The Scene
More than two weeks before the recent market meltdown, OpenAI’s head of global policy Chris Lehane told me he thought the success of DeepSeek was the AI industry’s “Sputnik moment.”
It was a dramatic way of describing a group of Chinese researchers who were able to build impressive open-source AI models, despite a lot of headwinds like US export controls on powerful chips. For the better part of a year, sector experts had been watching DeepSeek with admiration for its capabilities relative to its size.
After DeepSeek’s latest model release last week, venture capitalist Marc Andreessen echoed Lehane’s Sputnik line. His words and others’ panicked investors, and drained more than $1 trillion of value from AI firms and related companies Monday.
But that was the opposite reaction that Lehane was advocating a few weeks ago, when he said the US needs to do everything it can to attract the hundreds of billions of global investment dollars earmarked for AI.
In this article:
Reed’s view
Russia’s breakthrough in the ’60s Space Race did not signal the end of spending big dollars on space research. And DeepSeek’s success doesn’t mean spending billions of dollars on AI infrastructure, such as GPUs, is useless or will stop any time soon.
The US DeepSeek freakout is, instead, our greatest mass hallucination since… well… the drone fantasy in New Jersey a few weeks earlier, similarly amplified by heavy-breathing media reports. Axios reported that DeepSeek represented an “extinction-level event for venture capital firms that went all-in on foundational model companies.” New York Times veteran tech reporter Mike Isaac wrote that the tech companies’ “biz model stands to be vaporized along with the value of their shares, and the billions of dollars already invested.”
This hyperbole festered over the weekend, spreading on X and Reddit and a host of other social media platforms until it hit the stock market like a bomb on Monday, sending any public company making big bets on AI tumbling.
Part of the issue was timing. The DeepSeek R1 release happened in the waning hours of the World Economic Forum in Davos, a gathering known for its groupthink, and for being wrong. It also came on the heels of the blockbuster announcement by President Donald Trump that OpenAI, Oracle, and Softbank were teaming up for a potential $500 billion AI infrastructure investment in Texas.
But as the dust settles, the DeepSeek hysteria may reveal more about how little the market understands the AI industry.
Since the launch of ChatGPT, AI researchers have been figuring out creative ways to distill capabilities of larger models down into smaller, more efficient versions. Meta and the French national champion Mistral have been among the best at this game, launching open source models that have been popular in the AI research world and in some industries. Pressure from those providers has helped drive prices down quickly.
Simply put: Massive gains in efficiency are a given. We don’t yet know enough about DeepSeek’s R1 model to say exactly how it will perform in real-world scenarios. But even the most rosy predictions don’t really change the landscape.
That’s because appetite for AI is insatiable, especially in software development where the biggest issue today is “token limits” that get in the way of complex projects. The other bottleneck is capability. We are simply nowhere near a point where AI is good enough for everyday use for average people.
To get there, it will take many more breakthroughs and a lot more investment in infrastructure, like the $500 billion accelerated compute clusters in Texas.
DeepSeek represents an offering — and in the grand scheme of things, a somewhat small one — of some good ideas on how to make AI models more efficient.
Some of those ideas are already being used at top research firms like OpenAI and Poolside, according to conversations with AI researchers and executives at those places.
Step Back
DeepSeek’s success is also being interpreted as meaning China is “catching up” to the US on AI. The truth is, nobody in the AI industry was under any illusion that China wasn’t already close to the US in capabilities.
It’s always been possible that if the US “wins” the AI race, it will cross the finish line only slightly before its rival. But that is a preferable outcome to the opposite one, where China wins by a hair.
What changes now that DeepSeek released its R1 model? Probably nothing, but it could lead to calls for even tighter government restrictions. If DeepSeek was trained on lower-end processors (it’s currently the subject of debate), perhaps even those will be restricted.
The other question is whether open-source technology poses risks. There is a real and ongoing debate over whether the US should curb such developments in order to keep them from getting into China’s hands.
The truth is, these trade secrets are almost impossible to keep in the long run. It’s possible there is already some AI research that is classified and as things progress, there could be some kind of “Manhattan Project-like” program that keeps research papers under wraps until it’s deemed safe to release publicly.
If anything, the DeepSeek conversation shows what kind of research can be done with a relatively small amount of compute power. The US has been debating funding the National AI Research Resource, which would provide compute power to university-level researchers.
Doing so would encourage future DeepSeeks to happen inside the US, benefiting the overall domestic AI ecosystem.
Now What?
In the coming months, potentially even years, the big race is for AI inference, rather than training. Making more efficient AI models is just one part of the inference equation.
As we wrote about last week, the big hyperscalers like Google and Amazon are looking for efficiencies in every part of the technology stack, from chip design to cooling to the interconnects between GPUs.
And then there is the question of supply. Right now, there is simply more AI demand from consumers than anyone can provide.
It’s all about the tokens. Whoever can satisfy the insatiable market desires with the best capabilities at the lowest cost and latency is the winner.
DeepSeek is a free AI model. If it turns out to be the best model for those purposes, the hyperscalers will use it and profit from it. But they will still have to spend billions of dollars to build out the infrastructure to serve it.
As for Nvidia, its stock recovered a bit Tuesday. But plenty of companies are working on ways to dilute its key advantage: Cuda, its proprietary software that practically the entire AI industry is built on.
There’s also the possibility that somebody comes up with a breakthrough that allows AI inference and training to happen in some distributed way that utilizes all the latent compute power in the world that currently goes mostly unused. (That’s in the science fiction category right now.)
The AI foundation model companies like OpenAI and Anthropic need to keep innovating, staying ahead of the competition on both capability and cost. They are currently still ahead, but if someone comes out with more powerful models that are open source and can run efficiently on hyperscaler infrastructure, then they’re in serious trouble.
But even if that unlikely scenario occurs, the demand for that product will still require billions of dollars in GPUs, massive infrastructure investments, and probably more money to further the research.
And then there’s the question: If some company could leapfrog everyone in that way, why would it give away a secret sauce worth billions?