#010 - The looming AI debt wall
How soon till AI's easy money runs out?
You're reading Complex Machinery, a newsletter about risk, AI, and related topics. (You can also subscribe to get this newsletter in your inbox.)
I've been following coverage of commercial real estate's "looming debt wall" that stems from a mix of office vacancies and rising interest rates. (See Michele Wucker's excellent series on the topic if you need a catch-up.) All of this has me thinking about a similar, albeit smaller, issue that I'm calling "AI debt."
I'd also like to thank longtime friend, collaborator, and data science expert Noelle Saldana for reviewing early drafts of this newsletter.
You've probably heard the term technical debt before. Software dev teams take on tech debt when they skip over best practices in order to deliver a product faster. Similarly, a company accrues AI debt when leaders make bold, unfounded AI promises and cross their fingers that things pan out down the road.
These companies could take the safe path of developing an AI strategy, making sure that their ideas are within the realm of possibility, and so forth. But just like saving up to buy a house in cash, or following best practices when building software, doing that kind of AI prep work involves time and opportunity costs. So instead they load up on AI debt and watch the customers and investors roll in.
Debt isn't necessarily a bad thing. It lets you have your cake and eat it, too: you get the thing you want right now, even though you technically can't afford it just yet; and you have time to make good on your promise to repay. The problem is when you take on debt that you're unlikely to pay back. That kicks off a race against time in which luck, not skill, determines whether you can deliver.
Flavors of AI debt
Sometimes you luck out, and your leap of faith into AI actually turns into a working product. You faked it, then you made it. Congrats. (But please remember that this was a close call.) The rest of the time you're left with an AI system that reflects its true origins as a hasty gamble. Like when the product …
- … works, but it's an overpowered solution. You've developed AI models when you could have taken a software-based approach. Or you've created an AI chatbot when plain old search would have worked just fine. Whatever the case, you've shelled out a lot more money than you needed to, which means you'll have to work even harder (by selling a ton more of your product, or bending the truth with investors) to recoup your losses.
- … works, and you're getting traction, but the product is not defensible. Let's say your product is a thin layer on top of a provider's AI models, like OpenAI's ChatGPT. Your customers love it, and you're making money … but you're in real trouble if your provider changes their tune. You're also under threat because your customers might go straight to that provider to build their own solution.
- … doesn't work that well. Or not at all. That AI you bragged about? It's a flop. Now you have to paper over the holes with marketing and hand-waving. And you're spending even more money trying to get the AI to work in the meantime. If enough customers or investors determine that you're playing Fake It Till You Make It, they'll bail.
We've seen this before: the problem with cheap debt
AI is a hot technology that attracts attention from investors and buyers. Phrases like "the AI arms race" further encourage us to outspend the competition, so no one will question the money you throw at AI. (They'll probably be happier that you've taken the plunge.) All of this makes it easy to take on AI debt.
This should be a good thing, right? Econ 101 says that when the cost is low, we should load up!
Yes.
And also, no.
The problem with cheap debt is that it doesn't stay cheap forever. As the price of debt rises, it can make it harder to get new loans and push existing debt holders over the edge. Consider other fallouts from cheap debt:
- The 2008 Great Financial Crisis (GFC): Interest rates were low and lenders were loose on underwriting rules. That allowed homeowners to load up on adjustable-rate – sometimes interest-only – mortgages that were beyond their means. For a while they could refinance their mortgages to defer payments, but that opportunity collapsed when housing valuations stalled. Suddenly, the due date was now and the amount due was more than they could pay.
- Tech companies: Following the GFC, plummeting US interest rates led to infamous "Zero Interest Rate Protocol" (ZIRP) phenomena. Like, say, unprofitable startups loading up on VC money and loans in an attempt to corner their market by keeping prices artificially low. Now that interest rates are rising, that money is no longer cheap. Some of those companies have had to dramatically cut headcount or close up shop.
- The "Looming Debt Wall" of commercial real estate (CRE): The onset of the Covid-19 pandemic triggered a spike in remote work, gutting office landlords' revenues. The other shoe dropped when interest rates rose. Unlike traditional real estate, CRE loans are meant to be ongoing affairs in which holders refinance every now and then. If rates rise at the same time that revenue is low … then you have a big problem.
(I've greatly oversimplified the GFC and CRE stories for brevity. But the general points still stand.)
Change is on the horizon
The cost of AI debt will similarly rise at some point. As buyers develop a stronger understanding of what AI can and cannot do, and as they see the drawbacks as well as the benefits, they won't heap as many unrealistic hopes on it. That'll cool their appetite for the technology. Laws will catch up, imposing new regulatory burdens on what is currently a wide-open space of opportunity. Expect opex and capex to increase, too – either directly, as providers raise prices, or indirectly because they eliminate cheaper tiers of service.
Once this happens, companies will find it tougher to acquire new AI debt (make promises about new products) and tougher to pay off existing obligations (get more time and resources to make existing products work). The fallout will impact AI-related jobs, investors, and services.
What, specifically, will be the precipitating events for this turn of fate? I can't say for sure. AI's periodic name changes have helped it to dodge this reality but it can't play that card forever. History tells us that cheap debt eventually becomes expensive. And that's a painful shift for those who carry too much of it.
Making moves
Will the AI debt crash be as bad as 2008's GFC or today's looming CRE debt wall? Not so much. AI's connections don't run anywhere near as deep or wide as those of the real estate markets, so the damage should be reasonably contained.
Still, damage is damage. Hand-waving will only protect you for so long. The best move is to focus on the long-lasting AI use cases. Those that address real business needs and are competitively defensible will be able to pay off their debts. The rest are worth less than the time, money, and effort put into them – meaning that, in loan terms, they're underwater.
What's in a name?
Unrelated to AI debt…
Well, sort of. Maybe.
Apple, to much fanfare, has thrown its hat into the ring of Shoving AI Into Every Possible Crevice. It's quite a step up since iOS already contains a fair amount of AI – FaceID to unlock, indexing your photo album, all that.
To separate the Old And Busted from the new Hotness, Apple is boldly calling this latest wave "Apple Intelligence." (We've seen this before. Remember when companies rebranded NFTs for their loyalty programs while crypto was having a meltdown?) It's like they wanted to distance themselves from plain old artificial intelligence but still liked the acronym.
Can they really apply the "Apple" moniker if it's just a gateway to another provider? Apple Intelligence is mostly OpenAI under the hood. The best part is that they're paying the ChatGPT parent in my favorite currency, ExposureBucks™. Something that costs Apple nothing to provide and yet, somehow, still holds value for what is already the most famous company in the genAI space. It's like Oreo pushing hard to market its cookies – I don't see how it makes sense.
Then again, how much of the AI field makes sense? The Financial Times says that even Apple is grasping with this latest AI offering. We'll see whether the bold promises and slick presentations eventually lead to working, meaningful solutions.
Hmm.
So maybe Apple has taken on some AI debt, after all.
In other news …
- The SEC does not approve of AI-washing. Besides the charges of financial fraud, bankrupt startup Joonko is also on the hook for allegedly claiming that their AI could do things that it could not. (WSJ)
- The FTC is similarly unimpressed with AI-washing. (SEC blog)
- "Using AI to teach math" might raise an eyebrow. But "using the inner workings of AI as the backdrop for teaching math" sounds pretty cool. (Le Monde 🇫🇷)
- It was only a matter of time. Thanks to the rush of AI-generated music, we now have companies that claim to detect AI-generated music. Hopefully they do better than the first wave of AI plagiarism detection tools. (Les Echos 🇫🇷)
- Maybe "self-driving" is not the best description of a vehicle that rams a police car. Maybe. (Gizmodo)
- AI models – especially genAI models – involve randomness. But getting AI to change up a movie every time you watch it? That takes randomness to a new level.
The wrap-up
This was an issue of Complex Machinery.
Reading this online? You can subscribe to get this newsletter in your inbox every time it is published.
Who’s behind Complex Machinery? I'm Q McCallum. I think a lot about AI and risk, which I write about here.
Disclaimer: This newsletter does not constitute professional advice.