publication, availability, or use would be contrary to applicable law or regulation, or which would subject Bridgewater to any registration or licensing requirements within such jurisdiction. No part of this material may be (i) copied, photocopied, or duplicated in any form by any means or (ii) redistributed without the prior written consent of Bridgewater® Associates, LP.
The views expressed herein are solely those of Bridgewater as of the date of this report and are subject to change without notice. Bridgewater may have a significant financial interest in one or more of the positions and/or securities or derivatives discussed. Those responsible for preparing this report receive compensation based upon various factors, including, among other things, the quality of their work and firm revenues.