Investors woke up Monday morning to widespread declines in market leaders, most notably Nvidia (NVDA), which opened the day down more than 12% following the emergence of a Chinese AI model called DeepSeek, which seemingly threatens the entire premise of the AI bull market.
The current AI narrative is one of massive data centers run by mega-cap tech companies, driven by Nvidia’s advanced GPUs and proprietary architecture, all powered by newly built or recommissioned nuclear power plants.
Each step along that path has minted sky-high returns for stocks in the respective sectors.
Companies like Microsoft (MSFT), Amazon (AMZN), and Alphabet (GOOG) are or will be building the data centers.
Nvidia will provide the chips, with assists by foundry Taiwan Semi (TSM) using ASML’s (ASML) equipment.
And it will all be powered by nuclear energy from Constellation Energy (CEG) and companies like it.
Of note, CEG is trading lower by 20% on the day, TSM is down by 14%, and ASML is lower by 7%.
[text_ad]
Where DeepSeek disrupts this model, allegedly, is that it was trained at a cost of only $5.5 million, can be run locally (doesn’t require large data centers), and is beating the bleeding-edge models on benchmarks already.
Its open model also means it can be used commercially without restrictions.
The fear, for investors, is that if this is a “Sputnik moment,” as venture capitalist Marc Andreesen called it, and companies using DeepSeek don’t need the associated infrastructure required by existing U.S. models, the capital expenditures in chips, infrastructure, and power are simply wasted.
While I remain skeptical that existing generative AI models will evolve beyond their current capabilities (big-data autocomplete), and are thus worth the investment dollars chasing that, there are two critical concerns with DeepSeek that could render it practically unusable.
2 Reasons to Be Skeptical of DeepSeek
Chinese Data Quality & Collection
The first and most obvious reason to be skeptical of DeepSeek is its country of origin.
The model is subject to the same kinds of controls as the Chinese internet, which raises immediate questions about the reliability of its data.
Users have already found that it will not engage with prompts about subjects such as Tiananmen Square, which begs the question of what else the model may be omitting or ignoring.
Because generative AI models are black boxes (reach conclusions based on obfuscated logic), the quality of the training data is of the utmost importance.
The adage, “Garbage in, garbage out,” summarizes it well.
Regardless of how well-trained a model is, if its starting point is bad or censored data, it will not produce high-quality and reliable output.
Secondary to data quality is data collection by the Chinese government.
Hobbyists may find it acceptable to allow Chinese companies to feasibly collect information on prompts and outputs, but corporate users may be far more reticent.
The obvious workaround, using it locally without internet access, may be feasible, but we’ll need evidence of that.
Synthetic Data & Model Collapse
The biggest potential pitfall is the prospect of something called “model collapse.”
Researchers have found that the use of synthetic data, or input data that is created by AI models, can cause “irreversible defects” that render the models unusable. The link above from IBM offers a highly readable summary, but, in essence, because model-generated data is less diverse and robust, “bad” information gets propagated easily and overwhelms “good” information.
It’s akin to making a copy of a copy or, for Michael Keaton fans, Multiplicity.
This is not a threat that’s unique to DeepSeek, it’s a risk to all generative AI models, but it’s amplified by DeepSeek.
The use of synthetic data is a key component of the model’s framework.
In the event of model collapse, we would expect to see increasingly nonsensical outputs from DeepSeek as it gradually decays to uselessness.
Unfortunately for investors, whether DeepSeek has cracked the model collapse problem is a wait-and-see proposition.
The biggest threat to investors isn’t that DeepSeek usurps existing AI models (at least in my mind, for the reasons above), it’s that DeepSeek is a gut check on the AI narrative as a whole.
If DeepSeek can already do what more robust models do at a fraction of the price, what justifies the investments?
Further, if DeepSeek is the “best” model out there, what happens should it start spewing nonsense?
I already count myself among the ranks of AI skeptics, so this does not materially change how I’m approaching investing these days, but the apparent fragility of the AI narrative should prompt AI bulls to reconsider their theses and potentially take some profit off the table.
[author_ad]
