The AI Revolution and DeekSeek
The AI Revolution and DeepSeek
In this report, we will provide a brief background on the advancements of AI, the role of DeepSeek as a potential disruptor, and what this means for the markets and select industries.
Background:
Artificial Intelligence (AI) has been around since the dawn of the computer era. As computing power became cheaper and faster, companies increasingly used machine learning to improve operations and execution of their businesses The breakthrough to broad adoption was the introduction of large language models (LLMs) when OpenAI launched ChatGPT at the end of 2022. GPT stands for Generative Pre[1]trained Transformer and uses a variety of applications to understand and produce human language. This was a huge breakthrough versus earlier models requiring users to essentially be computer scientists. As the markets recognized the potential and companies started to invest in AI infrastructure, several stocks moved higher on expectations of large windfalls related to AI exposure. While US tech stocks dominated these gains, other companies also benefited as it became clear the adoption of AI would require the construction of large data centers that would consume vast amounts of electricity. The net result being strong gains in the US S&P 500 index for both 2023 and 2024 as multiples expanded on higher growth expectations. As of January 24, 2025 the S&P 500 Index closed at a P/E of 22.1x on next twelve months earnings expectations (NTM) versus a 10-year average of 18.2x , which benefited from much lower interest rates that led to TINA (there is no alternative) for most of those years. Importantly, at the end of last year, the top 10 stocks in the S&P 500 traded at a 64% premium to the remaining 490 stocks according to JP Morgan, largely due to elevated expectations from AI. The conventional wisdom was that US technology stocks dominated AI and this justified the large premium of US equity markets relative to the rest of the world as these stocks would reap the biggest economic rewards from AI as they built large moats around their businesses by deep-pocketed spending on infrastructure. However, many market observers had become concerned about valuations given very lofty expectations.
Enter DeepSeek: In 2022, before ChatGPT, technology stocks drove the US stock markets down as increased spending lowered free cash flow (FCF) and returns, leading the companies to cut capital spending and slash costs in 2023. However, with the launch of ChatGPT and as interest in AI accelerated, the market started rewarding higher spending. For 2025, the projected capital spending for Alphabet, Amazon, Meta, and Microsoft is roughly $300 billion, or twice the 2022 spending that triggered stock corrections. What can go wrong? Enter DeepSeek, the ChatGPT disrupter. According to reports, a Chinese AI startup called DeepSeek , that is just over one year old, built AI models that oƯered comparable performance to the best chatbots. The kicker was DeepSeek claims it used older generation Nvidia chips costing $5.6 million versus the best chatbots using the latest generation Nvidia chips costing ~$100 million. Immediately, this triggered questions about the value of the large spending of the US technology companies and what return on investments (ROI) investors might get from this spending. Thus far, investors have been willing to accept a lower FCF generation in the short term on expectations that future FCF would be substantially higher as the large investments generate strong results. However, if a young Chinese AI start-up can create a competitive LLM at a fraction of the costs, it undermines the foundation of US tech leadership in AI and the US companies’ ability to create big moats around these businesses. As a result, tech/growth stocks corrected, some more violently than others, wiping out ~$1 trillion on Wall Street on one January 27th.
What do we know about DeepSeek and its capabilities?
DeepSeek was founded by Chinese national Lian Wenfeng, who first founded a hedge fund that used AI for trading. For DeepSeek, he hired mostly fresh graduates with a mission to seek out diƯerent ways to do things. Rather than scanning “the entire internet” for answers to a question, DeepSeek uses short cuts to find the right areas to hunt for the answers. The approach used by US LLMs requires vast data processing capabilities and is power hungry, raising demand for high-end computer chips, data centers and electricity. DeepSeek undermines this approach as it requires less overall processing and power at much lower costs. In essence it is a distilled version of the prior LLMs, but critically it used the information from the prior LLMs to get down to a slimmer version. This raises the question how dependent was the model on prior LLM’s? Already OpenAI claims DeepSeek violated their terms of service when using their LLMs by unlawfully exfiltrating large amounts of data to train the DeepSeek LLM. Furthermore, if all models slim down, will future versions degrade if newer “foundation models” are not being built? On the other hand, if the claims are true, this is a great example of “Creative Destruction” coined by Austrian economist Schumpeter. However, DeepSeek has limitations, especially when questioned about China, China’s Communist Party, its leaders, and politics. This could be a different concern as answers on issues outside of China could also become colored by the views of the country’s political leadership. On the other hand, the lower cost of the DeepSeek chatbot could lead to much quicker adoption of LLMs globally accelerating the benefits from AI, but also the risks from AI such as lost freedoms due to surveillance, identity fraud, etc.
While the market reacted violently initially to the claims by DeepSeek, others have questions. For example, the company provides the costs of the Nvidia chips they used, but some wondered if these were the actual chips used and if they used more advanced, expensive chips. Furthermore, were the cost claims the actual costs, or did the Chinese government provide significant “foundation research” for free since the costs of LLMs are more than just computer chips. Another question is the timing of the release just as the US has changed administration with President Trump critical of China and threatening strong tariƯs on imports from China. Furthermore, should the US dominate AI while putting severe restrictions on exports of high-end computer chips to China, it will also allow the US military to hold a substantial technological edge on the Chinese military. DeepSeek could blunt this and could be seen as hand grenade thrown to undermine US confidence in its tech/AI superiority. Finally, other companies/countries are trying to build out their own AI capabilities/LLMs but have been limited by steep costs. DeepSeek is a seismic shift in their favor if the costs and technological claims are true and DeepSeek used older Nvidia chips not subjected to export controls.
What does this mean for the markets and industries?
Mark Andreessen of Netscape fame and Venture-Capital founder called this a Sputnik moment, and we agree as it will have large implications for market expectations for select industries. From a big picture perspective, though, the much lower costs, if validated, could substantially accelerate the adoption of LLMs with the consumers being the biggest winners as they see much more value in using the models once costs are less of a hurdle. Ultimately, competition is good for the consumer.
Tech and Communications:
US tech companies have dominated global tech since the early 1990’s, led the internet revolution, smart phone revolution, and were expected to dominate the future AI revolution. If the claims are substantiated, DeepSeek has potentially pulled the rug under the US companies AI feet lowering the money barriers to entry substantially. This will undermine the companies spending on capex to create large moats around their businesses if future LLMs require much less energy-intensive data processing, less needs for data centers, cloud storage, and electricity consumption. Hence, companies must re-evaluate spending plans and investors must question the assumptions and ROI behind these investments. Ultimately, future top line growth might, emphasize might, be less than previously expected along with smaller margins leading to lower future earnings and FCF. This will negatively impact multiples, leading to a correction in stock prices, but only if the claims are true. Alternatively, it could become a teaching moment where the industry uses the information from DeepSeek to build their own, slimmer LLMs at lower costs, allowing for earlier and more widespread adoption. This would support strong revenue growth, and earnings expectations, aiding stock prices.
Utilities and Industrials:
Utilities have seen their fundamentals change in the last few years. Between 2005 and 2023, electricity demand in the US grew 11.4%, or a CAGR of 0.6%. With the prospect of a massive build out of power[1]hungry data centers and electrification of the car fleet, electricity demand was expected to see a substantial shift upward. To meet future demand, the utility industry would have to build out generating and transmission capacity. Since most of the demand would go directly to future data centers, the utility companies could earn higher returns as their company clients will be willing to pay a higher price for reliable power than traditional consumers through regulated businesses. As a result, utilities with large non-regulated generating assets and independent power producers (IPPs) have seen earnings expectations rise as they have announced expansion plans supported by long-term oƯtake contracts, including bringing back mothballed nuclear plants. However, if the need for data centers will be much less due to DeepSeek, the demand for electricity will also be less. This will lead to reduced earnings expectations for certain utility companies, likely leading to a correction in their stock prices. However, it is a big if. For industrials, the issue is the same as for utilities as these companies would do the construction and buildout of data centers and power lines leading to higher future earnings growth. Similarly, this is now being questioned leading to potentially lower future expectations and share prices.
Energy Companies:
Another beneficiary of the buildout of data centers would be energy companies to feed electricity generating assets. While some of the electricity can come from renewables, they are unfortunately intermittent and do not work when the sun does not shine, or the wind does not blow. Data centers require reliable energy with 100% uptime. The best and cleanest source is nuclear, but this is mostly tapped out, so the second alternative is natural gas. However, the production of natural gas might not be exactly where the data centers are located so pipelines for natural gas will be required. Already, the US has an extensive natural gas network, but future demand will require an expansion of the network leading to growth for the midstream companies as they add capacity. Furthermore, US natural gas producers should also benefit as the increased demand for natural gas should lead to higher prices and margins. Already, the demand for US natural gas is set to increase from higher LNG exports as more plants are completed over the next three years. Increased demand from data centers adds to this demand growth.
REITs:
A few REITs specialize in building and managing data centers and clearly this is a big opportunity for them that could be reduced.
Final Observations:
The US stock market has performed exceptionally well in the last two years, led by companies linked to AI either directly (tech and communications) or indirectly. This has led to lofty expectations for many stocks. As a result, we are not surprised by the correction and volatility. History tells us the market usually observes a 10% intra-year correction every year. The trigger this time was the surprise findings by DeepSeek, but the current correction has been rather small except for a dozen or so high-multiple stocks. We find this healthy. However, the arrival of DeepSeek will have a longer lasting impact on expectations and valuations if the current claims are correct.
Another dimension is geopolitical. In Cold War 2.0, the US and Europe see China as a bigger threat than Russia. During Cold War 1.0, the Soviet Union posed a military threat to the West, but never an economic/manufacturing threat. Cold war 2.0 is diƯerent since the West has outsourced much of its manufacturing to China and is much more dependent on trade with China. As a result, a potential conflict could be extremely harmful economically even if military clashes are minor. If China has, or will, achieve parity in AI with the US/West, the probability of a conflict rises as China could see themselves less dependent on the West than the West is on China for many key items, including health care products and medicines. To prevent China achieving parity on AI, the US could put in place much more severe export restrictions as the DeepSeek LLM relied heavily on western technology. While this could ensure the US keeps the edge in AI, it would also negatively impact sales for technology companies given the importance of the Chinese market.
Conclusion:
If what we know about DeepSeek is true, and its chatbot can compete with the best of the world’s leading chatbots at a fraction of the cost while consuming less electricity, it will dramatically change the competitive landscape for US technology companies. It will also change the outlook for those companies that stand to benefit from higher spending on infrastructure, if DeepSeek results in lower spending. Instead of viewing spending as a barrier to entry, investors will ask tougher questions demanding a roadmap for how the spending will improve ROI and FCF, rather than being dilutive. In the end, this will impact valuations of both tech companies and those dependent on the spending from tech companies. For now, more rigorous testing and replications of the findings by DeepSeek are required to verify the claims, especially the costs. If replicated, expect more volatility in the future based on this news.