But the reason is just that the development of this technology is very expensive. I and my co-founders at Anthropic were among the first to document what we call the scaling laws for AI systems. The scaling laws show that when you pour more compute into AI systems, you get better performance across a wide range of cognitive tasks. That's gone from stuff that was really simple and primitive back in 2019 to today where models are climbing the curve of how well they can do at coding. As the models get better, the returns get better and better. Every new model has cost more, but also generated much more revenue than the previous model.
So OpenAI, Oracle, and SoftBank just announced a $500 billion joint venture with the backing of the Trump administration called Project Stargate. How do you interpret that announcement and what was visually being presented there with Amazon next to Sam Altman, next to Larry Ellison, all behind Donald Trump? And should we expect a similar announcement, less by Trump maybe, from Anthropic in the future?
Speaking to the policy aspects of this, one thing I've been saying is about the increasing power of AI systems that has so much promise but also presents risks. We have a number of ideas that we've advocated for. I've gone to the Hill, to the previous administration, and to the incoming administration to talk about the importance of that. One topic is energy provision, providing that energy inside the United States so we don't have to go to countries we trust less for critical security infrastructure around AI systems. Provisioning more energy is a good thing from that perspective. On the particular project, it was a bit hard to make heads or tails of it. It seemed a bit chaotic. It's not clear how much money is actually involved, how much was committed, or which company actually had the money. It was presented along with the Trump administration, but it's not clear what the role of the government actually was.
Well, just to follow up, do you worry that the government here is picking a winner and could advantage OpenAI?
This has always been our perspective. If we do that, there will be enough of the right things to go around. If we make ourselves substantively helpful, things will work themselves out. Some players may be more flashy about it or see it as more of a political thing than others. We work with our cloud partners, and for many years we've been thinking about how to scale. For years I've been saying $100 million in 2023, $1 billion in 2024, $10 billion, $100 billion. So we've been planning accordingly.
Well, let's keep talking about policy, because one of the first things that President Trump has done is to cancel the previous administration's AI executive order. What do you make of that and what should replace it?
First, we should get the facts straight. There's been a lot of this totemic thing around the executive order. It was actually a relatively simple thing that didn't do much. What it did do I thought was good, but it didn't do much. The main thing it introduced from the perspective of frontier model development was reporting requirements on models above a certain scale—basically, how much compute did I put into it, what kinds of testing did I do. All completely voluntary. I think it used up probably one day's time of one person at Anthropic per year to comply. So I think it was a good thing, but also not a big deal that it's been taken away. Things that are much more important are the Biden administration's export controls on chips going to China.
Dario, your essay, you talk about the potential for this technology to accelerate life sciences and do incredible things. But there's also the geopolitical risk. How should we think about this moment?
My view, and I've been saying over the last few days, is that I'm becoming more confident that we may be only two or three years away from AI systems being better than humans at almost all tasks. I've called this in my essay 'Machines of Loving Grace' a country of geniuses in a data center. And if you have that, you can do a number of extremely positive things: curing diseases, extending the human lifespan. I said in that essay I wouldn't be surprised if after having these models for a few years, we were able to double the human lifespan. That sounds extreme, but as a former biologist, I think it's within the realm of possibility. But along with all these great things come risks. These systems can do anything you could do on a computer, and eventually build robots and project themselves into the physical world. There are questions about the balance of power. Our economic system, military system, and security system are based on assumptions about how power is balanced. The risks I see are all around ways that balance could be disrupted. I worry about the autonomous risk of AI systems—the idea that we make highly autonomous systems smarter than a Nobel Prize winner that act over many days. Are they doing what we want them to do? One question is, what are they going to do? Are they a threat to everyone on Earth? And can individuals use them to do dangerous things, like terrorism? And can countries, particularly authoritarian countries, do especially dangerous things with them? Our view on policy is derived from that. We've been big on export controls on chips going to China, because I'm very worried about what an authoritarian country would do with that kind of power. I'm very worried about 1984 scenarios or worse. The energy provision we've talked about is the other side: we don't want the chips to go to China, but we do want to build as much energy as we can here. We want to make this go right so that we have these resources in the United States. Even as we guard against the dangers of authoritarian countries, we have to worry that the systems we ourselves build pose dangers. If we deploy one of these systems or if someone steals them, could a single person do much more harm? Going back to Bill Joy's essay 'Why the Future Doesn't Need Us,' he talked about this idea that as technology empowers individuals, a single person could cause immense harm. For that reason, we've always been an advocate of testing and measurement of AI systems for national security risks, as done in the U.S. AI Safety Institute.
On the topic of China and the risk of countries that maybe don't share the US's same values, have you seen the latest DeepSeek model that has taken the world by storm? A very competitive model coming out of a Chinese company. What do you make of it?
It's been reported that DeepSeek has something like 50,000 H100 chips, though you can never be sure what's true or not. Other players are planning to build in the next year high hundreds of thousands of chips in 2025, probably millions in 2026, and potentially tens of millions in 2027. We're at a moment with these new scaling-oriented reasoning models where there's a switch-over point. Because we're switching paradigms, it's possible to make strong models in this new paradigm without a huge amount of compute. At the same time, China has been able to smuggle mid-tens of thousands of chips. I think the export controls are important to maintain a lead. If they get parity, we'll have to think of another strategy for a security equilibrium where several players are at parity. I'm hopeful we can avoid that with strong export controls. I hope this is a bipartisan thing and that the Trump administration continues this. I know there are a number of folks in the Trump admin who understand the logic and are supportive.
