The internet ran out. Your transaction data didn't.

The exhaustion of high-quality public training data is forcing a reckoning about where the next generation of AI capability comes from. For banks and health systems, the answer has been accumulating in their own systems for decades.

JPMorgan's Contract Intelligence system, known internally as COIN, can review commercial loan agreements and extract key provisions in seconds. The same work previously required approximately 360,000 hours of lawyer and loan officer time each year. What made COIN effective is not the sophistication of its architecture. It is the depth of JPMorgan's proprietary document corpus: decades of commercial loan agreements, legal precedents, and internal annotations built through the actual practice of one of the world's largest financial institutions. No amount of public internet data could substitute for that specific, structured, domain-grounded knowledge base. The competitive moat JPMorgan had built over decades of financial operations turned out to be exactly the asset that made their AI credible in production.

Most organizations haven't fully processed what that means for them. The sentence that's worth sitting with in 2026 is this: the supply of high-quality public training data for large language models is, by credible research estimates, running out. The banks and health systems that have proprietary domain data accumulated over decades are sitting on an asset they may not realize has become scarce.

The data wall is not a future problem

Epoch AI estimates an 80% confidence interval for the exhaustion of quality-filtered public text between 2026 and 2032. The front of that window is this year. Common Crawl, the nonprofit web dataset that has served as the foundational corpus for virtually every major language model, has been processed, filtered, and incorporated across multiple training generations. The marginal value of each new crawl is declining because an increasing share of what's on the internet is itself AI-generated.  

Frontier labs have responded by shifting substantially toward synthetic data, model-generated content used to extend training corpora. Gartner predicted in 2021 that by 2024, 60% of data used in AI and analytics projects would be synthetically generated, up from roughly 1% in 2021, a forecast that proved directionally accurate as adoption accelerated. This shift is rational as far as it goes. But synthetic data has a fundamental dependency: it is only as good as the real knowledge it is anchored to. And in regulated industries, where the value of training data comes from its grounding in actual domain practice: real transactions, real patient encounters, real claims. Synthetic generation without a strong human-knowledge anchor drifts from what's actually true.

What the new york times lawsuit actually revealed

The New York Times lawsuit against OpenAI, filed in December 2023, is widely discussed as a copyright dispute. It is also a data dependency story. The complaint documented the degree to which high-quality journalism (carefully reported, edited, fact-checked) had been central to what made the models capable of nuanced language and reasoning. What the litigation made visible is that the best text on the public internet came from a relatively small number of high-quality, professional sources, and those sources are now contested, paywalled, or simply exhausted as a new signal after years of incorporation.

The lesson for financial services and healthcare is not primarily about copyright. It's about the structure of knowledge. The text that made frontier models genuinely useful for complex reasoning was produced by humans who knew things, in contexts where accuracy mattered, under conditions where being wrong carried a consequence. A journalist's article, a legal brief, a clinical protocol: these forms of knowledge are valuable precisely because they were produced under conditions that enforced quality. Synthetic equivalents don't inherit those conditions by construction.

What domain-specific data actually means in banking and healthcare

For a health system, the irreplaceable training asset is not publicly available medical literature. It is the structured, longitudinal, outcomes-linked clinical data that has accumulated through decades of patient care: the EHR records that connect treatment decisions to patient outcomes, the clinical notes that capture physician reasoning in real cases, the operational data that reflects how care actually gets delivered. Health systems using Epic, for example, are embedded in a data network that represents some of the most valuable clinical training data that exists. Most of those health systems haven't operationalized it as a strategic AI asset.

For a bank, the irreplaceable asset is the transaction history, credit performance data, and customer behavior patterns that have accumulated across economic cycles, including the cycles where standard models fail. A model trained on a bank's actual loan portfolio, including the defaults and the recoveries and the edge cases that fell outside standard underwriting categories, understands credit risk in a way that a model trained on publicly available financial text cannot replicate.

The organizations that will have the most durable AI capability in financial services and healthcare are not the ones with the most access to general-purpose models. They are the ones that have built a disciplined strategy around their proprietary domain data, understanding what they have, ensuring it is structured and verified well enough to be useful for training, and partnering in ways that activate it without losing the governance and trust that make it valuable. The data wall is closing the gap between organizations with that strategy and those without it faster than most senior leaders have recognized.

For a mid-sized regional bank or a community health system that hasn't treated its accumulated operational data as a strategic AI asset, the practical question is immediate. The frontier labs that are training the models everyone will use in two years are actively seeking the kind of high-quality, domain-grounded, human-verified data that those organizations hold. The window in which that data has maximum leverage (as a negotiating asset, as a training advantage, as the foundation for genuinely differentiated AI capability) is not indefinitely open. First-party data strategies that haven't been started are increasingly difficult to catch up on.