What happens when the addressable market for a groundbreaking technology suddenly seems 1/20th as large as investors previously thought?
That question roiled AI semiconductor stocks on Monday morning following a stunning announcement by DeepSeek, a Chinese tech start-up, which claimed to have developed an impressive large-language AI model requiring a small fraction of the computing power employed by more established competitors, such as OpenAI.
What are the claims?
According to DeepSeek, its latest model was trained on just approximately 2,000 H800 Nvidia GPU chips, at a total cost of around $5 million—roughly 1/20th of the $100 million it reportedly cost to train OpenAI’s models. DeepSeek also claims that its R1 model matches the performance of OpenAI’s o1 model while charging just 3% to 5% of the cost of what OpenAI charges for its premium service.1
Semiconductor investors—suddenly fearing that far less hardware might ultimately be needed to power the AI revolution than previously assumed—took a “sell-first-ask-questions-later” approach: By mid-morning on Monday, shares of Nvidia, Broadcom and Marvell Technology had all fallen at least 10%.
How did DeepSeek do this?
In very basic terms, most large-language models are trained in two steps: (1) supervised learning (teaching the model to recognize patterns in text and images), and (2) reinforcement learning (fine-tuning through feedback on performance).
DeepSeek said it introduced a novel approach to model training by skipping the first step and instead using only reinforcement learning. Based on little prior knowledge, R1 allegedly learned to reason by continuously trying, failing and improving through feedback.
According to DeepSeek, eliminating the first training phase didn’t just cut down on the cost and time required to train R1, it also enhanced the model’s flexibility and reduced the inherent biases that can come with being trained on pre-labeled data.
What does this mean for equity markets?
While this story rapidly evolves and leading AI players scramble to replicate and verify DeepSeek’s claims, we expect significant volatility across the semiconductor sector in the near term as the market seeks to gauge the potential fundamental impact on the broader AI ecosystem.
Meanwhile, updated capex outlooks from Microsoft and Meta continue to be robust, and we believe the recently announced $500 billion Project Stargate—a joint venture among OpenAI, Softbank and Oracle to build out AI infrastructure in the U.S.—should continue to support demand for AI chips and equipment over the coming quarters. At the same time, we note there was already plenty of debate about the overall demand for AI computing power, and DeepSeek’s salvo, in our view, will likely increase the jitters. We expect to get more meaningful color on overall AI investment when large hyper-scalers, including Amazon and Google, report earnings over the coming weeks.
Over the mid-term, we believe demand for advanced “inferencing” algorithms—and the next generation chips to power them, such as the GB200/300—will keep gathering steam with the rise of agentic AI. (For more on this trend, see our recent post The Potential Power of AI Agents.)
All things considered, and acknowledging how rapidly this story is moving, we believe that the AI arms race will continue to support strong demand for semiconductors, especially for more efficient ASIC chips and data center networking technologies.
As a result, we expect the U.S. to continue restricting China’s access to advanced AI chips and semi manufacturing tools, while also seeking to beef up its own AI infrastructure, including Project Stargate.
And don’t forget…
Progress, history reminds us, is rarely linear, and that journey can be painful in patches.
For all the market volatility in the wake of DeepSeek’s announcement, we believe it’s important to point out that, if the company’s claims prove true, its discoveries could have an enormously positive impact on the broader economy—in the form of lower capital costs, wider AI adoption and greater efficiency across many industries.
Furthermore, while mega-cap tech firms have accounted for a vast majority of cap ex associated with AI, the potential arrival of a new breed of more-affordable, open-source models—built on less massive and pricey hardware—could help support a broadening of the overall equity market. Not only do value and non-tech cyclicals appear to be holding up well amid the mega-cap rout, government bonds have also gotten a boost on the promise of broader productivity growth and potential disinflation. (For more details on the potential broadening of the equity market, please see our 1Q2025 Equity Market Outlook.)
Ultimately, we believe AI’s evolution toward lower computing costs and greater efficiency will lay the foundations of a stronger, more resilient global economy—encouraging news, in our view, for long-term investors.