A little over one month ago on December 25th, a Chinese AI lab little known in the US dropped some new weights, following with the model card and paper the next day.
Four days ago, NVIDIA was worth around $3.6T. At one point today, it was down to $2.9T – still an astronomical sum, to be sure, but a market capitalization representing an almost equally epic market correction.
Just what happened in those thirty-three days?
The Chinese lab, of course, was DeepSeek. Keen observers realized immediately that its qualities aside, the real import were the economics it represented. Per Willison’s numbers, DeepSeek v3 was a model some 40% larger than Meta’s Llama 3.1, but trained on roughly 9% as many GPU hours.
This matters because GPUs are both expensive and difficult to source. The best hardware, in fact, is theoretically unavailable in China because of the United States government’s chip ban. The DeepSeek team responded to this challenge by deeply seeking efficiencies, and apparently found them.
Seven days ago, the DeepSeek team released a new model, R1, a reasoning model comparable to OpenAI’s o1. That’s when things began to move quickly, because not only were DeepSeek’s training costs transformative economically, it was now bumping up against the performance of the best models the US had to offer. And unlike those models, DeepSeek’s were open. All carried the MIT license, which would theoretically make them legitimate open source software, but certain versions were trained on Llama, which is not open source software, and which in turn means that models trained on it cannot be considered open source. Likewise, we don’t have the original training data.
Regardless of whether they meet the technical definition, however, DeepSeek dropped models that were or claimed to be truly open, highly capable and game changers from an efficiency and thus cost standpoint. It took a couple of days for the market to evaluate some of these claims, and while there is much that we still don’t know about the models, engineers who’ve taken them apart in detail have come away impressed – to the point that there have been rumors of near panic within the leaders of those engineers at large, public AI shops on sites like Blind.
Which is why, when the market opened today, the bottom fell out for anything tangentially related to AI, with the NASDAQ closing down 3.1%. NVIDIA was far from the only tech company hammered by the market’s freak out today, but they were the most prominent because they are the proverbial 800 pound gorilla in the market for AI chips. With great success comes great visibility, for better and for worse.
While this decimation will likely prove to be a short term overreaction, and more temperate market corrections should be forthcoming in the coming days, if DeepSeek’s claims continue to be validated, this is an inflection point from an industry standpoint. Many have been examining the higher level industry and geopolitical implications from this news – if you’re a Stratechery subscriber, as one example, Ben Thompson has a good FAQ here – and we’ll all continue to sift through the fallout for weeks and months to come.
One aspect that has not seemed to attract much attention, however, are the implications for enterprise buyers and their relationships with the large, existing model providers such as Anthropic, AWS, Google, Microsoft and OpenAI.
While enterprise AI efforts’ biggest problem to date has not been the technology but understanding where and how best to apply it, they have had two critical concerns with respect to AI’s large, broadly capable foundational models.
- First, and most obviously, is trust. Enterprises recognize that to maximize the benefit from AI, they need to be able to grant access to their own, internal data. Generally speaking, however, they have been unwilling to do this at scale. Vendor promises notwithstanding, one common pattern of adoption has been a limited proof of concept executed on a public model, and then going to production with a privately hosted and managed equivalent that has the data access it requires.
-
Second has been cost. Enterprises have been shocked, in many cases, at the unexpected costs – and unclear returns – from some scale investments in AI. While bringing AI back in house has offered some hope of cost reductions, internal capabilities are expensive and GPUs, as mentioned, have been difficult to acquire. This is presumably why AWS CEO Matt Garman said at reInvent, “On prem data environments are not well suited for AI. They’re just not.”
Enterprises that want to embrace AI, in other words, have reasons to want to do so on their own infrastructure. But that has posed its own set of challenges, challenges which have led many enterprises to scale back their ambitions and turn their eyes from large, expensive foundational models to small, more cost efficient and easily trained alternatives. An approach which has been compelling to users who are employing AI tactically to solve a narrow, discrete problem rather than strategically.
DeepSeek, however, challenges these core assumptions.
- What if enterprises didn’t have to rely on closed, private models for leading edge capabilities?
- What if training costs could be reduced by an order of magnitude or more?
- What if they did not require expensive, state of the art hardware to run their models?
DeepSeek’s most advanced model has been available for seven days. And as stated, there is a great deal of testing and experimentation ahead – and doubtless many enterprises will have concerns for geopolitical reasons about a model trained in China on unknown data sources. But if DeepSeek’s technical and efficiency promises hold, the challenge for AI vendors may not just be for ultimate model supremacy, but for the enterprise market they’ll need to justify their sky high valuations.
Disclosure: AWS, Google and Microsoft are RedMonk customers. Anthropic, DeepSeek, OpenAI and NVIDIA are not currently customers.
