That’s the provocative title of an opinion piece in today’s Times, and the writer makes a pretty convincing case. Surprised I haven’t seen this more often.
It’s pretty simple, really. Nobody knows how high is up (like the telecom guys in 1999 lagging endless miles of fiber, sort of) when it comes to AI.
The other guys in the game: Facebook, Google, Amazon, Microsoft have profitable businesses which can “subsidize” the buildout for, well, ever, practically. Open AI is dependent on funding.
Which they have, in the last funding round of any business ever, in history: 40 Billion. (2nd place goes to Ant Group, a Chinese fintech company, at $18 billion.)
But there’s no profit yet, nor realistic expectation of profit in the near term. And the usual Silicon Valley mantra of “build it and they will come”, or “scaling law” (it will become more powerful, but also exponentially more expensive) may not happen timely. Or at all, in an investment lifetime.
So maybe an IPO could be the bridge? The largest IPO in history is $30 Billion - for Saudi Aramco which had a profitable model, and proven assets in the ground.
And Open AI has said it will spend $1.4 Trillion building data centers and other things as will be necessary to the business.
Isn’t ChatGPT the consumer-facing product of Open AI?
I have spent many hours with ChatGPT over the past week. It has written several estate documents for me to enable DH to use the WA State estate tax exclusion should I pre-decease him. This has saved many hundreds of dollars in lawyer fees. I do plan to bring the documents to my attorney to glance over to make sure they are solid but they look OK to me – and to Perplexity.
Payment to OpenAI for this valuable service = $0.
I also consulted ChatGPT for a detailed description of a new nonstick large skillet to replace mine which is wearing out. Again, lots of detailed recommendations and reviews. Finally, I clicked through on a link to a J.C. Penney skillet which I ordered. Cost to me for ChatGPT’s advice: $0. The pan was on sale with free shipping. How much of a kickback will J.C. Penney pay Open AI?
Will that pay off $1 trillion of data centers?
Wendy
I would guess zero. Do they even know you bothered to look at an Open AI review? Or did it facilitate the purchase? If they do, then they might get a payment, but it will take a lot of those to pay for a trillion dollars worth of data centers but the ongoing training costs, not to mention signing up (and babysitting) retailers.
It’s entirely possible. But I think that the article misses an important point, and one that I like to emphasize when people get a little sanguine about what might happen if the AI bubble bursts.
What did the article miss? Let me give one more quote:
Signaling the scale of capital that he believes he needs, OpenAI has committed to spending $1.4 trillion on data centers and related infrastructure. Even if OpenAI reneges on many of those promises and pays for others with its overvalued shares, the company must still find daunting sums of capital. However rich the eventual A.I. prize, the capital markets seem unlikely to deliver.
The important point to remember is that the buildout of the data centers is not being financed by equity, for the most part. It’s being financed by debt. Meta and Facebook and Microsoft generally aren’t paying for their infrastructure buildouts primarily from earnings, but by borrowing a ton of money.
So Open AI doesn’t need to raise a trillion dollars in its own actual money, either earnings or equity. It needs to raise a trillion dollars in debt. Just like all the other participants. None of them can fund this through current operations - it’s just too much.
I don’t disagree that several of these AI efforts are likely to go bust - mostly because the industry sure looks like it’s heading towards commoditization very fast. Interestingly enough, despite being independent ventures, almost all of the major AI companies have converged on models with very similar architectures and capabilities. There’s at least a seven or eight out there that all do pretty much the same thing, to the same level of competence, in the same way. There is no economic reason why we would spend a trillion dollars each to build more than half a dozen versions of the same basic architecture…..so we probably won’t. Which means that several of them will fall by the wayside.
I’m just not sure that OpenAI will necessarily be one of the ones that does simply because it lacks a revenue stream, since it probably has more than enough equity capital and backing from MSFT to continue accessing the debt markets for as long as the other companies. The field will be winnowed as some companies have their debt spigots turned off. I just don’t know that OpenAI is in any more danger of that, or in danger of it happening earlier, than any of the companies that have big revenues….
While true, I think it misses one big point. The debt that Open AI issues will be backed only by the physical assets it develops, and possible by (hopeful) future revenue streams.
The debt issues by Meta et al are backed by the same physical type assets (data centers) but also by currently existing cash flow of gigantic consumer corporations, as well as the same future revenues that Open AI is promising.
I don’t disagree with your analysis that they’re “mostly the same” even if different, and that some kind of consolidation is inevitable (I anticipate a spin out and shared ownership between a couple of the majors in a NewCo sort of venture), but that still begs the question: why do business with Open AI which has nothing more than Sam’s good word that it will be worth something someday?
And therein lies the problem. If it truly had value it would have a cost. Put differently, how much would you have paid OpenAI for that service, and at what point would you have elected to use the attorney (whom you are still going to pay to look over the AI plan)?
Except they’re not, exactly. The debt obligations aren’t being issued by Meta. They’re being issued by the SPE’s that run the data centers, backstopped by the expectation that Meta will make lease payments to the data centers going forward.
Fundamentally, the same assumptions are in play - AI is going to be hugely profitable for this particular company, and in a short enough time that incurring losses pending profitability will be manageable. If that assumption breaks down, and AI stays unprofitable for a long enough time and/or this specific company decides to give up their AI ambitions…well, it’s going to have the same effect on the debt spigot for both companies. If Meta isn’t making money on AI, they’re going to have to stop incurring debt for AI - or more likely, they’ll stop being able to access the debt markets in a way that makes sense.
Again, I think the general idea that some AI efforts will go belly up is sound. I just think that the article overlooks the fact that the AI infrastructure efforts of the major companies are being done in separate shell entities, for the most part, that are pretty much just as removed from their main companies as OpenAI is from its major investor Microsoft.
I agree, although I do know quite a few people in the IT world (as well as a few in other fields) that have paid subscriptions since there are more extensive features and searches there. According the OpenAI, they brought in an estimated $20 billion in 2025, which was up $10 billion from their estimates at the beginning of the year.
What happened to Meta when they spent $70+B of cash, mostly using their existing cash flow, on the “metaverse” concept? And so far nearly zero revenue and zero profit (well, an accumulated deficit of $70+B).
OpenAI will exist for as long as its investors are willing to pump money into it. If that willingness continues through the generation of free cash flow then they might continue to exist. But if that willingness ends before enough cash flow is generated, then the existence may end abruptly. But many new businesses were like this, for example Uber was entirely supported by investors and new investment, until recently, and they still have a substantial accumulated deficit that they can work down until they produce a true net profit to those investors. And SiriusXM went over 15 years before they overcame their accumulated deficit. Plenty of other examples.
Well, if you clicked on their affiliate link, they will get a small payment. But if you found the item on your own after seeing their recommendation, then they likely get nothing.
And it will probably not pay off the data centers. Of course, a couple of decades ago, I was convinced that advertising couldn’t possibly pay for the capital Google required, so I didn’t invest in Google (now Alphabet). That was a mistake of course.
This is a valid point - people, especially fans, can continue to dump money into a losing private venture for far longer that may be advisable. You listed some examples that eventually paid off. In contrast, I am reminded of Star Citizen. I was a backer about 8 years ago.
In development for over a decade, is still raising money, and has raised over $900 million in crowdfunding by 2025.
By contrast, World of Warcraft cost less than $100 million to develop.
Yes, but I assumed the lease payments were in some way backstopped. Who else would lend the cash to a “private entity” without some assurance of payment.
Is it that there are others lending to the SPE without any assurance of payback? And if so, what absurd interest rate must they be getting for their capital? (I almost wrote stupidity, but thought better of it.)
I don’t think that’s the case. If the hyperscalers were guarantors on the loans, then these things would be showing up on their balance sheets as debt obligations. The whole point of these arrangements is that they not appear as debt to the main company.
All of this is just my “lawyer who is curious and reads things on the internet” level of financial knowledge, but here goes. The SPE investors (such as Blue Owl) put in a non-trivial amount of equity, which is part of it. The data centers themselves are collateral to the loans, typically - these are technically secured debt obligations like a mortgage. The hyperscaler’s role is to provide a set-term lease commitment to the data center, which then supports the appraised value of the collateral. So you have a secured debt instrument supported by collateral with a good loan-to-value ratio, which gets you a good rating - but that “value” part of the LTV is only as good as the hyperscaler’s lease commitment. The debtholder doesn’t have the ability to go after the assets of the hyperscaler itself in the event the debt isn’t serviced - they have the ability to foreclose on the data center and receive lease payments.
That level of remove is what might keep OpenAI on similar footing with the other big hyperscalers in this type of transaction. They just need enough capital to keep current with lease obligations on their existing centers so that they can keep getting debt for the new ones, which they should be able to do with $40 billion plus whatever more they can get from MSFT and the open markets for as long as people are willing to issue debt into the AI Thing.