As Labor Day has now passed, it’s safe to say that 2025 will go down as yet another ‘noisy’ year for investors. Things are rarely boring in the stock market, but it was expected that 2025 would at least be a more ‘normal’ year than prior ones. Investors anticipated a good-but-not-too-hot economy, solid corporate earnings growth, a decline in inflation, and interest rate cuts to provide a supportive backdrop for stocks.
Those assumptions aren’t completely out the window now, but investor confidence has been seriously tested in 2025. The most tumultuous period was in April when the S&P 500 fell nearly 20% as higher-than-expected reciprocal tariffs (which were quickly paused/delayed) caused investors to reassess their outlook, to put it mildly.
A less remembered period of market volatility occurred this year in late January, when investors feared a disruption to the AI investment theme. On January 20th, seemingly out of the blue, a little-known Chinese AI company released an AI model called “DeepSeek”. Surprisingly, it displayed performance very close to leading AI models created by American tech companies. The catch? DeepSeek did it with cheaper and fewer GPUs, with the overall cost rumored to be ~90% cheaper than domestic AI models. The implication was huge for investors. Did American large-language models really need the highest-end GPUs, namely from NVDA? If so, how many GPUs do they actually need? Will improved AI training methods hurt demand for chips?
Regarding GPUs, since ChatGPT was unveiled in late 2022, Nvidia has been the GPU leader underpinning the AI revolution. Without their chips, it is/was generally understood that the large language models developed by OpenAI, Google, Anthropic, etc. wouldn’t be possible. Large-scale purchases of Nvidia GPUs were necessary to build out massive data centers that could handle the computing demand. As a result, NVDA’s stock has surged in the past 18 months, leapfrogging Tesla, Meta, Amazon, Google, and Microsoft to become the largest company in the world by early 2025. It quickly lost that title as DeepSeek news hit the stock market. NVDA lost roughly $600b in market value in a single day—a record loss for an individual stock. It stabilized afterward, but NVDA underperformed the market until the April lows. Since then, however, a combination of improved overall market sentiment and reaffirmed GPU spend from all the biggest tech players (MSFT, META, GOOG) propelled NVDA stock to rally 80% and become, once again, the largest company in the world.
How is it possible these tech names are doubling down on their GPU/AI capex given the new information about DeepSeek? Well, the pessimistic answer is that eventually they will, they just haven’t had enough time to digest and adjust capex plans accordingly. A more optimistic answer is that we are seeing an economic phenomenon called “Jevons’ Paradox” playing out. During the Industrial Revolution, it was noticed by British economist William Jevons that even as coal-burning steam engines became more efficient, overall coal consumption still skyrocketed. At first, it seemed irrational (hence the ‘paradox’ name), but in the real world, it makes sense as demand isn’t a static thing. As the cost of something useful decreases, demand increases, and the increased demand actually outstrips the efficiency gains in the first place. English coal consumption is the original example, but modern ones include fuel-efficiency gains in automobiles (gasoline consumption grew up until the 2020 pandemic), cheaper air travel (we use air travel more each year), more efficient LED light bulbs (we use more lights and keep them on longer). Perhaps a new example will turn out to be the increased efficiency of AI models leading to more AI applications, increased overall usage, and continued high demand for GPUs.
To be clear, we don’t yet know what the ‘right’ answer is. We are likely in the early innings of the AI story and these trends take years to play out. What seems clear, though, is that the world is increasingly adopting AI and that the companies that provide this AI are finding more efficient ways to provide it. Regardless of investor outcome, efficiency gains in AI should be seen as a natural evolution in an exciting and new industry. In fact, every transformative industry that has come before has seen similar efficiency gains: railroads, autos, electricity, personal computing/semiconductors, telecom, and air travel.