Dave Lee: Anthropic Brings 'Anti-Sell' Strategy to Its IPO

Joining us now is Dave Lee, a Bloomberg Opinion columnist.

Look at all that money.

That's an awful lot of money, and it balances things out, doesn't it? This technology could end the world as we know it, change jobs, and change the economy. But people looking at this S-1 filing wanted to know how the business was doing. The company reported $16.3 billion in revenue in the first six months of this year. It will need to bring in much more money than that to justify all its spending. Other companies are facing the same challenge. Still, this is a major moment. Someone described it as a test of whether investors believe in the AI trade, and I think that's a fair description.

Are we living in a strange world where people are simply saying, 'Okay, they've explained the risks, and we still want in'? There's a lot in this filing. The US midterm elections are coming soon, people are pushing back against data center construction, and supply chain problems and other issues could slow things down. Are we living in a strange world where people still can't wait to invest in this?

It is a strange world, and I guess it's also Dario Amodei's world. He's the CEO of Anthropic, and he has built his public image around talking about the worst possible outcomes of AI. That's what he's been doing. Anthropic itself was founded because its leaders were unhappy with what was happening at OpenAI, so they left and started another company.

But what is really different between the two companies?

Anthropic would say it cares more about safety, although some people question that claim. This focus on safety appears throughout its S-1 filing. In several places, the company tells potential investors that it will slow down if it believes its AI is not safe. That could reduce its revenue and might even put the whole company at risk if its leaders do not believe they can make the technology safe enough.

I see this S-1 filing as a kind of 'anti-sell.' The company refers to this idea in the document itself. When Anthropic is trying to hire someone, it tells the person about all the things that could go wrong and all the reasons the job might not be right for them. Many people decide to work somewhere else. Others say, 'I think I can handle this.' I think Anthropic wants to find investors who react in the same way. It wants investors who trust Dario Amodei, Daniela Amodei, and the rest of the leadership team. These investors must be willing to wait while the company spends huge amounts of money on more computing power, continues to invest heavily in research and development, and explores many ways to use its technology. The company will need time to do all this, and I think its prospectus is meant to find investors who understand that.

In the next hour, we're going to speak with Paul Sweeney. He spent 20 years on Wall Street and lived through the financial crisis, as well as the dot-com boom and bust of the 1990s and early 2000s. He reminded us that it took 10 years for the Nasdaq to return to its level from before the dot-com bubble burst.

When we look for historical comparisons, is SpaceX the only real example? It's a huge company, it's not profitable, and its potential market overlaps with Anthropic's in some ways.

The only example in what sense?

The only other company we can compare with Anthropic in terms of the size and scope of its S-1 filing and IPO.

One interesting thing about SpaceX is that, as we've seen Elon Musk discuss on Twitter over the past few days, it has other businesses besides AI. It has rockets and Starlink, its satellite internet service, which is also growing.

Maybe it will have cars soon, too.

Maybe it will have cars once all these businesses come together. That's the general idea.

And robots.

Right, although that seems more difficult. That's a separate story. But I think the point he was making about other companies, such as the so-called neoclouds, is important. These companies are set up mainly to provide computing power to AI companies. It's hard to run a business like that if you have nothing else to support it.

SpaceX has rockets. Meta has its other businesses, including advertising and all its apps. Google has many different businesses, too. I wonder how difficult things will be for Anthropic because AI is its entire business. It's a company focused almost entirely on one thing. When it releases a product, it needs to charge customers for it right away. In recent days, Anthropic has launched a competitor to Meta's AI assistant, Muse, called Claude Duet or Claude Duo. I've actually forgotten the exact name. Anthropic cannot offer a product at a loss and support it with money from other businesses in the same way Meta can, because AI is its whole business. So while SpaceX may be a useful comparison in some ways, it has more ways to move money between its businesses, something Elon Musk is very good at.

Unlike OpenAI, which uses advertising in the free version of ChatGPT, Anthropic does not do that.

Anthropic has made a strong commitment not to use ads, hasn't it? Its Super Bowl ads were very effective and a little cheeky, and they certainly annoyed Sam Altman. The company repeated its promise in this prospectus, saying it would not use advertising. I'd love to know what Anthropic will think about this five years from now. Will it stick to saying it never wants ads, only to decide later, like Netflix did, that maybe advertising is worth trying?

Still, Dave, I want to understand how all this will work. There are many large language models, and then there's what China is doing. A lot of new technology is coming at us. Will we live in a world where Claude is good at one thing, Gemini is good at another, and Chinese open-source models are good at something else, even for people who just want travel advice? I don't understand how this will eventually play out. Do we need all these different models?

That's a really interesting question. Early investors in these companies, especially OpenAI, often believed that one company would eventually become the clear winner. But things aren't turning out that way. Different models are proving better at different tasks. I recently had breakfast with the CEO of a successful legal AI company. He told me that his company switches between models every couple of weeks, depending on which one performs best at particular tasks.

So their product is built in a way that lets them switch between models easily? They can swap out the brains behind the product?

Yes. For example, one model might be better at writing a legal document, while another might be better at finding references to support its claims. Which model does each task best can change when new versions are released and the technology improves. The CEO said his company can switch models by changing a single line of code. So when I look at Anthropic spending billions of dollars, only to be the preferred model for perhaps two weeks, I wonder how difficult that business will be.

But who owns all the data these companies collect? Where does it come from? To make their models better, they need to keep taking in more information. I think about stories published by Bloomberg or medical research documents, for example.

Or the things people upload to these tools. We're helping make all these models better, too.

Yes, all of that information is part of the issue.

At some point, won't someone say, 'Wait a minute, that data belongs to a particular person or company'?

I think some people would say we're already at that point. Consider the music and articles that AI companies collect from the internet. There are several lawsuits between companies that provide content and AI developers. For example, The New York Times is suing OpenAI. There's also another problem: eventually, there might not be enough new data left to keep training these models. One solution some AI companies are exploring is synthetic data. That means creating new information for the models to learn from instead.