The familiar 'this time is different'

If Anthropic gets the target valuation it wants, Anthropic and SpaceX together will be worth almost as much as the entire South Korean stock market

In 2013, Demis Hassabis of DeepMind (DM) set up an interesting test for the sundry billionaires wooing him. London-based DM was already hailed as a boutique frontrunner in artificial intelligence (AI), with an array of brilliant researchers. Elon Musk wanted to buy the company. Google wanted to buy it. Facebook wanted to buy it. They all had mega offers.

 

Dr Hassabis was uninterested in the money, or rather, he considered it to be a low priority. He wanted to ensure he could raise enough cash to permanently fund DM. Beyond that, the cash was irrelevant as far as Dr Hassabis was concerned.

 

His real priority was to be allowed to run DM without hindrance. He wanted to keep its focus on solving critical problems, regardless of the return on capital employed. His ultimate goal was (and remains) “solving intelligence” in order to create artificial general intelligence (AGI).

 

So, in one-on-one conversations with Larry Page and Mark Zuckerberg, Dr Hassabis, a five-time world champion gamer, did something that might be considered a little evil. Of course, he discussed AI and his goals of AGI. But he also discussed other “hot technologies” like 3D printing, virtual reality, augmented reality, and self-driving cars. He was looking for their reactions.

 

Mr Page and Mr Zuckerberg were extremely enthusiastic about AI. But Mr Zuckerberg was equally enthusiastic about the other stuff. And that is why Dr Hassabis decided to go with Google’s offer. He figured Mr Zuckerberg didn’t understand that AI would be bigger than any other technology, and that AI would subsume all the other technologies discussed at those meetings. Circa 2026, it is hard indeed to imagine virtual reality or augmented reality systems, or self-driving cars, or 3D printing systems, or any other technology for that matter, which doesn’t have AI in it somewhere.

 

Subsumption is the argument that comes up when people voice their misgivings about the AI bubble. By any standards prevailing in any previous historical cycle, AI is already experiencing the mother of all bubbles.

 

Anthropic is looking at a  initial public offering (IPO) at a valuation of $2 trillion. SpaceX successfully engineered an IPO at a $1.8 trillion valuation, and that wasn’t because it builds rockets or runs satellite broadband services. It is because SpaceX is the holding company for SpaceXAI, which builds “frontier AI models” for Elon Musk. OpenAI is avoiding going public because it can easily raise $30 billion-plus in private capital on the basis of a few phone calls, and Sam Altman is said to believe OpenAI could pick up a valuation of several trillion if the excitement brews for a while longer.

 

Those numbers exceed the gross domestic product (GDP) of South Korea ($1.9 trillion) and Taiwan ($980 billion), even though both nations are critical cogs in AI infrastructure. AI-related infrastructure under planning today will cost over $10 trillion over the next five years. That would be roughly 7-8 per cent of 2025 global GDP. So capex on AI would annualise to around 2 per cent of global GDP.

 

No earlier technology has ever sucked up that much in the way of resources. For context, the internet market-leader Amazon went public in 1997 at a valuation of $438 million. The 1997 GDP of South Korea was $590 billion. At the peak of the internet bubble, Amazon was valued at $26 billion in 1999. It subsequently dropped to $5.5 billion by 2000.

 

Using South Korea as a benchmark may seem unscientific. But so are bubbles. In 1989, the total value of real estate in Tokyo exceeded the total value of real estate in the state of California, although California is larger than all of Japan and just as prosperous.

 

In an apple-to-apples comparison, Korea’s market capitalisation is $4.5 trillion. This includes AI-critical chaebols such as Samsung, Hynix, and LG Energy. Assuming Anthropic gets the target valuation it wants, along with SpaceX, it will be valued at roughly $4 trillion. Are those two companies really worth as much as the entire South Korean stock market?

 

Set against such comparisons, the subsumption argument is the only one that makes some sense. AI’s valuations are justified if and only if AI ends up running all key technology, orchestrating all innovation, and ensuring everything runs at peak efficiency. It will then also be in a position to hold the world to ransom, and it might just do that if Anthropic is correct in warning us about AI’s tendency to deploy blackmail.

 

Every time there is a bubble, the phrase “This time it’s different” does the rounds. If it is indeed different this time, it will be due to subsumption. But what if it isn’t and the bubble collapses? The fallout will be catastrophic.

 

The writer is a New Delhi-based independent journalist