The compounding enterprise

On July 20, a federal judge gave final approval to the largest copyright settlement in U.S. history, requiring Anthropic to pay $3,000 per work for roughly 500,000 books it pulled from pirate libraries to train its models. The court had already ruled that training AI on copyrighted text is fair use, but the settlement was decided on the provenance of the data, and Anthropic agreed to destroy the pirated files. The author argues that this settlement exposes the liability of unverified data in AI systems, coining the term "Verification Debt," and advocates for enterprises to build "compounding intelligence" through curated, governed data flywheels.

On July 20, a federal judge gave final approval https://arstechnica.com/tech-policy/2026/07/judge-approves-anthropics-1-5-billion-copyright-settlement-with-authors/ to the largest copyright settlement in U.S. history. Anthropic will pay $3,000 per work for roughly 500,000 books it pulled from pirate libraries to train its models. Here’s the detail most of the coverage missed: The court had already ruled that training AI on copyrighted text is fair use. The capability was legal. The provenance was not. Read that again. The industry’s defining legal battle wasn’t decided on what the model could do. It was decided on whether anyone could account for where its knowledge came from. And buried in the settlement is a stronger signal: Anthropic agreed to destroy the pirated files. A frontier model’s corpus can now shrink by court order. Subtract half a million books from the foundation and ask yourself: Is the model your teams rely on tomorrow as capable as it was yesterday? This isn’t a story about one company. The settlement set no binding precedent, and Google, Meta, OpenAI and Midjourney are still in the dock. Meanwhile, every organization deploying AI is accruing the same liability in miniature: Data you can’t trace, sources you can’t verify, outputs you can’t attribute. I call this Verification Debt. It never appears on the balance sheet — until it appears all at once, with a court date attached. Some of us will remember July 20 as the day the GenAI bubble popped. The day generic models were exposed as a depreciating, legally contested input. The right response isn’t caution. It’s ambition. The enterprises that win the next decade will stop renting intelligence — and start compounding their own. For the past year, I’ve been building frameworks and systems for companies chasing the same vision: The AI-native organization. The executives I work with have already bought in. What frustrates them is the gap between that vision and the daily reality of unreliable, generic models that know everything in general and nothing about their business in particular. Closing that gap comes down to two ideas. The first is the data flywheel . Jim Collins gave us the metaphor and Amazon turned it into an operating model: Every use of a system improves its data, and better data improves every future use. Each turn of the wheel lowers the cost — and raises the value — of the next one. The second is compounding intelligence . Think of it as compound interest applied to what your organization knows. When knowledge is captured, verified and fed back into the systems making decisions, every decision inherits the full history of the decisions before it. Intelligence stops depreciating and starts earning interest. Now the catch, and it’s the real lesson of July 20: Compounding doesn’t care about the sign. A flywheel fed unverified data compounds errors Collins’ “Doom Loop” just as efficiently as a governed one compounds value. Hallucinations become tomorrow’s context. Drift becomes tomorrow’s baseline. I call this the feral flywheel, and the pirated corpus at the center of the settlement is what one looks like at industry scale. A verified flywheel is different by design. Its answers are grounded in curated, human-authored knowledge. Its scope is deliberately limited. Its outputs trace back to their sources. And when the data goes stale, somebody knows. Which brings me to the sentence I ask every executive team to sit with: Curation and data governance aren’t overhead on the flywheel. They are the flywheel. “The model is no longer the product,” tweeted https://x.com/gdb/status/2057670776803996110?lang=en OpenAI’s Greg Brockman — and on a recent episode of the 20VC podcast https://www.youtube.com/watch?v=OxFyVcO1Yow , Perplexity CEO Aravind Srinivas explained why he’s right. If you’re just reselling raw model tokens, Srinivas argues, “you have no business because the model will get commoditized.” The business belongs to whoever takes the model, grounds it in valuable context, orchestrates it with the right tools and delivers one unified system. In other words: The durable asset isn’t the model. It’s the system around it — and the context you feed it. The settlement sharpens the point. If every competitor is renting the same commoditized models — with the same generic knowledge and the same legal exposure baked in — the only defensible asset left is the proprietary flywheel you feed them. Your experts’ judgment. Your decision history. Your verified corpus. If that sounds like your RAG pilot, it isn’t. Retrieval is a one-way pipe. A flywheel closes the loop — verified outcomes flow back into the corpus — so the system answering questions in the fourth quarter is smarter than the one you deployed in the first. This isn’t theoretical. I’m watching it play out in a place few people expected: corporate training. Companies are quietly replacing legacy training and development with digital twins — governed AI versions of their own experts. The economics are hard to argue with. A course library starts aging the day it’s recorded. A twin stays current because it’s connected to the living, curated knowledge of the organization. And a twin meets every learner where they are. The same expertise translates on demand — technical depth for engineers, plain language for the board, precise wording for legal. That puts established platforms like Udemy, LinkedIn Learning and KnowBe4 squarely in the displacement zone. Just-in-time learning from your own verified corpus is beating just-in-case content licensed from someone else’s. Notice what the early winners have in common. They aren’t buying better models than their competitors. They’re building better systems around the same models — systems that get smarter with every use because the organization’s own knowledge is doing the compounding. Here’s a prompt for your next board meeting: If the model is now a commodity, what durable system is our institution building around it — and would a competitor recognize it as a moat? Most AI business cases project ROI. Flywheel economics build it in. The logic is clear. A flywheel that adds verified value on every turn is ROI-positive by definition. Each governed cycle lowers the marginal cost of the next trusted decision and raises the value of the underlying data asset. The only losing move is feeding it ungoverned data. None of this is free. Curation takes money and governance takes headcount, up front. The difference is what happens next: The spend amortizes, and the asset appreciates. Three returns show up fastest. But the deepest return is one I wrote about in my book, The CISO on the Razor’s Edge: Curing corporate amnesia. Every organization suffers from it. Expertise walks out the door. Decisions get re-litigated because nobody remembers why they were made. The same lessons are repurchased, year after year, from consultants and course vendors. I’ve watched million-dollar decisions stall for weeks, waiting on knowledge the company already had. A data flywheel is institutional memory with a compounding rate of return. It consolidates knowledge trapped in silos and delivers it back at the moment of need. The visible results are faster cycle times and accelerated just-in-time learning. The invisible result is bigger: An organization whose baseline rises every quarter instead of resetting. The good news is that you don’t start with a big program. You start with one governed turn of the wheel. Three moves, in order. None of this requires waiting on the next frontier model. That’s the point. The model is a commodity; your knowledge is not. The AI industry just learned the price of unverified data: $3,000 per work. The question for every CIO is simpler, and it deserves a place on the board agenda. What is your organization’s knowledge worth — and are you compounding it, or renting someone else’s? This article is published as part of the Foundry Expert Contributor Network. Want to join?