If the U.S. is to compete effectively in this new speeded-up AI race, it should keep its regulators reined in.
Assuming that the numbers being claimed by the Chinese AI firm, DeepSeek, are accurate, it cost only an estimated $5.6 million to develop one of its latest models. That’s compared with an average for advanced models of around $1 billion.
DeepSeek [had already] said in a late-December report that it used a cluster of more than 2,000 Nvidia chips to train its other V3 model, compared with the tens of thousands of chips that are normally used for training models of a similar size.
As I discussed yesterday, this raises questions about all the billions that have been (or were to be) invested in the computing power then thought necessary to take AI forward. That installed computing power will be put to use, but will it be so valuable? What now happens to Stargate (a planned $500 billion), to Microsoft’s plans to invest $80 billion in data centers this year, to Meta’s plans to spend $65 billion over the same period?
DeepSeek’s model is open-source (even if with a Chinese “core,” which is something to think about).
As the Wall Street Journal notes:
This could help give more small businesses access to AI tools at a fraction of the cost of closed-source models like OpenAI and Anthropic, which Amazon has backed. There are advantages to such closed-source systems, especially for privacy and national security. But open-source can foster more collaboration and experimentation.
To Marc Andreesen, DeepSeek’s RI is “AI’s Sputnik moment.”
Could be. In any event, it may mean, as I mentioned yesterday, a switch where AI’s profit opportunity changes from being the creation of AI computing power to the use that is made of that power. There’s no reason why the U.S. should not do well at this, but the move way from fueling progress with billions of dollars has removed one of the competitive advantages the American sector enjoyed. To use the jargon, the “moat” created by the need for those billions has largely dried up.
Two plusses are that the sudden news from China has shown the pointlessness and counter-productive nature of EU-style AI regulation or the AI rules introduced by the Biden administration (via an executive order that has now been revoked). The emergence of DeepSeek also demonstrates (again) how poorly suited antitrust is suited for a sector in which the dominant can easily become obsolete.
If the U.S. is to compete effectively in this new speeded-up AI race, it should keep its regulators reined in.