OK, DeepSeek happened. OK, it convulsed stocks. That was then, this is now. If the DeepSeek announcement was an earthquake on Wall Street, where will the rubble fall hardest, and what might be left in a condition not only better than expected but better than before DeepSeek came along? Those are the questions we’ll look at today.
Let’s start with what I think is the most important point here. While there might be a winner in a race to nowhere, the winner can’t expect much out of it. In the world of AI, OpenAI and the whole AGI (artificial generalized intelligence) thing has captivated those looking for entertainment, but they don’t offer any clear path to monetization any time soon. Whether DeepSeek is great at the sort of online generative AI chatbots we hear so much about, then, is kind of irrelevant. What matters? Is it great at something with real potential to generate ROI.
Enterprises have told me from the first that, to them, the value of AI lies mostly in what it might bring to existing applications. Their hot-buttons are largely in the area of business analytics, but you can see that there’s also value in almost any other area, in almost every vertical, to the embedding of expert intelligence in applications. There are big benefits here, so that covers the notion of the “R” in “ROI”.
On the “I” side, we have less clarity. Nobody should seriously believe that an AI expert that takes 800 GPUs to run, ten times that many to train, and needs a river to cool is going to be economical. Entertaining, maybe, but not economical. Yet that’s what the majority of our AI seems to have been directed to producing. I’ve heard AI players bragging about the number of GPUs their model needed, when any CIO would tell them that the goal should be to use fewer rather than more. We could embed simple-model AI agents in everything, and we can’t do that with megamodels.
What DeepSeek has done is illustrate how vulnerable the AI hype wave is to ROI reality, or even a tale that conveys a risk to that reality. If AI needs to be minimalist to succeed broadly, then the “cloud model” of enormous GPU data centers is a losing proposition, and that seems to be what people have been chasing. At least that’s what we’re hearing about. Truth be told, we don’t really know for sure that DeepSeek can even do what it seems to promise, but we can be darn sure that the threat alone is enough to send chills through our current AI thinking, and that almost surely means someone will move to realize the DeepSeek threat in an effort to counter it. This will change AI, and impact everyone in it, so let’s look at each of the elements of AI to assess the impact.
Top of our list is the AI model players, like OpenAI, Google, and Meta. I’ve blogged for some time that inside the AI skunk works of all these players is a set of initiatives that have the same goal as DeepSeek. That’s because I find it impossible to believe that enterprises haven’t shared the same thing with them all as they’ve shared with me. Contained, limited-resource, AI agents have been what enterprises wanted from the first. IBM has been the only player in the whole market to not only recognize but articulate that, but I’m sure it’s been widely known.
Why then have the model players continued to measure success by counting GPUs? Because it generated good ink, publicity, because cloud players were good early sales opportunities, and because supporting a race by selling racing gear is smart even if the race is to nowhere. The real AI opportunity is complicated, and nobody wants to wait to realize benefits in a financial world where only the next earnings report matters.
Now, these guys will have to shift, because VCs and other vendors are likely to go after the DeepSeek opportunity, and so the big model players will have to do the same. The risk is that if they seem to do that, they validate the impact of DeepSeek and may hurt their current position. These companies face a risk, but watch Anthropic in particular because they seem to be doubling down on the AI hype side of the picture.
Next, we have the AI chip players, most notably NVIDIA, and this is the group of players we can expect to see impacted by DeepSeek the most. In fact, most of the market debacle really came down to this space in general, and NVIDIA in particular. However, DeepSeek’s impact isn’t to destroy the space but to populize it, which means there will be winners too.
In the near term, NVIDIA is at risk, because pressure on the public generative AI giants will put near-term chip spending under pressure. NVIDIA actually makes lower-end chips, including ones used by DeepSeek, but they’re lower-profit, than the keystone GPUs the big guys buy. I’m sure they’ll push the lower-end stuff, but they can’t seem to be abandoning their current cash cow.
Not so for players like AMD and Broadcom, and even Intel. If populist AI is the future, then cheap AI chips are essential, and players with no real exposure to the hype side of AI have everything to gain by pushing to get traction, which means pushing for competing low-cost AI models too. Somebody could find gold here, so this is a space to watch. Even NVIDIA’s risk is more near-term than long-term.
How about the software players? IBM, Oracle, Intuit, SAP…you know who I mean. This group could find themselves with nowhere to go but up, but a misstep on the ladder could be very hurtful. IBM, who alone has seen the AI truth, has an uncluttered path upward. Oracle, having bet a lot on AI training via the cloud, needs to shift its focus to enterprise-hosted AI, and if they mess that up they’re inviting trouble. IBM is ramping up its own initiatives there. Intuit has major risks to navigate because AI agents could make new entrants into taxes and bookkeeping practical, and they’ve been raising their prices on all their stuff to boost profit growth. But AI agents could open the door to a simplification of both spaces.
Platform software vendors likely have a net upside here, not only because a more populist version of AI would mean more locally hosted agents, which means more platform tools to host them, but because AI operations support offers a potential differentiator. Operations missions for AI are camel’s-nose opportunities, a way of getting AI into place without navigating a massive transformation. If the AI is agent-based, and requires modest resources, it can expand from any early ops outpost.
For the server vendors, this move is probably neutral-to-positive. It’s not yet clear how much new hardware would be needed for local AI hosting, but I do think that at least the shift to agent AI will promote local hosting of AI in the form of data center clusters.
Network players, notably Cisco, are a mixed bag. Self-hosting AI generates AI-cluster opportunity, but in general, AI in itself is neutral on network impact. However, network operations has the same opportunity for AI integration as IT operations, and it is likely to gain credibility through the linkage with agent AI. This would favor Juniper in the space, since Juniper has the most mature and best AI position.
Realistic AI, I believe, would benefit us more than hyped AI, but just because we’ve proven with DeepSeek that hype AI is vulnerable doesn’t mean it won’t continue on. Here we are on Thursday, and much of the market mess DeepSeek created on Monday seems forgotten. Yes, stocks aren’t back to the level they were on Friday of last week, but the panic is over. But not the impacts.
Wall Street loves bubbles, not just because they make money when the bubble is expanding (what, in Street lingo is called “alpha” or appreciation in value) but because they make money when it bursts, because professionals make money on any change in valuation (“beta” in Street-speak). If I’m right, the Street and both AI prospects and suppliers will have to navigate this same sort of panic again, because what’s driven AI up to now is almost all hype. There is a value proposition for AI, perhaps a powerful one, but that’s not what we’re hearing about now. The question is what, if any, player in the AI space has known this was hype all along and has been working behind the scenes to fund that AI reality. Whoever does emerges strong, and those who don’t will surely suffer.