主权AI：为什么各国都在打造自己的模型

This is a massive vulnerability. We’ve got to control our own stack. It’s not just about self-defining the culture, but about controlling the information space. Do we build? Do we partner? What do we do?

The United States in AI right now has world leadership. Instead of colonization, what we have now, I think, is foundation model diplomacy. This big structural revolution is both a threat and an opportunity. And, Guido, I want to talk about sovereign AI, AI, and geopolitics. Let’s start with the news. Our partner Ben is in the Middle East right now. What happened, and why is it so important?

What happened is that the kingdom announced that it’s going to build its own local hyperscaler, or AI platform, called HUMAIN. I think what’s notable is that, as opposed to the status quo of the cloud era, they’re viewing the AI era as one where they’d like the vast majority of AI workloads to run locally.

If you think about the last 20 years, the way the cloud evolved was that the vast majority of cloud infrastructure basically existed in 2 places: China and the United States. The United States ended up being the home for the vast majority of cloud providers serving the rest of the world. That doesn’t seem to be the way AI is playing out, because we have a number of frontier nations that are raising their hands and saying, “We’d like infrastructure independence.”

The idea is that we’d like our own infrastructure that runs our own models, and that gives us the autonomy to build the future of AI independent of any other nation, which is quite a big shift.

I think the headline numbers are somewhere in the range of $100 billion to $250 billion worth of cluster buildout that they’ve announced, of which about 500 megawatts seems to be the atomic unit of these clusters that they’re building. A number of countries, with the kingdom being the most recent, have been announcing what we could think of as sovereign AI clusters. That’s a pretty dramatic shift from the pre-AI era.

I think it’s spot on. Many geopolitical regions are reflecting back on what happened in previous big tech cycles. Wherever the technology is built, and whoever controls the underlying assets, has a tremendous amount of power to shape regulation, shape how this technology is being used, and put themselves in a position for the next wave that comes out of it.

In the Industrial Revolution, having oil was important, and now having data centers is important. I think it’s a very exciting development.

You can often tell why something is important to somebody by the semantics that people use to communicate a new infrastructure project. They’re being called AI factories. They’re not being called AI data centers.

There are 2 ways to respond to that. One train of thought would be, “That’s just branding—the marketing people doing their thing. Under the hood, this is really just data centers with slightly different components.” The opposing view would be, “Actually, no, this is not just marketing.” If you look under the hood—if you X-ray the data center itself—very little of it is the same as it was 20 years ago.

The big difference in active components, of course, is GPUs. Today, if you look at the average 500-megawatt data center and what percentage of the capex required to build—or operate, rather—that data center went to GPUs,

I think we’re also seeing a specialization. The kind of data center you build for a classic, CPU-centric workload and what you build for a high-density AI data center look very different. You need liquid cooling to the rack. You need a very different energy supply, close to a power plant. You want to lock in that energy supply early on.

We’re also seeing a change in consumer behavior. Classically, you want a very full stack that has lots of services to help enterprises build all these applications. We’re seeing more and more enterprises that are actually comfortable with just building on top of a simple Kubernetes abstraction or something similar, and basically cherry-picking a couple of Snowflake- or database-type services on the side to complement that.

I think there’s a new world. The technical components in an AI factory are completely different from those in a traditional data center. Then there’s the question of what it does. Historically, a lot of the workloads that traditional data centers handled were hosted workloads for enterprises or developers, whoever it might be, where most of the data sets and workloads were not particularly opinionated. By “opinionated,” I mean they weren’t necessarily subject to a ton of cultural oversight.

You could argue that was not the case with China, where China wanted full oversight over those workloads. But for the better part of the 2000s, the vast majority of enterprise workloads didn’t need decentralized serving.

What’s different about AI seems to be that these models aren’t just compute infrastructure. They’re cultural infrastructure. They’re trained on data that has a ton of embedded values and cultural norms. That’s the training step.

Then, when you have inference, which is when the models are running, you have all these post-training steps that steer the models to determine whether to say something or not, and whether to refuse the user or not. That last mile is where, over the last year, it has become more and more clear that countries want the ability to control what the factories produce, or don’t produce, within their jurisdiction.

That urgency didn’t exist to the same extent with traditional cloud workloads, because of the cultural factors, or because of concerns around independence or resilience. My sense is that there are 2 things going on, but one is the rise of capabilities in these models. They’re now well beyond what we’d consider the early toy stage of a technology.

You have foundation models literally running in defense, healthcare, and financial services. ChatGPT has about 500 million monthly active users making real decisions in their daily lives. I think that makes a lot of governments say, “Wait a minute. If we are dependent on some other country for the underlying technology on which our military, defense, healthcare, financial services, and daily citizens’ lives are driven, that seems like a critical point of failure.”

I think it’s not just about self-defining the culture, but about controlling the information space. Today, we’re starting to see models replacing search. You no longer go to Google; you go to ChatGPT, and it comes back with an answer. If there’s a historical fact that doesn’t show up in the Chinese model but does show up in the U.S. model, that is the reality that people grow up with.

If you write an essay in school in the future, many of our essays will be graded by an LLM. In fact, in school, something that may be truthful may be graded as wrong because whoever controlled the model decided that it should not be part of the training corpus.

Is your expectation that this is going to play out? To what extent is it going to play out? On the cloud, as we mentioned, there’s the Chinese internet and the Western-rest-of-the-world internet. How widespread is this sovereign AI idea going to become?

If you look at the Industrial Revolution, oil was the foundation of a lot of the technologies. You needed oil reserves in order to participate. I think it will be a little bit the same thing. If you want to build industry in a particular country, export things, drive development, and really harness the power that comes with that, you need the corresponding reserves.

AI data centers are a little bit like these oil reserves, with the big difference being that you can actually construct them yourself if you have the necessary investment dollars and the willpower to do it. But I think they will be the foundations for building all the layers on top that ultimately determine who wins this race.

Talk more about the implications of what this means. Is this something that the United States should be excited about? Are there now winners across the board in all of these local efforts? Talk about some of the big implications here.

This big structural revolution is both a threat and an opportunity. The United States in AI right now has world leadership. That’s an opportunity. Hanging on to it won’t be easy, as it is in every tech revolution.

Don’t we want people to be dependent on us in the same way that they were in the cloud revolution, or do we benefit somehow from it being more decentralized?

The world is not one place, so complete centralization won’t happen. Being the leader is good. Having strong allies that also have that technology is also very valuable. So it’s probably a balance of those that we’re looking for.

We’re clearly in an unstable equilibrium right now, and so Guido’s right that the arc of humanity and history is such that things will probably shake out until there’s a stable equilibrium. The question is, what is the stable equilibrium?

One way to reason about it is to look at historical analogies. After World War II, when Europe was completely decimated, there was a group of really enterprising people in the private sector and the public sector who got together and said, “We can either choose to turn our backs on Europe and adopt a posture of isolationism, where we mostly focus on a postwar, American-only agenda, or we can try to adopt a policy where we know that if we don’t help out our allies, somebody else will.”

Yeah. And so they came up with this idea called the Marshall Plan, right? A number of leading enterprises in the US got together, like GE and General Motors, and literally subsidized the massive reconstruction of Europe. That helped a lot of European economies quickly get back on their feet.

At the time, there was a ton of criticism of the Marshall Plan because it was viewed almost as a net export of capital and resources. But what it did end up doing was solidifying this unbelievable trade corridor between the US and Europe for the next 50 years, which really kept China out of that equation for 70 years. Yeah, 70 years, really.

And so I think we have a choice, which is to either approach it the way we would apply the Marshall Plan for AI, right, and say, “A stable equilibrium is certainly not one where we just turn our back on a bunch of allies, because China definitely has enough compute resources to try to export great models like DeepSeek to the rest of the world. So what do we want our allies on: DeepSeek or Llama?” That’s sort of what it comes down to at the model level of the stack.

The reality is that a number of countries are not waiting around to find out. The ones that certainly have the ability to fund their own sovereign infrastructure are rushing to do it right now. And what does that mean for the nationalization debate, or how do you see that playing out?

You know, Leopold Aschenbrenner, formerly of OpenAI, in his famous report talked about how, if this thing becomes as critical to national security as we think it will be, at some point governments aren’t just going to let private companies run it. They’re going to want to have a much more integrated approach to it. Where do you stand on the likelihood of that, and what does that mean for the feasibility of regulation in a world where it’s much more decentralized?

I probably have a strong opinion on that. I grew up in Germany, right, so benefiting from the Marshall Plan. One lesson I took away from that is that I think any kind of centralized, planned approach does not work.

To some degree, the Eastern Bloc—East Germany versus West Germany—was a nice A/B test: central planning versus a free-market economy. What works better? I think the results speak for themselves.

So I think having the government drive all of AI strategy, whether it’s a Manhattan-style project or an Apollo project—pick your favorite successful project—I can’t see that working. You probably need a highly dynamic ecosystem of a large number of companies competing.

There are some areas where I think the government can have a hugely positive effect. On the research side, we’ve seen that again and again: funding fundamental research that is not quite applied enough yet for enterprises to pick up is very valuable. I think it can help in terms of setting good regulation. Bad regulation can easily torpedo AI, as we’ve seen.

And so I think there’s a strong role for government to lead this and to direct this. There’s no master plan at the end of the day that you can make that has all the details. That has to come from the market.

I don’t agree with Leopold Aschenbrenner’s point of view, actually. The history of centralized planning at the frontier of technology is not great, barring a few situations that were essentially brief sprints of war. Arguably, even the Manhattan Project—which is the analogy I think he uses in his piece—we now know had leaks. It was literally a cordoned-off facility in Los Alamos or wherever, and there were still spies.

For anyone who has had both the pleasure and displeasure of working in any large government system, it’s a pipe dream.

The good and the bad news is that, in a sense, it doesn’t really matter where the model weights are. It matters where the infrastructure that runs the models is. In a sense, inference is almost more important.

A year ago, we were in a pretty rough spot, I would say, with the arc of regulation. There were a number of proposals in the United States to try to regulate the research and development of models versus the misuse of the models. I think that, luckily, we’ve moved on.

Just a year before DeepSeek came out, a number of tech leaders in Washington were testifying that China was 5 to 6 years behind the US, with confidence, on the record. And then DeepSeek came out 26 days after OpenAI put out the frontier model. That just shattered all of those arguments, and the fact that it was an MIT-licensed model meant that every other country had access immediately.

So the calculus has changed, right? I think it means that the only way to win is to build the best technology and out-export anybody else. If the question is whose math the world is using, we’d love for it to be American math.

My view is that we are much better off embracing the ability for other countries to serve their own models. Ideally, the best product wins, and the best models just come from the US and our allies.

Yeah. Is that the new age of LLM diplomacy that we’re entering here?

Ben had a great talking point about this at, I think, FII Riyadh last year. He said something to the effect of: because these models, like we discussed earlier, are cultural infrastructure, you don’t want to be colonized in the digital era, in cyberspace. I think that’s pretty spot-on.

Yeah. Instead of colonization, what we have now, I think, is foundation model diplomacy.