In July I published the AI Risk Weather Report (part 1, part 2), a wide-lens look at the ups and downs of the genAI space. That lays the groundwork for playing the "What if?" game in risk analysis: you toss out several potential scenarios, figure out how they might come to be, then determine how you'd handle them.

A chat with longtime friend and co-conspirator Scott Robbin – an experienced software developer who deserves credit for the strong contributions he's made to today's newsletter – raised one what-if I hadn't considered:

What if the genAI model providers lose software development as a use case?

What, indeed.

Any other field would simply fall back to something else. No big deal. But right now the entire genAI-industrial complex holds all of its eggs in one basket. Not by choice, but because "enhancing software development" is the only basket it can find. Beyond that, genAI's broad-brush applications are "crime" and "creating excitement for genAI" – hardly enough to support the field should software dev take a hike.

In fact, when you compare all of the investment to the outcomes, the genAI-industrial complex is underwater even with software development as an anchor use case. A developer exodus might sink it for good.

Friends for now

It helps to see why genAI-backed software development – that mix of vibe coding, agentic coding, and code assistance – has become so popular in the first place.

For one, bots can handle large contexts better than most humans. That gives them a leg up on understanding new, unfamiliar codebases, and they can pass explanations along to human interlocutors. Two, the ability to generate code at scale enables a shorter time-to-market for apps and features. Three, that speed offers improved experimentation because you can quickly build working prototypes. (That software doesn't need to be perfect. You're not releasing a finished app in that case; your goal is to get something that kinda-works into the hands of a prospect.)

Despite these strong reasons to like genAI The Technology, developers' relationship with genAI The Industry is on thin ice.

Let's start with in-house software teams. Gains from those agentic super powers are highly concentrated among the senior-level practitioners who can effectively delegate to a team of bots. That's a relatively small set. But everyone on the team is using genAI. Including the junior devs who are probably seeing a lot of trial-and-error. The combined LLM token burn rate exposes their employers to unstable pricing and other such subscription nonsense. CFOs will eventually step in to trim the fat. Every cut they make results in less money flowing to model providers' pockets.

Then you have software development studios, which build bespoke apps for other companies. These developers are subject to the same token price concerns as their in-house equivalents. They also work with clients who expect them to pass along the savings from all that agentic magic. This combination of rising costs and thinner margins weakens studios' relationship with model providers.

The last group is the DIY crowd – the companies that figure they can vibe code their way to a solution in lieu of hiring professionals to build it for them. Some of these apps will turn out just fine. But a number are already being turned over to dev studios for cleanup, because the original vibed-up versions turned out to have holes. And because the creators learned that there is more to running an app than generating the code. And because foundation model pricing tacitly targets professional developers – the revenue-generating kind who see their LLM subscription as a cost of doing business – which makes it less appetizing for smaller DIY players.

The icing on the cake is that both dev studios and in-house software teams are seeing diminished performance gains from newer foundation models. They're content to stick with the older, cheaper versions that are still Good Enough™. For them, the money and effort spent on building those newer models has gone to waste.

Swinging the axe

Professional developers are too hooked on LLMs to stop using them altogether, mind you. Should token economics crash into hardware economics, they'll find alternatives. And therein lies a big problem for the foundation model companies:

The ace in every professional dev shop's back pocket is the option – and, importantly, the skill – to self-host an open-weight LLM.

Such a move would allow dev teams to keep their code-generation productivity gains while cutting the foundation model providers out of the revenue loop altogether.

This is when you'll tell me that hardware costs are out of control precisely because of demand from model companies. Yes. But your average CFO would still prefer the fixed costs of purchasing their own internal server cluster to the variable, unstable costs of per-token billing. Particularly sharp-minded shops would find a way to rent out their unused capacity to niche players and hobbyists. As they say: the cloud is just someone else's computer.

You might also tell me that the foundation model companies have other fans. Again, yes. The genAI hype machine has stoked Corporate FOMO™ in every executive suite in the country, which is why we have companywide usage mandates and tokenmaxxing leaderboards. But eventually that will all give way to a simple matter of bubble mechanics.

Your typical asset bubble or mania breaks down because someone finally snaps out of their reverie. One true believer sees the world for what it is, the word spreads, and eventually everyone's beating a path to the exit. If enough developers break away from model providers' services but keep pulling those genAI-backed gains, that will lead people in other departments to ask So what are foundation model companies good for, anyway? Aren't they just a thin paywall over some GPUs? Shortly followed by Can we get IT to set up some LLM machines for us, too?

Importantly, this framing doesn't challenge the value of genAI as a technology. It challenges the value of companies that keep asking for money when they don't seem to provide much. Even AI-hungry execs can get behind ditching the big-model subscriptions once they understand the cost savings and risk mitigation provided by an in-house LLM setup.

Digging a hole

The picture is particularly grim once you consider how foundation model companies have backed themselves into a financial corner. They're hemorrhaging cash while providing little to the customers most likely to slow the bleeding.

It's something of an open secret that they're under-charging for LLM tokens. Subsidizing products this way is reminiscent of ZIRP-era startups, such as food delivery and rideshare, which kept prices artificially low early on in order to gain market share and raise prices once they were the only game in town. When it comes to genAI model providers, though, a sharp price hike would likely result in a short-term revenue bump followed by a sharp drop-off. In the time between, developers would either ferret out efficiencies in their usage or jump ship to the aforementioned self-hosting setup.

Then you have the expensive and risky prospects of foundation model development. Every mathematical model, from a linear regression to a homegrown neural network, represents an amount of up-front work that may not pay off in the form of a workable system. Building an LLM dramatically scales both the effort and the chances of failure, with the added scrutiny of widespread media coverage and a vocal user base. If that new model's a flop, or even just a mild improvement, it's not just a PR problem. It's a big hole in the balance sheet. Right next to all of the other holes in the balance sheet, because foundation model providers are building an expensive product while they're likely hurting for cash. I say "likely" because their finances are officially private; but from the outside, we can all see that they're loading up on debt to fund their habit.

Most of that money is earmarked for datacenter buildouts. Those are expensive facilities, planted in anticipation of future demand for a technology that has insufficient present-day demand and still needs to prove itself. Those datacenters also face public backlash, which makes them less likely to produce return on the investment.

To top it off, both OpenAI and Anthropic are preparing to go public. At least one journalist figures they're trying to IPO not so much to raise cash for future growth, but to backfill the holes they've dug to date.

Looking for an out

Just maybe, a Deus Ex Machina moment could pull the foundation model companies out of this tailspin. It happened with Java (hello, enterprise web development) and JavaScript (sneaking back in while Microsoft Silverlight and Adobe Flash duked it out over market share). Some deity could grace genAI foundation model companies with hardware improvements to offset some of the risks of model development. That, and new use cases could cover revenue lost to self-hosting software developers. Perhaps someone will cook up new user interfaces, too, as we've gotten as much as we can out of that plain old text-entry box. (And it was never that good to begin with, really.)

Barring that, expect these players to take some last-ditch steps to stay in the game. They'll likely create new subscription tiers in order to capture more of that price-insensitive market of corporate behemoths and well-heeled true believers. They may also try partnerships or bundle deals, similar to what you see with streaming services or airline-hotel pairings. One particularly desperate move would be to further subsidize token prices, to force competitors out of the game. The last provider standing can own all of the (very small!) pie and live on. But that only works if a single company is willing to white-knuckle the discounts. If they all chase each other to match prices, it's game over.

And then..?

So what if developers pull away, and the foundation model companies take a tumble?

In a perfect world we'd say that only the companies themselves would need to worry. But this is not a perfect world.

Remember that the tendrils of the genAI spread deep and wide. Between their massive debt holdings and inflated stock market valuations, a collapse would send shock waves throughout the financial system. (And if that happens after the OpenAI and Anthropic IPOs, that would likely erase portions of Main Street investor portfolios.) Those datacenters won't exactly disappear, either. I could go on, but I already wrote up one possible take on a post-crash world last October. You can read that if you're up for it.

It's far more interesting to talk about the potential upsides. For starters, much of the genAI hype would fade. That would pave the path to clearer thinking on use cases. We'd also see greater interest in smaller, task-specific models (those general-purpose behemoths are just showpieces, anyway) and models in smaller places (like on iDevices). Homegrown ML models, many of which are still in operation, would regain their well-deserved share of the spotlight. They've been quietly powering fraud detection systems and spam classifiers while genAI did its song-and-dance routine.

And then there's the last, subtle point that holds a large impact. If we split the foundation model providers into "only provides access to LLMs" and "has other lines of business," well … Leagues of unhappy software developers may drive the former to ruin, thereby eliminating some noisy competitors for the latter.

Time will tell.

In other news …

For more links to recent news, and with a slightly broader scope, I encourage you to check out my other newsletter. It's a weekly, curated drop of what I've been reading.