When we think of the “cloud,” we often imagine data floating invisibly in the ether. But the reality is far more tangible: the cloud is located in huge buildings called data centers, filled with powerful, energy-hungry computer chips. Those chips, particularly graphics processing units (GPUs), have become a critical piece of infrastructure for the world of AI, as they are required to build and run powerful chatbots like ChatGPT.

[time-brightcove not-tgx=”true”]

As the number of things you can do with AI grows, so does the geopolitical importance of high-end chips—and where they are located in the world. The U.S. and China are competing to amass stockpiles, with Washington enacting sanctions aimed at preventing Beijing from buying the most cutting-edge varieties. But despite the stakes, there is a surprising lack of public data on where exactly the world’s AI chips are located.

A new peer-reviewed paper, shared exclusively with TIME ahead of its publication, aims to fill that gap. “We set out to find: Where is AI?” says Vili Lehdonvirta, the lead author of the paper and a professor at Oxford University’s Internet Institute. Their findings were stark: GPUs are highly concentrated in only 30 countries in the world, with the U.S. and China far out ahead. Much of the world lies in what the authors call “Compute Deserts:” areas where there are no GPUs for hire at all.

The finding has significant implications not only for the next generation of geopolitical competition, but for AI governance—or, which governments have the power to regulate how AI is built and deployed. “If the actual infrastructure that runs the AI, or on which the AI is trained, is on your territory, then you can enforce compliance,” says Lehdonvirta, who is also a professor of technology policy at Aalto University. Countries without jurisdiction over AI infrastructure have fewer legislative choices, he argues, leaving them subjected to a world shaped by others. “This has implications for which countries shape AI development as well as norms around what is good, safe, and beneficial AI,” says Boxi Wu, one of the paper’s authors.

The paper maps the physical locations of “public cloud GPU compute”—essentially, GPU clusters that are accessible for hire via the cloud businesses of major tech companies. But the research has some big limitations: it doesn’t count GPUs that are held by governments, for example, or in the private hands of tech companies for their use alone. And it doesn’t factor in non-GPU varieties of chips that are increasingly being used to train and run advanced AI. Lastly, it doesn’t count individual chips, but rather the number of compute “regions” (or groups of data centers containing those chips) that cloud businesses make available in each country.

Read More: How ‘Friendshoring’ Made Southeast Asia Pivotal to the AI Revolution

That’s not for want of trying. “GPU quantities and especially how they are distributed across [cloud] providers’ regions,” the paper notes, “are treated as highly confidential information.” Even with the paper’s limitations, its authors argue, the research is the closest up-to-date public estimate of where in the world the most advanced AI chips are located—and a good proxy for the elusive bigger picture.

The paper finds that the U.S. and China have by far the most public GPU clusters in the world. China leads the U.S. on the number of GPU-enabled regions overall, however the most advanced GPUs are highly concentrated in the United States. The U.S. has eight “regions” where H100 GPUs—the kind that are the subject of U.S. government sanctions on China—are available to hire. China has none. This does not mean that China has no H100s; it only means that cloud companies say they do not have any H100 GPUs located in China. There is a burgeoning black market in China for the restricted chips, the New York Times reported in August, citing intelligence officials and vendors who said that many millions of dollars worth of chips had been smuggled into China despite the sanctions.

The paper’s authors argue that the world can be divided into three categories: “Compute North,” where the most advanced chips are located; the “Compute South,” which has some older chips suited for running, but not training, AI systems; and “Compute Deserts,” where no chips are available for hire at all. The terms—which overlap to an extent with the fuzzy “Global North” and “Global South” concepts used by some development economists—are just an analogy intended to draw attention to the “global divisions” in AI compute, Lehdonvirta says. 

The risk of chips being so concentrated in rich economies, says Wu, is that countries in the global south may become reliant on AIs developed in the global north without having a say in how they work. 

It “mirrors existing patterns of global inequalities across the so-called Global North and South,” Wu says, and threatens to “entrench the economic, political and technological power of Compute North countries, with implications for Compute South countries’ agency in shaping AI research and development.”

Nearly 200 workers inside Google DeepMind, the company’s AI division, signed a letter calling on the tech giant to drop its contracts with military organizations earlier this year, according to a copy of the document reviewed by TIME and five people with knowledge of the matter. The letter circulated amid growing concerns inside the AI lab that its technology is being sold to militaries engaged in warfare, in what the workers say is a violation of Google’s own AI rules.

[time-brightcove not-tgx=”true”]

The letter is a sign of a growing dispute within Google between at least some workers in its AI division—which has pledged to never work on military technology—and its Cloud business, which has contracts to sell Google services, including AI developed inside DeepMind, to several governments and militaries including those of Israel and the United States. The signatures represent some 5% of DeepMind’s overall headcount—a small portion to be sure, but a significant level of worker unease for an industry where top machine learning talent is in high demand.

The DeepMind letter, dated May 16 of this year, begins by stating that workers are “concerned by recent reports of Google’s contracts with military organizations.” It does not refer to any specific militaries by name—saying “we emphasize that this letter is not about the geopolitics of any particular conflict.” But it links out to an April report in TIME which revealed that Google has a direct contract to supply cloud computing and AI services to the Israeli Military Defense, under a wider contract with Israel called Project Nimbus. The letter also links to other stories alleging that the Israeli military uses AI to carry out mass surveillance and target selection for its bombing campaign in Gaza, and that Israeli weapons firms are required by the government to buy cloud services from Google and Amazon.

Read More: Exclusive: Google Contract Shows Deal With Israel Defense Ministry

“Any involvement with military and weapon manufacturing impacts our position as leaders in ethical and responsible AI, and goes against our mission statement and stated AI Principles,” the letter that circulated inside Google DeepMind says. (Those principles state the company will not pursue applications of AI that are likely to cause “overall harm,” contribute to weapons or other technologies whose “principal purpose or implementation” is to cause injury, or build technologies “whose purpose contravenes widely accepted principles of international law and human rights.”) The letter says its signatories are concerned with “ensuring that Google’s AI Principles are upheld,” and adds: “We believe [DeepMind’s] leadership shares our concerns.”

A Google spokesperson told TIME: “When developing AI technologies and making them available to customers, we comply with our AI Principles, which outline our commitment to developing technology responsibly. We have been very clear that the Nimbus contract is for workloads running on our commercial cloud by Israeli government ministries, who agree to comply with our Terms of Service and Acceptable Use Policy. This work is not directed at highly sensitive, classified, or military workloads relevant to weapons or intelligence services.”

The letter calls on DeepMind’s leaders to investigate allegations that militaries and weapons manufacturers are Google Cloud users; terminate access to DeepMind technology for military users; and set up a new governance body responsible for preventing DeepMind technology from being used by military clients in the future. Three months on from the letter’s circulation, Google has done none of those things, according to four people with knowledge of the matter. “We have received no meaningful response from leadership,” one said, “and we are growing increasingly frustrated.”

When DeepMind was acquired by Google in 2014, the lab’s leaders extracted a major promise from the search giant: that their AI technology would never be used for military or surveillance purposes. For many years the London-based lab operated with a high degree of independence from Google’s California headquarters. But as the AI race heated up, DeepMind was drawn more tightly into Google proper. A bid by the lab’s leaders in 2021 to secure more autonomy failed, and in 2023 it merged with Google’s other AI team—Google Brain—bringing it closer to the heart of the tech giant. An independent ethics board that DeepMind leaders hoped would govern the uses of the AI lab’s technology ultimately met only once, and was soon replaced by an umbrella Google ethics policy: the AI Principles. While those principles promise that Google will not develop AI that is likely to cause “overall harm,” they explicitly allow the company to develop technologies that may cause harm if it concludes “that the benefits substantially outweigh the risks.” And they do not rule out selling Google’s AI to military clients.

As a result, DeepMind technology has been bundled into Google’s Cloud software and sold to militaries and governments, including Israel and its Ministry of Defense. “While DeepMind may have been unhappy to work on military AI or defense contracts in the past, I do think this isn’t really our decision any more,” one DeepMind employee told TIME in April, asking not to be named because they were not authorized to speak publicly. Several Google workers told TIME in April that for privacy reasons, the company has limited insights into government customers’ use of its infrastructure, meaning that it may be difficult, if not impossible, for Google to check if its acceptable use policy—which forbids users from using its products to engage in “violence that can cause death, serious harm, or injury”—is being broken.

Read More: Google Workers Revolt Over $1.2 Billion Contract With Israel

Google says that Project Nimbus, its contract with Israel, is not “directed at highly sensitive, classified, or military workloads relevant to weapons or intelligence services.” But that response “does not deny the allegations that its technology enables any form of violence or enables surveillance violating internationally accepted norms,” according to the letter that circulated within DeepMind in May. Google’s statement on Project Nimbus “is so specifically unspecific that we are all none the wiser on what it actually means,” one of the letter’s signatories told TIME. 

More from TIME

At a DeepMind town hall event in June, executives were asked to respond to the letter, according to three people with knowledge of the matter. DeepMind’s chief operating officer Lila Ibrahim answered the question. She told employees that DeepMind would not design or deploy any AI applications for weaponry or mass surveillance, and that Google Cloud customers were legally bound by the company’s terms of service and acceptable use policy, according to a set of notes taken during the meeting that were reviewed by TIME. Ibrahim added that she was proud of Google’s track record of advancing safe and responsible AI, and that it was the reason she chose to join, and stay at, the company.

The tech industry is currently embroiled in a heated debate over the future of AI: should powerful systems be open-source and freely accessible, or closed and tightly monitored for dangers?

On Tuesday, Meta CEO Mark Zuckerberg fired a salvo into this ongoing battle, publishing not just a new series of powerful AI models, but also a manifesto forcefully advocating for the open-source approach. The document, which was widely praised by venture capitalists and tech leaders like Elon Musk and Jack Dorsey, serves as both a philosophical treatise and a rallying cry for proponents of open-source AI development. It arrives as intensifying global efforts to regulate AI have galvanized resistance from open-source advocates, who see some of those potential laws as threats to innovation and accessibility.

[time-brightcove not-tgx=”true”]

At the heart of Meta’s announcement on Tuesday was the release of its latest generation of Llama large language models, the company’s answer to ChatGPT. The biggest of these new models, Meta claims, is the first open-source large language model to reach the so-called “frontier” of AI capabilities.

Meta has taken on a very different strategy with AI compared to its competitors OpenAI, Google DeepMind and Anthropic. Those companies sell access to their AIs through web browsers or interfaces known as APIs, a strategy that allows them to protect their intellectual property, monitor the use of their models, and bar bad actors from using them. By contrast, Meta has chosen to open-source the “weights,” or the underlying neural networks, of its Llama models—meaning they can be freely downloaded by anybody and run on their own machines. That strategy has put Meta’s competitors under financial pressure, and has won it many fans in the software world. But Meta’s strategy has also been criticized by many in the field of AI safety, who warn that open-sourcing powerful AI models has already led to societal harms like deepfakes, and could in future open a Pandora’s box of worse dangers.

In his manifesto, Zuckerberg argues most of those concerns are unfounded and frames Meta’s strategy as a democratizing force in AI development. “Open-source will ensure that more people around the world have access to the benefits and opportunities of AI, that power isn’t concentrated in the hands of a small number of companies, and that the technology can be deployed more evenly and safely across society,” he writes. “It will make the world more prosperous and safer.” 

But while Zuckerberg’s letter presents Meta as on the side of progress, it is also a deft political move. Recent polling suggests that the American public would welcome laws that restrict the development of potentially-dangerous AI, even if it means hampering some innovation. And several pieces of AI legislation around the world, including the SB1047 bill in California, and the ENFORCE Act in Washington, D.C., would place limits on the kinds of systems that companies like Meta can open-source, due to safety concerns. Many of the venture capitalists and tech CEOs who celebrated Zuckerberg’s letter after its publication have in recent weeks mounted a growing campaign to shape public opinion against regulations that would constrain open-source AI releases. “This letter is part of a broader trend of some Silicon Valley CEOs and venture capitalists refusing to take responsibility for damages their AI technology may cause,” says Andrea Miotti, the executive director of AI safety group Control AI. “Including catastrophic outcomes.”

---

The philosophical underpinnings for Zuckerberg’s commitment to open-source, he writes, stem from his company’s long struggle against Apple, which via its iPhone operating system constrains what Meta can build, and which via its App Store takes a cut of Meta’s revenue. He argues that building an open ecosystem—in which Meta’s models become the industry standard due to their customizability and lack of constraints—will benefit both Meta and those who rely on its models, harming only rent-seeking companies who aim to lock in users. (Critics point out, however, that the Llama models, while more accessible than their competitors, still come with usage restrictions that fall short of true open-source principles.) Zuckerberg also argues that closed AI providers have a business model that relies on selling access to their systems—and suggests that their concerns about the dangers of open-source, including lobbying governments against it, may stem from this conflict of interest.

Addressing worries about safety, Zuckerberg writes that open-source AI will be better at addressing “unintentional” types of harm than the closed alternative, due to the nature of transparent systems being more open to scrutiny and improvement. “Historically, open-source software has been more secure for this reason,” he writes. As for intentional harm, like misuse by bad actors, Zuckerberg argues that “large-scale actors” with high compute resources, like companies and governments, will be able to use their own AI to police “less sophisticated actors” misusing open-source systems. “As long as everyone has access to similar generations of models—which open-source promotes—then governments and institutions with more compute resources will be able to check bad actors with less compute,” he writes.

But “not all ‘large actors’ are benevolent,” says Hamza Tariq Chaudhry, a U.S. policy specialist at the Future of Life Institute, a nonprofit focused on AI risk. “The most authoritarian states will likely repurpose models like Llama to perpetuate their power and commit injustices.” Chaudhry, who is originally from Pakistan, adds: “Coming from the Global South, I am acutely aware that AI-powered cyberattacks, disinformation campaigns and other harms pose a much greater danger to countries with nascent institutions and severe resource constraints, far away from Silicon Valley.”

Zuckerberg’s argument also doesn’t address a central worry held by many people concerned with AI safety: the risk that AI could create an “offense-defense asymmetry,” or in other words strengthen attackers while doing little to strengthen defenders. “Zuckerberg’s statements showcase a concerning disregard for basic security in Meta’s approach to AI,” says Miotti, the director of Control AI. “When dealing with catastrophic dangers, it’s a simple fact that offense needs only to get lucky once, but defense needs to get lucky every time. A virus can spread and kill in days, while deploying a treatment can take years.”

Later in his letter, Zuckerberg addresses other worries that open-source AI will allow China to gain access to the most powerful AI models, potentially harming U.S. national security interests. He says he believes that closing off models “will not work and will only disadvantage the U.S. and its allies.” China is good at espionage, he argues, adding that “most tech companies are far from” the level of security that would prevent China from being able to steal advanced AI model weights. “It seems most likely that a world of only closed models results in a small number of big companies plus our geopolitical adversaries having access to leading models, while startups, universities, and small businesses miss out on opportunities,” he writes. “Plus, constraining American innovation to closed development increases the chance that we don’t lead at all.”

Miotti is unimpressed by the argument. “Zuckerberg admits that advanced AI technology is easily stolen by hostile actors,” he says, “but his solution is to just give it to them for free.”

A large majority of American voters are skeptical of the argument that the U.S. should race ahead to build ever more powerful artificial intelligence, unconstrained by domestic regulations, in an effort to compete with China, according to new polling shared exclusively with TIME.

[time-brightcove not-tgx=”true”]

The findings indicate that American voters disagree with a common narrative levied by the tech industry, in which CEOs and lobbyists have repeatedly argued the U.S. must tread carefully with AI regulation in order to not hand the advantage to their geopolitical rival. And they reveal a startling level of bipartisan consensus on AI policy, with both Republicans and Democrats in support of the government placing some limits on AI development in favor of safety and national security.

According to the poll, 75% of Democrats and 75% of Republicans believe that “taking a careful controlled approach” to AI—by preventing the release of tools that terrorists and foreign adversaries could use against the U.S.—is preferable to “moving forward on AI as fast as possible to be the first country to get extremely powerful AI.” A majority of voters support more stringent security practices at AI companies, and are worried about the risk of China stealing their most powerful models, the poll shows. 

The poll was carried out in late June by the AI Policy Institute (AIPI), a U.S. nonprofit that advocates for “a more cautious path” in AI development. The findings show that 50% of voters believe the U.S. should use its advantage in the AI race to prevent any country from building a powerful AI system, by enforcing “safety restrictions and aggressive testing requirements.” That’s compared to just 23% who believe the U.S. should try to build powerful AI as fast as possible to outpace China and achieve a decisive advantage over Beijing.

The polling also suggests that voters may be broadly skeptical of “open-source” AI, or the view that tech companies should be allowed to release the source code of their powerful AI models. Some technologists argue that open-source AI encourages innovation and reduces the monopoly power of the biggest tech companies. But others say it is a recipe for danger as AI systems grow more powerful and unpredictable. 

“What I perceive from the polling is that stopping AI development is not seen as an option,” says Daniel Colson, the executive director of the AIPI. “But giving industry free rein is also seen as risky. And so there’s the desire for some third way. And when we present that in the polling—that third path, mitigated AI development with guardrails—is the one that people overwhelmingly want.” 

The survey also shows that 63% of American voters think it should be illegal to export powerful AI models to potential U.S. adversaries like China, including 73% of Republicans and 59% of Democrats. Just 14% of voters disagree. 

A sample of 1,040 Americans was interviewed for the survey, which was representative by education levels, gender, race, and the political parties for whom respondents cast their votes in the 2020 presidential election. The margin of error given for the results is 3.4% in both directions.

So far there has been no comprehensive AI regulation in the U.S., with the White House encouraging different government agencies to regulate the technology themselves where it falls under their existing remits. That strategy appears to have been put in jeopardy, however, by a recent Supreme Court ruling that limits the ability of federal agencies to apply broad-brushstroke rules set by Congress to specific, or new, circumstances.

“Congress is so slow to act that there’s a lot of interest in being able to delegate authorities to existing agencies or a new agency, to increase the responsiveness of government” when it comes to AI policy, Colson says. “This [ruling] definitely makes that harder.”

Even if federal AI legislation seems unlikely any time soon, let alone before the 2024 election, recent polling by the AIPI and others suggests that voters aren’t as polarized on AI as they are on other issues facing the nation. Earlier polling by the AIPI found that 75% of Democrats and 80% of Republicans believe that U.S. AI policy should seek to prevent AI from quickly reaching superhuman capabilities. The polls also showed that 83% of Americans believe AI could accidentally cause a catastrophic event, and that 82% prefer slowing down AI development to account for that risk, compared to just 8% who would like to see it accelerated.

Hanging on the wall of Anthropic’s offices in San Francisco in early May, a stone’s throw from the conference room where CEO Dario Amodei would shortly sit for an interview with TIME, was a framed meme. Its single panel showed a giant robot ransacking a burning city. Underneath, the image’s tongue-in-cheek title: Deep learning is hitting a wall. That’s a refrain you often hear from AI skeptics, who claim that rapid progress in artificial intelligence will soon taper off. But in the image, an arrow points to the robot, labeling it “deep learning.” Another points to the devastated city: “wall.”

[time-brightcove not-tgx=”true”]

The cheeky meme sums up the attitude inside Anthropic, an AI company where most employees seem to believe that AI progress isn’t slowing down—and could potentially result in dangerous outcomes for humanity. Amodei, who was on the cover of TIME last month as his company was named one of TIME’s 100 Most Influential Companies of 2024, says Anthropic is devoted to studying cutting-edge AI systems and developing new safety methods. And yet Anthropic is also a leading competitor to OpenAI, releasing powerful tools for use by the public and businesses, in what even Amodei himself has worried might be a perilous race that could end badly.

On June 20, Anthropic fired its latest broadside in that race, releasing the latest version of its Claude chatbot: Claude 3.5 Sonnet. The model, according to the company, sets new industry standards on reasoning, coding and some types of math, beating GPT-4o, the most recent offering from OpenAI. “With today’s launch, we’re taking a step towards what we believe could be a significant shift in how we interact with technology,” Amodei said in a statement on June 20. “Our goal with Claude isn’t to create an incrementally better [large language model] but to develop an AI system that can work alongside people and software in meaningful ways.”

TIME’s cover story about Amodei last month delved into the question of whether Anthropic can succeed at its safety mission while under pressure to compete with OpenAI, as well as tech giants like Microsoft, Google and Amazon—the latter two of whom are significant investors in Anthropic. Today, TIME is publishing a longer excerpt from the interview conducted in early May for that piece. It has been condensed and edited for clarity.

Anthropic is the youngest “frontier” AI lab. It’s the smallest. Its competitors have more cash. And in many ways, it’s the most expressly committed to safety. Do you think of yourselves as underdogs?

Certainly all the facts you cite are true. But it’s becoming less true. We have these big compute deals. We have high single-digit billions in funding. Whereas our competitors have low double-digits to, somewhat frighteningly, triple-digits. So I’m starting to feel a little bit awkward about calling ourselves the underdogs. But relative to the heavyweights in the field, I think it’s absolutely true.

The AI “scaling laws” say that as you train systems with more computing power and data, they become predictably more capable, or powerful. What capabilities have you seen in systems that you’re not able to release yet?

We released our last model relatively recently, so it’s not like there’s a year of unreleased secrets. And to the extent that there are, I’m not going to go into them. But we do know enough to say that the progress is continuing. We don’t see any evidence that things are leveling off. The reality of the world we live in is that it could stop any time. Every time we train a new model, I look at it and I’m always wondering—I’m never sure in relief or concern—[if] at some point we’ll see, oh man, the model doesn’t get any better. I think if [the effects of scaling] did stop, in some ways that would be good for the world. It would restrain everyone at the same time. But it’s not something we get to choose—and of course the models bring many exciting benefits. Mostly, it’s a fact of nature. We don’t get to choose, we just get to find out which world we live in, and then deal with it as best we can.

What did you set out to achieve with Anthropic’s culture?

We have a donation matching program. [Anthropic allows employees to donate up to 25% of their equity to any charity, and will match the donation.] That’s inclined to attract employees for whom public benefit is appealing. Our stock grants look like a better deal if that’s something you value. It doesn’t prevent people from having financial incentives, but it helps to set the tone.

In terms of the safety side of things, there’s a little bit of a delta between the public perception and what we have in mind. I think of us less as an AI safety company, and more think of us as a company that’s focused on public benefit. We’re not a company that believes a certain set of things about the dangers that AI systems are going to have. That’s an empirical question. I more want Anthropic to be a company where everyone is thinking about the public purpose, rather than a one-issue company that’s focused on AI safety or the misalignment of AI systems. Internally I think we’ve succeeded at that, where we have people with a bunch of different perspectives, but what they share is a real commitment to the public purpose.

There’s a real chance that Donald Trump gets elected at the end of this year. What does that mean for AI safety?

Look, whoever the next president is, we’re going to work with them to do the best we can to explain both that the U.S. needs to stay ahead of its adversaries in this technology, but also that we need to provide reasonable safeguards on the technology itself. My message is going to be the same. Obviously different administrations are going to have different views, and I expect different [government] policies depending on the outcome of the election. All I can really do is say what I think is true about the world.

One of our best methods for making AI systems safer is “reinforcement learning,” which steers models in a certain direction—to be helpful and harmless—but doesn’t remove their dangerous capabilities, which can still be accessible through techniques like jailbreaking. Are you worried by how flimsy that method seems?

I don’t think it’s inherently flimsy—I would more say that we’re early in the science of how to steer these systems. Early in that they hallucinate, early in that they can be jailbroken. Early in the sense that, every time we want to make our model more friendly sounding, then it’ll turn out that in some other context, it gives responses that are too long. You’re trying to make it do all the good things, and not all the bad things, but that’s just hard to do. I think it’s not perfect, and I think it’ll probably never be perfect. We have a Swiss cheese model: you can have one layer that has holes in it, but if you have 10 layers of Swiss cheese, it’s hard for anything to fly through all 10 layers. There’s no silver bullet that’s going to help us steer the models. We’re going to have to put a lot of different things together.

Is it responsible to use a strategy like that if the risks are so high?

I’d prefer to live in an ideal world. Unfortunately in the world we actually live in, there’s a lot of economic pressure, there’s not only competition between companies, there’s competition between nations. But if we can really demonstrate that the risks are real, which I hope to be able to do, then there may be moments when we can really get the world to stop and consider for at least a little bit. And do this right in a cooperative way. But I’m not naive—those moments are rare, and halting a very powerful economic train is something that can only be done for a short period of time in extraordinary circumstances.

Meta is open-sourcing models that aren’t too far behind yours, in terms of capabilities. What do you make of that strategy? Is it responsible?

Our concern is not primarily about open-source models, it’s about powerful models. In practice, what I think will happen with most of these companies—I suspect it will happen with Meta, but I’m not certain—is that they’ll stop open-sourcing below the level where I have concerns. I ultimately suspect that open-source as an issue is almost a red herring. People treat open-source models as if they undermine our business model. The truth is, our competitors are anyone who can produce a powerful model. When someone hosts one of these models—and I think we should call them open weights models—they get put on the cloud and the economics are the same as everything else.

Microsoft recently said it is training a large language model in-house that might be able to compete with OpenAI, which they’ve invested in. Amazon, which is one of your backers, is doing the same. So is Google. Are you concerned that the smaller AI labs, which are currently out in front, might just be in a temporary lead against these much better-resourced tech companies?

I think a thing that has been prominent in our culture has been “do more with less.” We always try to maintain a situation where, with less computational resources, we can do the same or better than someone who has many more computational resources. So in the end, our competitive advantage is our creativity as researchers, as engineers. I think increasingly in terms of creative product offerings, rather than pure compute. I think you need a lot of pure compute. We have a lot, and we’re gonna get more. But as long as we can use it more efficiently, as long as we can do more with less, then in the end, the resources are going to find their way to the innovative companies.

Anthropic has raised around $7 billion, which is enough money to pay for the training of a next-generation model, which you’ve said will likely cost in the single-digit billions of dollars. To train the generation after that, you’re starting to have to look at raising more cash. Where do you think that will come from?

I think we’re pretty good for the one after the next one as well, although in the history of Anthropic, we’re going to need to raise more. It’s hard to say where that [will come from]. There are traditional investors. There are large compute providers. And there are miscellaneous other sources that we may not have thought of.

One source of funding that doesn’t get talked about very much is the government. They have this scale of money—it might be politically difficult, but is that something you’ve thought about, or even had conversations about?

The technology is getting powerful enough that government should have an important role in its creation and deployment. This is not just about regulation. This is about national security. At some point, should at least some aspects of this be a national project? I certainly think there are going to be lots of uses for models within government. But I do think that as time goes on, it’s possible that the most powerful models will be thought of as national-security-critical. I think governments can become much more involved, and I’m interested in having democratic governments lead this technology in a responsible and accountable way. And that could mean that governments are heavily involved in the building of this technology. I don’t know what that would look like. Nothing like that exists today. But given how powerful the thing we’re building is, should individuals or even companies really be at the center of it once it gets to a certain strength? I could imagine in a small number of years—two, three, five years—that could be a very serious conversation, depending on what the situation is.

Do you think the current level of inequality in the world is too much?

I would say so. We have billions of people who still live on less than $1 per day. That just doesn’t seem good.

One of the things you hear from someone like Sam Altman is, if we can raise the floor for those people, then he’s comfortable with the existence of trillionaires. How do you think about the power imbalance that comes with that level of inequality?

There’s two separate things there. There’s lifting the floor, and then there’s power within the world. I think I’m a little less sanguine about this concentration of power. Even if we materially provide for everyone, it’s hard for democracies and democratic decision-making to exist and to be fair, if power is concentrated too much. Financial concentration of power is one form of power, and there are others. But there were periods in U.S. history—look at the Gilded Age—where industrialists essentially took over government. I think that’s probably too extreme. I’m not sure the economy can function without some amount of inequality. But I’m concerned we might be in one of those eras, where it’s trending towards being too extreme to be healthy for a democratic polity.

Ideas around guaranteed basic income—if we can’t think of anything better, I certainly think that’s better than nothing. But I would much prefer a world in which everyone can contribute. It would be kind of dystopian if there are these few people that can make trillions of dollars, and then the government hands it all out to the unwashed masses. It’s better than not handing it out, but I think it’s not really the world we want to aim for. I do believe that if the exponential [rate of AI progress] is right, AI systems will be better than most humans, maybe all humans, at doing most of the things humans do. And so we’re really going to need to rethink a lot. In the short run, I always talk about complementarity—we should make sure humans and AIs work together. That’s a great answer in the short run. I think in the long run, we’re really going to need to think about, how do we organize the economy, and how humans think about their lives? One person can’t do that. One company can’t do that. That’s a conversation among humanity. And my only worry is, if the technology goes fast, we’ll have to figure it out fast.