Demystifying DeepSeek: Five things every Advisor should know

#1: Why did markets react so sharply to the DeepSeek news?

The market’s reaction was sharp, but it wasn’t immediate—DeepSeek had released an earlier version of their model in December, which industry analysts were already evaluating. It wasn’t until people in finance started to test out the model themselves that a consensus was reached: DeepSeek was disruptive and had the potential to compete with the best U.S. A.I. apps. That narrative—which was overly negative, in my view—spread rapidly the weekend before the selloff, resulting in the Monday plunge.

Importantly, it was not a widespread selloff in Technology—the drop was largely concentrated in Nvidia and other semiconductor companies like Broadcom, which are heavily exposed to A.I. capex spending. The magnitude of the overall market decline was a function of Nvidia and Broadcom’s market capitalization—if, for instance, Nvidia makes up 10% of the market, then a 17% stock decline translates to an almost 2% overall market tumble. It was concentrated losses in those two names that really drove the broader market down.

Nasdaq-100 January 2025 Performance vs. Megacaps

Source: Nasdaq, Bloomberg. Data as of January 31, 2025.

#2: Is it realistic that the DeepSeek model was trained for approximately $6 million USD as claimed?

Some of the headline numbers that have been reported are misleading. The supposed $6 million training cost was calculated by multiplying the 3 million graphics processing unit (GPU) hours required to train DeepSeek’s V3 model by a cost of around $2 per GPU hour (3 million hours x $2 per hour = $6 million). But that was only the cost of training this particular version of the DeepSeek model—it does not include work done on previous versions. As such, we can state definitively that the overall training cost was more than $6 million USD, as that figure does not encompass the several years of development work that preceded the V3 model.

There is still much we don’t know about DeepSeek’s development process, including whether it benefitted from the unauthorized use of competitors’ data. One thing we do know, however, is that they’ve been at this for years. In 2021, prior to the U.S. government’s ban on the export of certain Nvidia chips to China, DeepSeek acquired about 2,000 leading-edge Nvidia processors. It’s also been speculated that they currently have at least 10,000, and perhaps as many as 50,000, Nvidia chips at their disposal, which they cannot admit publicly since the company would have had to go around the U.S.’s export restrictions to acquire them. Overall, given the cost of high-end hardware and the capital expenditures associated with setting up data centres, maintenance, and so on, some experts are estimating that the real all-in cost of building DeepSeek’s models was closer to $2 billion USD.

#3: How does DeepSeek change the global A.I. landscape?

In recent years, the budgets for A.I. development in the U.S. have been virtually limitless because they’ve been largely funded by some of the biggest companies on the planet, including Microsoft, Google, Meta Platforms, and Amazon. They are all committed to winning the ‘A.I. race’ and don’t have to worry as much about export limitations or massive capital expenditures because of their enviable financial resources.

Prior to DeepSeek’s sudden emergence, the assumption was that the leading A.I. models in the U.S. were ahead of the leading models in China. What caught analysts by surprise was the new efficiencies that DeepSeek was able to incorporate in the model’s design and training, resulting in a model comparable to the leading U.S. players at what may have been a lower price point than a model like OpenAI’s ChatGPT. As mentioned earlier, we don’t know for certain how much less expensive DeepSeek’s development and training process may have been compared to something like ChatGPT. But it is likely that the leading U.S. developers will begin to implement some of these new efficiencies, if they haven’t already. That means that training A.I. models could be cheaper and faster in the future—which would be a tailwind for everyone.

#4: Ultimately, is this good news for Tech and A.I.?

The upside for many of the top A.I. players is that the adoption of some of the architectural improvements pioneered by DeepSeek could lower the cost of training A.I. models going forward. In fact, OpenAI was already in the process of independently developing and introducing efficiencies to make models more inference compute-heavy rather than training-heavy. What does that mean in layman’s terms? For the past few years, the A.I. race has primarily been about using the best hardware and the most data—the entire vastness of human knowledge, essentially—to train the most expensive models. What OpenAI and some other A.I. companies are now doing is trying to find a balance between all of that up-front training, which is very expensive, and a more “human” way of thinking through problems. In this newer, inference compute-heavy approach, a well-trained foundational model is able to apply reason to questions, simultaneously evaluating and cross-checking potential answers rather than simply relying on a massive knowledge database. This shift was already underway and has only been accelerated by DeepSeek.

The shift to an inference computing approach could be bullish news for players like Nvidia, because powerful GPUs will still be required to do these kinds of calculations on an ongoing basis. The question is: will the timing of revenue and earnings growth look different if less money is being spent up front on intensive model training, and more is spent further down the road as A.I. is adopted more broadly and demand rises? In the long term, greater demand for A.I. will mean greater demand for inference computing, which is a bullish indicator for chipmakers like Nvidia. Companies like Microsoft, Amazon, and Google also fund or operate leading cloud computing platforms, which is where much of this inference computing will be happening, so these developments could be bullish for them as well. Apple is another company that could benefit, as more efficient and competent A.I. models will be able to be run on devices like iPhones and MacBooks without using up all of the devices’ resources.

#5: Knowing what we do now, how can investors play the A.I. theme?

We believe the Nasdaq-100^®, which is accessible via the BMO Nasdaq 100 Equity ETF Fund, is well-suited for this unique market environment due to its long-running leadership in innovation, as well as its demonstrated ability to outperform in volatile markets.

Most notably, this was evident in March 2020 during the early days of the pandemic, when the Nasdaq-100 had the smallest drawdown and quickest recovery across the three major U.S. benchmarks, outperforming both the S&P 500 and the Dow Jones Industrial Average. It happened again in March 2023 during the U.S. regional banking crisis, as the Nasdaq-100 outperformed the Dow Jones and S&P 500 by five-to-seven percent after the crisis had reached its climax.¹ In that instance, the index was helped by the fact that, by design, it has no exposure to Financials.

Historically, Tech has been viewed as more volatile, more speculative, and riskier than broad markets. But that perspective is rooted in an era when Tech was not yet central to the economy the way it is now. Back then, Tech companies just didn’t look as healthy from a fundamentals standpoint as they do currently. We believe that the Nasdaq-100 is well-positioned for the next several years, especially if interest rates stay higher for longer or if inflation picks back up. That’s because roughly half of the index has either no debt or enough cash on their balance sheets to earn more in interest income than they spend on servicing debt. High rates, therefore, are accretive for the index, which is not something you could have said 10 or 20 years ago.

Nasdaq-100 vs. S&P 500: Significantly less leveraged

Source: Nasdaq, Bloomberg. Data as of November 29, 2025.

*Management Expense Ratio (MER) is as of September 30, 2024.

BMO Global Asset Management, as of January 31, 2025.

Please contact your BMO Global Asset Management wholesaler for any additional support and guidance.

¹ Source: Nasdaq, Bloomberg.

² All investments involve risk. The value of a Mutual Fund can go down as well as up and you could lose money. The risk of a Mutual Fund is rated based on the volatility of the Mutual Fund’s returns using the standardized risk classification methodology mandated by the Canadian Securities Administrators. Historical volatility doesn’t tell you how volatile a Mutual Fund will be in the future. A Mutual Fund with a risk rating of “low” can still lose money. For more information about the risk rating and specific risks that can affect a Mutual Fund’s returns, see the BMO Mutual Fund’s simplified prospectus.