Does the world really want what Sam-AIam is selling?
Fans of Dr Seuss will know by heart the key stanzas of Green Eggs and Ham: Do you like Green eggs and ham? I do not like them, Sam-I-am.
I do not like Green eggs and ham.
For those who have never had to read a bedtime story, allow me to explain. An irrepressible little creature, Sam-I-am, spends the entirety of the book pitching green eggs and ham — on the face of it, an unappetising dish — to a sceptical and increasingly irascible larger creature. With every page, the pitch grows more elaborate. On a boat? With a goat? In the rain? On a train? Surely, there must be some context in which green eggs would be appealing fare. By the time Sam prevails, his hapless victim inhabits a scene of chaos.
When you come to think of it, there is often someone called Sam trying to sell you something you don’t initially want. In the 1920s, as I learnt from Andrew Ross Sorkin’s page-turner 1929, it was Sam Crowther whose article “Everybody Ought to Be Rich” — by buying stocks with margin credit (loans from brokers to investors) — captured the hubris of the Roaring Twenties. A few years ago it was Sam Bankman-Fried with his crypto exchange FTX. “I have an abundance mentality,” Bankman-Fried declared at the height of his fame. “I want FTX to be a place where you can do anything you want with your next dollar. You can buy bitcoin … You can buy a banana.”
And you could also have bought green eggs and ham — until FTX blew up and Sam landed in jail.
A lot of the applications of generative artificial intelligence (AI) remind me of green eggs and ham.
Take OpenAI’s Sora 2.0. With a few prompts, you can generate soft-porn videos of scantily clad Manga girlelves.
This is also one of the ways Elon Musk tries to sell XAI’s Grok. But why would I want to watch such videos, any more than I want to eat green eggs and ham? Financial history can help us here. If you’re unsure if there’s an AI bubble, just refer to the historian Charles Kindleberger’s five-stage model:
1 Displacement: Some change in economic circumstances creates new and profitable opportunities for certain companies.
2 Euphoria or overtrading: A feedback process sets in whereby rising expected profits lead to rapid growth in share prices.
3 Mania or bubble: The prospect of easy capital gains attracts first-time investors and swindlers eager to defraud them.
4 Distress: The insiders discern that expected profits cannot possibly justify the now exorbitant price of the shares and begin to take profits by selling.
5 Revulsion or discredit: As share prices fall, the outsiders stampede for the exits, causing the bubble to burst altogether.
We are currently at stage 3.
It is impossible to read the first part of Sorkin’s 1929, with its shameless self-promoters, big-spending bankers and irrationally exuberant retail investors — above all, with its rampant leverage — without being reminded of our own times. We too easily forget that underpinning the stock market boom of the Twenties were the tech stocks of the day: RCA, for example, the company that encapsulated the possibilities of mass entertainment on radio, vinyl and celluloid.
Today, we’re being offered something even more alluring than the cornucopia of the Jazz Age. According to the median forecast of Ezra Karger’s Longitudinal Expert AI Panel (LEAP), by 2030 more than 18 per cent of American work hours will be AI-assisted. By 2040, AI will be as important to this century as electricity or the car were to the previous one. Indeed, there is a one in three chance that AI is going to rank alongside the printing press as a technology that “changed the course of human history”.
Even if AI falls short of that, Reuters reported last week that 97 per cent of listeners cannot tell AIgenerated and human-composed songs apart. The song topping the country charts, Breaking Rust’s Walk My Walk, is yet more AI slop.
AI, or rather the promise of AI, is now the principal driver of both the US economy and the stock market. Between a sixth and two fifths of the rise in GDP over the past year is attributable to investments in computer and communications equipment, including chips, data centres, grid upgrades and AI software.
The US fund manager Ruchir Sharma estimates that AI companies account for 80 per cent of the gains in US stocks this year. The blogger-economist Noah Smith notes that “more than a fifth of the entire S&P 500 market cap is now just three companies — Nvidia, Microsoft and Apple — two of which are basically big bets on AI”. The so-called Magnificent Seven (those three plus Alphabet, Amazon, Meta and Tesla) account for more than a third of the S&P. Quarterly capital expenditure (capex) by these companies now exceeds $110 billion, roughly three times what it was two years ago. A huge proportion of the total — nearly two fifths — consists of purchases by everyone else of Nvidia’s graphics processing units (GPUs).
The standard analogy is with the dotcom bubble that peaked in 2000. The standard counterargument is that the value of Nvidia is much lower relative to the company’s earnings than was true of the tech giant Cisco 25 years ago. Unlike most other stock markets (except Japan’s), the growth in US market cap reflects rising earnings, not just rising valuations.
Moreover, in the late 1990s capex, much of it in fibre optic cables, ran far ahead of demand. The same is not true of demand for GPUs. Nvidia, which this week announced record revenue growth, cannot keep up with AI-driven demand for additional “compute”. Nor can the American power grid. One of the unintended consequences of the AI capital expenditures boom is that US electricity bills are up 7 per cent this year.
Add to that the sheer speed of adoption of AI. More than 18 billon messages are sent to OpenAI’s flagship large language model, ChatGPT, every week. The rate of adoption is far higher than that of the world wide web in the 1990s.
In short, AI is changing the 2020s economy faster than the internet changed the 1990s economy. In a recent paper, Stanford University’s Erik Brynjolfsson and co-authors show that, since the widespread adoption of AI, workers aged 22-25 in the most AI-exposed occupations, such as legal services, “have experienced a 13 per cent relative decline in employment, even after controlling for firm-level shocks”. Talk to anyone in investment banking and they will tell you that their programmes to hire entry-level analysts are being slashed.
For all these reasons, 19th-century US railroads may be a better analogy than 1990s telecoms. Think of today’s capital expenditures on data centres being like that on railways 150 years ago. And there’s the rub. Two things can be true at the same time: a) AI could be as economically worthwhile an investment as railroads and b) we could still experience at least one stock market crash along the way to its general adoption.
The problem arises, as Noah Smith has argued, if — as happened in 1873 and 1893 — investors suddenly realise that their returns won’t be quite as rapid as they had expected. If AI investors realise the same, or that returns won’t accrue to the companies doing all the big-ticket investment, a crash is likely. Moreover, history tells us that the economic hit will be proportionately larger depending on how much of the capital expenditures are being financed by debt, as opposed to equity of cash flow from other sources.
Established big tech remains safer
Clearly, the “hyperscalers” formerly known as Big Tech — Microsoft, Amazon, Meta and Alphabet — can finance the lion’s share of their capital expenditures from free cash flow because they already run hugely profitable software, cloud-computing, search and platform businesses. And equally clearly, they’re likely to continue to invest in Nvidia as long as Jensen Huang’s chip design and software remain state of the art. But OpenAI is another matter.
According to the Wall Street Journal, Sam Altman “recently told employees that OpenAI wanted to build 250 gigawatts of new computing capacity by 2033 … a plan that would cost over $10 trillion by today’s standards”. That would also be equivalent to a third of current US peak energy usage.
Altman’s pitch is to promise global economic transformation if he can get enough computing capacity. But he adds a non-zero risk that artificial general intelligence wipes out humanity, which somehow makes it even more impressive.
Yet OpenAI is not quite ten years old; it used to be a non-profit; its corporate governance has sometimes resembled a soap opera; its flagship product ChatGPT is only three years old; and its burn rate (the amount of money it loses each quarter) may be the highest in history. How does Altman propose to pay for 250 gigawatts of new computing capacity? The answer is only partly by taking out bank loans ($4 billion to date). Nor is it by issuing bonds. But it still involves debt of a kind — from just about everyone else in the AI game.
Altman has signed a $22.4 billion cloud contract with CoreWeave. He has signed a $38 billion deal with Amazon Web Services. He has agreed to buy Broadcom’s custom chips and networking equipment. The only hitch is that “OpenAI is in no position to make any of these commitments”, as one analyst told the Financial Times last month. Why? Because while Altman says the company’s annualised revenue is “well more” than $13 billion, its losses in the last quarter amounted to $12 billion. Edward Zitron is one of a number of analysts sceptical about OpenAI’s path to profitability.
The company’s claim that revenue will grow to $100 billion by 2028 seems implausible. It would certainly be unprecedented.
Some of OpenAI’s financing is being provided by Microsoft, with which it has a revenue-sharing agreement. There are also deals with Google and Nvidia. Perhaps the most important part comes from Larry Ellison’s Oracle, one of the parties to the Stargate project, a joint venture announced in January to invest $500 billion in AI infrastructure for OpenAI.
These deals are complex. The Nvidia-OpenAI deal, for example, entails an agreement from Nvidia to lease up to five million of its chips to OpenAI. In return, Nvidia will invest up to $100 billion in OpenAI over time to help the company pay for the chips. In that way, Nvidia serves as both an investor in and supplier to OpenAI.
Similarly, OpenAI may be on the hook to pay more than $20 billion to CoreWeave, but it also owns part of CoreWeave, “having made a $350 million equity investment in the company before its initial public offering”.
Other deals involve similar circular financing. The Wall Street term for this is “roundabouting”. The phrase “house of cards” also comes to mind, as clearly anything that caused a significant equity market correction would pose serious problems for the stability of this structure.
“ There has been the realisation that ChatGPT is more of an upgrade on Google Search than a productivityraising miracle
What might lead investors to revise downwards their expectations about the money to be made out of generative AI? I can think of four good reasons for disappointment:
From sell-off to shock to recession
First, the realisation that ChatGPT is more of an upgrade on Google Search than a productivity-raising miracle. The bulk of ChatGPT use is by people seeking practical guidance, information or technical help. By contrast, according to an MIT study, 95 per cent of organisations are getting zero return on their AI investments. That’s because employees are using it to generate what the Harvard Business Review has dubbed “workslop”.
Second, OpenAI has serious competition. A year ago, its share of the generative AI market was 86.6 per cent. Today it’s 72.3 per cent. Google’s Gemini is gaining rapidly. And Anthropic is beating OpenAI when it comes to enterprise AI.
Third, all the US AI players face competition from China’s open-source models, which have rapidly outstripped their US counterparts when it comes to worldwide adoption. More and more US companies, such as Airbnb, are quietly using Chinese models because they are cheap. When Jensen Huang says “China is going to win the AI race” that seems like bad news for Sam-I-am.
Finally, if GPUs are serving implicitly as collateral for AI debt, that’s a potential headache, too. Just ask Michael Burry, who made his name with the “big short” of US real estate before the global financial crisis. Unlike railroads, GPUs are short-lived assets with a useful life of perhaps five years. Or is it eight? Or two? No one knows.
As I said, mild disappointment can cause a crash even when the technology is awesome and the investment will ultimately be worth it for society as a whole. That was scant consolation to those who lost their shirts on railroad securities in 1873 and 1893. How far an equity sell-off today would cause a wider shock, even a recession, depends on the extent of the financial contagion. Step forward Oracle, which has about $96 billion of long-term debt, up from $75 billion a year ago, with a potential total of $290 billion by 2028, according to Morgan Stanley. Oracle’s debtto-equity ratio has surged to 500 per cent, compared with Amazon’s 50 per cent and Microsoft’s 30 per cent.
The price of Oracle credit default swaps (a derivative you buy to protect you against a company failing to honour its debts) has doubled since September.
It’s all fine, just fine, in current financial conditions, which are as easy as they have been in three years. But you have to ask yourself: are the 1920s calling to ask for their financial history back? Certainly, when I read in The Wall Street Journal that “the fates of the world’s biggest semiconductor and cloud companies — and vast swathes of the US economy — [are tied] to OpenAI, essentially making it too big to fail”, I have only one response: I do not like them, Sam-I-am.
I do not like Green eggs and ham.
Comments
Post a Comment