The Farm League, the Judas Goat, and the Bill That Always Comes Due

I've been watching the tech industry for nearly thirty years now, and I'll tell you something that still genuinely bothers me: the same play keeps getting run, the players keep falling for it, and the people running it seem to sleep just fine, just fine.

The latest example landed in my lap a couple of weeks ago. I logged into Claude[1], poked around as I usually do, and spotted something new in beta: Claude Design. Took about thirty seconds to figure out what I was looking at. That, right there, is an intercontinental ballistic missile with Figma's name stencilled on the nose cone.

And the thing is, Figma saw it coming. They just couldn't do a goddamn thing about it.

Let me back up and explain the pattern, because it has a few moving parts.

The frontier AI labs — Anthropic, OpenAI, and the rest — are burning extraordinary amounts of money. We're talking hundreds of billions of dollars across the industry in infrastructure, training runs, and compute that would make a defence contractor blush. None of them are profitable. Not even close. So to generate the appearance of a functioning business while they figure out whether the technology can actually pay for itself, they opened up API access and priced it below cost. Significantly below cost. Eric Wise, who writes The Learning Foundry, puts the number at somewhere between eight and thirteen dollars of compute for every dollar of subscription revenue Anthropic collects. That's not a pricing strategy, that's a subsidy programme.

Now, cheap API access is catnip for builders. Startups showed up, raised venture money, hired engineers, and built real products on top of these models. Cursor built a coding assistant that developers genuinely loved. Figma wove AI capabilities into a design platform that was already the industry standard. These weren't toy projects. These were businesses with customers, revenue, and product-market fit that the labs themselves hadn't yet managed to demonstrate.

Which brings us to the part of this story I find particularly difficult to be polite about.

The moment a startup demonstrated that a use case actually worked — that real humans would pay real money for an AI-powered product — the labs had a decision to make. They could continue being infrastructure providers and watch the application layer be owned by someone else. Or they could look at what their customers had built on their backs, conclude that was actually a pretty good product, and build their own version.

They picked the second option. Every time.

Cursor showed the world that AI pair programming was a real thing people wanted. In 2025, Anthropic launched Claude Code. Figma proved that AI-native design tooling had a market. Last week, Anthropic launched Claude Design. The labs didn't need to do the hard, expensive work of product discovery, of finding out whether the idea was viable. Their paying customers did that for them. What a deal.

This is where the farm league analogy fits perfectly. If you follow baseball, you know how the minor leagues work. The farm system exists to develop talent for the major league clubs. The players do the hard work of getting good, and when they're ready, the big club calls them up — or doesn't, depending on whether they've got someone better. The farm team isn't a partner. It's a pipeline. The relationship looks collaborative right up until the moment it isn't.

The startups building on top of foundation models were operating as the farm league for the frontier labs, whether they knew it or not. Developing the product intuition, validating the customer demand, and handing the insight to the big club every time they hit the API.

But here's where it gets a bit darker, and where the term "Judas goat" earns its place in this conversation.

A Judas goat, if you haven't encountered the term, is a trained animal used in livestock operations to lead other animals into a slaughter facility. The goat is spared. The flock is not. The goat isn't malicious — it probably doesn't know what it's doing — but the effect is the same regardless of intent.

In platform economics, the Judas goat companies are the ones that get celebrated. Featured at keynotes. Co-marketed. Given early API access and warm introductions. Figma publicly bragged about their Claude Code integration earlier this year. That post is still up, which makes for somewhat uncomfortable reading now. Cursor was openly held up as a showcase for what you could build with Claude. The labs needed those stories — they needed proof that the API business was real, that the ecosystem was healthy, that developers should come build on their platform. The startups needed the visibility for fundraising and customer acquisition.

Both sides got something out of the arrangement. But only one side was gathering intelligence the entire time.

Because when a startup builds on your platform, you see everything. Which user segments are growing fastest. Which workflows create genuine stickiness. What the friction points are. What pricing the market will bear. You don't need to spy on anyone — the usage data tells the whole story. The startup is doing their product discovery with their own money, their own engineering team, their own customer relationships. And the signal flows directly to the infrastructure provider who will eventually compete with them.

If your ethics instict is beginning to twinge at this point, congratulations, you are human.

The conference keynote is worth watching as a leading indicator. Pay attention to when the ecosystem partners stop appearing on stage. When the slide deck shifts from "look what our amazing partners have built" to "look what our team has built, inspired by what we saw in the market" — that's usually about six months before the product announcement that ends someone's business.

Now, I want to be fair about intent, because the ethics here are genuinely complicated.

I don't think most of the people who built these platforms sat in a room and said "let's cultivate a bunch of startups, extract their market knowledge, and then destroy them." That would require a level of deliberate malevolence that I don't think is actually the right frame. What I think happened — at Twitter, at Facebook, at the LLM labs — is something more mundane and in some ways more troubling.

The founders mostly believed in the open ecosystem, at least initially. Then the capital stack arrived, and capital stacks have opinions. Nobody in a board meeting says "let's screw the ecosystem." What happens is the CFO presents a slide showing margin leakage to third parties. The board asks why those companies are making money on your infrastructure. Someone coins a phrase like "vertically integrating the value chain" and suddenly a philosophical position is a business decision dressed in neutral language.

Frances Haugen's[2] revelations about Facebook inspired me to think about this. I remember reading them and genuinely asking myself whether I could have sat in those rooms and participated in what was clearly, at some level, a decision to prioritise engagement metrics over human wellbeing. I don't think I could have. I can't really separate my personal ethics from my professional ones — never have been able to, which has occasionally been professionally inconvenient. But it is abundantly clear that a substantial number of my brethren in tech are just fine with it. Completely fine. Taking the money, sleeping soundly, updating their LinkedIn with the promotion.

The Facebook situation had documented internal resistance — people inside who believed that killing the developer ecosystem was strategically stupid, not just ethically wrong. They lost that argument. The growth team won[3]. The Zynga situation is almost textbook: Zynga's dependency on Facebook was listed as a risk factor in Facebook's own IPO S-1. Think about that for a moment. The platform had decided, before the ink was dry on the public offering, that the ecosystem dependency was a liability to be managed away. Zynga just hadn't been told yet.

The LLM version of this has one additional wrinkle that makes it, if anything, more ethically fraught.

Twitter and Facebook ran this play after they were already dominant and profitable. The LLM frontier labs are doing it while they are deeply, structurally unprofitable, underwriting the subsidy with investor capital. The startups building on that subsidised API pricing were, in effect, getting their unit economics handed to them on a temporary basis. The pricing made their business models look viable. When Anthropic raised API costs last summer — abruptly, with limited warning — it sent shockwaves through Cursor's user base almost overnight. The business hadn't changed. The subsidy had.

This is the bill that always comes due. The labs needed the ecosystem to prove the use cases. The ecosystem needed the cheap API to prove the business models. Both of those things were true at the same time, and neither side had to say anything dishonest for the whole thing to be a trap.

The product management lesson here is one of the oldest in the discipline, and it's almost embarrassing that it still needs to be said.

A defensible moat matters. If your core differentiator sits on infrastructure you don't control, you don't have a moat. You have a dependency. Those are very different things. Cursor's answer to this problem appears to be workflow depth and developer habit formation — genuine stickiness built around the experience rather than the underlying model. That might be enough. Might. Figma's real moat was always the collaborative workflow and the file format lock-in, not the AI layer — which is probably why Adobe was willing to pay a very large number to acquire them, before regulators had other ideas. But Claude Design isn't attacking the core workflow. It's attacking the AI-native layer that Figma just finished building on top of someone else's foundation. And that will likely be enough to stall Figma's rise.

If you're a product leader at a company where the AI capability is the product — where you're essentially a smart wrapper around a foundation model you don't own — this is the moment to have an honest conversation with your board about what your actual moat is[4]. Not what it should be or could be. What it is right now. And whether the company that provides your core capability has any reason not to build what you built, now that you've shown them it works.

The farm league develops the talent. The major club calls it up when it's ready.

The Judas goat leads the way in.

The bill always comes due. TANSTAAFL

1 - I use Antrhopic's Claude. I pay for it, because it does a superlative job for a specific task, and to me that is easily worth $20 a month. I am not a coder though, and I do not use any of the code generation bits, but some of my peers do. At best I am code adjacent.

2 - one has to assume that you shouldn't aspire to be a Product Manager who is a whistle-blower. All props to Frances on this.

3 - if I ever wake up at my job and find that there is a "Growth Team" that has captured the center of gravity, that is the day I pull the rip-cord.

4 - about a year or a year and a half ago, the circuit of product management influencers were raving about how easy and great it was that all you needed was an idea, an AI coding tool, and a grind ethic and you could start building a MRR growth engine.