Tech giants Amazon, Meta, Alphabet, Microsoft, and Oracle have been taking on enormous new debt in a race to build out their artificial intelligence ventures in the last year, fueling Wall Street fears of a bubble capable of disrupting the entire economy.
In this edited conversation, Andy Wu, Arjun and Minoo Melwani Family Associate Professor of Business Administration at Harvard Business School, explains why AI hyperscalers — firms that operate, or will need to operate, massive, global data centers — are taking on enormous liabilities and whether investors are right to worry about a possible AI bubble.
Why are generative AI firms fundraising so aggressively?
Generative AI is perhaps the most exciting technology since the rise of the internet. That excitement has attracted a significant amount of attention from private equity, venture capital, and public equity investors.
I agree with the consensus about the long-term value creation potential of generative AI. But achieving that long-term vision requires a capital-intensive infrastructure buildout. We need more data centers, more chips, and more electricity to handle the escalating computing needed to both create frontier AI models (training) and use them (inference).
“While generative AI can do amazing things, it is also perhaps the most wasteful use of a computer ever devised.”
Andy Wu
While generative AI can do amazing things, it is also perhaps the most wasteful use of a computer ever devised. If you do 1+1 on a calculator, that’s one calculation. If you do 1+1 in generative AI, that is potentially a trillion calculations to get an answer. That consumes a huge amount of chip capacity and electricity.
And so, many companies are attempting to build out that capacity by buying chips, building data centers, and, in some cases, even buying and building nuclear power plants to power those facilities.
The issue is that someone has to incur the fixed cost of the buildout today for the potential of long-term profit in the future. That long-term profit is hypothetical and has not been realized yet. The companies leading this buildout have taken on significant debt alongside unprecedented levels of equity financing.
As the market for cloud computing grows, it becomes more competitive and, in some ways, less attractive.
If you asked me five years ago, I would have said the market for public cloud infrastructure was only big enough for three hyperscalers. But now that there’s so much more demand for computing, more companies can reach the economies of scale to be viable.
Just a few years ago, Oracle was not a part of the conversation. But the growing market has dramatically lifted Oracle and allowed it to become economically viable and competitive with Microsoft, Google, and Amazon.
Several companies, and especially the neoclouds that specialize in renting out GPUs [graphics processing units], have borrowed significant amounts premised on hypothetical cash flows in the future. So they’re borrowing money now to build a data center that they expect to get paid for by somebody else in the future.
For instance, OpenAI has promised $100 billion contracts to several of its vendors. OpenAI today does not generate anywhere near the amount of revenue to pay for any of that.
Those vendors have raised money to build data centers on the assumption OpenAI is going to pay them $100 billion later. If OpenAI cannot grow revenue fast enough to meet those commitments, several of those vendors will be underwater financially.
Is that buildout truly needed right now?
The industry faces two contradictory timing problems. On one hand, from a long-term perspective, my view is that the scale of buildout is absolutely necessary to facilitate AI. If anything, we’re probably too slow: not just on the data center side, but especially on the electrical grid.
But on the other hand, the risk right now is the gap between the long-term vision of AI and whether or not the growth will materialize fast enough to pay for the buildout. The subtlety here is that these companies can end up underwater if AI grows fast but less rapidly than they hope for.
Why such apprehension over AI borrowing and spending?
First, it’s become apparent how much money has been borrowed. Financing losses with equity investment is one thing. But defaulting on debt has much more disruptive consequences for the companies involved and our economy as a whole.
Second, there are unusual “circular financing” arrangements between customers and suppliers that have drawn attention. To some, it appears that Nvidia is paying its customers to buy its products. Certainly, there’s some scenario where vendor financing is justifiable, but it certainly raises eyebrows here.
More generally, the bigger issue is downstream. For the potential customers of the data centers — the companies training models or running inference — there’s no short-term scenario in which they are economically viable given how costly it is today. The customers of the data centers are not themselves profitable, and they have no immediate way of generating enough revenue to cover the cost of compute.
“What’s critical to understand, but overlooked by most users, is that generative AI has a significant variable cost.”
Andy Wu
What’s critical to understand, but overlooked by most users, is that generative AI has a significant variable cost. It costs OpenAI real money every time we ask ChatGPT something and ChatGPT responds.
OpenAI CEO Sam Altman once joked saying, “Please” and “Thank you” to ChatGPT costs them millions of dollars. For now, as AI applications grow their customer base and usage, they lose more money. Growth itself does not fix the economics.
Do you see signs of an AI bubble?
In my research on technology strategy, I often look back at the history of the technology industry for hints on how to think about the future.
Technology regularly goes through these ups and downs. The dotcom bubble is the most famous, but in recent years, we’ve had a work-from-home bubble with Peloton and Zoom. We’ve had a bunch of crypto bubbles. There was a virtual reality bubble. In the mid-2010s, we had a gig economy bubble.
It’s easy to get overexuberant about technology.
I would define a bubble in technology as when there’s a significant mismatch between the vision for potential value creation and the current reality of value capture. In other words: Everyone can imagine how useful the technology will be, but no one has figured out yet how to make money. This mismatch puts companies currently operating in a very difficult position.
Regardless of the long-term legitimacy of their offering, they have real financial obligations they have to meet today that they may not be able to meet. What separates a hype cycle that goes away without much fanfare versus a truly destructive bubble is the amount of leverage and risk being taken by the investors and vendors.
Given AI’s importance, what effect could an AI bust have on the U.S. economy?
Big tech is largely insulated from the risks of this. They’ve taken a shrewd and conservative strategy for AI. They positioned themselves well to benefit from the rise of AI, but they don’t stand to lose that much if AI grows slower than anticipated.
Why won’t Big Tech lose much if AI falters?
First, Microsoft has mostly outsourced AI to a third party, OpenAI. Second, Amazon will support anybody’s AI model, seemingly indifferent to the specifics. Third, Meta spent billions of dollars building an open-source AI model that they hand out for free to the world.
If you take those three facts in conjunction, what that’s saying is that these companies don’t really think that core AI technology is a meaningful business in and of itself.
“If you take those three facts in conjunction, what that’s saying is that these companies don’t really think that core AI technology is a meaningful business in and of itself.”
Andy Wu
Instead, they’re focused on profiting from all the adjacencies to AI. I often use a gold rush analogy: OpenAI, Anthropic, and xAI are out there digging for gold.
Nvidia is the consummate shovel seller, designing the chips needed by the gold diggers. And Meta is the consummate jewelry maker: Meta’s social media, advertising, wearables, and metaverse businesses stand to benefit from advancements in generative AI, wherever it comes from and whenever it comes. Microsoft does a bit of shovel selling and jewelry making, but the key thing is they’re not stuck digging for gold.
Certainly, it’s plausible that Amazon and Microsoft and Google might make less money on their cloud computing than they ideally would like if AI growth slows or declines, but they would not end up in financial distress. They still have plenty of customers absent AI.
Who is most exposed if AI fizzles out?
The model builders and the neoclouds, because they’re entirely dependent on a very particular growth trajectory of AI.
What should AI investors keep in mind?
As John Maynard Keynes allegedly once said, the markets can remain irrational longer than you can remain solvent.
But there is an important nuance that Keynes missed. Any ambitious vision for a new technology rests on faith in the unproven, so backing it inherently demands a degree of irrationality. If the market can keep the faith to persist, it buys the necessary time for the technology to mature, for the costs to come down, and for companies to figure out the business model.
In other words, if the market can remain irrational long enough, the vision eventually becomes the reality.