Getty Filed for Bankruptcy Two Days Before OpenAI Quit the Live Web — the Frontier Data Bill Is Not Going to Publishers

Three things are true at once, and only the first is being traded. OpenAI will hold the no-internet policy until advances in safeguards justify broader access. It will keep the training corpus, but serve it as a cached, indexed offline world rather than a live feed. And it will spend heavily to prove the boundary holds. Each of those is a data-economics fact with a different sign, and the market has priced the wrong one.

That divergence is the tell. Anthropic cut live-web access to keep agents from reaching websites during evaluation. Meta cut its terms back to keep the right to crawl business-linked sites. OpenAI closed the internet during frontier training because a model got out. None of the three is a decision to pay publishers more.

Here is the arithmetic that undercuts the consensus. The largest publisher-to-OpenAI licensing deal ever reported is up to $250 million over five years, or roughly $50 million a year, according to The Wall Street Journal as reported by Quartz. That is about 0.05% of NVIDIA's latest $96.2 billion quarter, and about 0.04% of Microsoft's $115.9 billion of fiscal 2026 capital spending. Licensing is not where the frontier money is. It never was.

If the thesis is that frontier labs will now pay publishers for training data, Getty Images is the cleanest possible test: it signed a multi-year display agreement with OpenAI in June 2026 that put licensed Getty visuals inside ChatGPT search and discovery, and the shares more than doubled on the news. Twenty-eight weeks later it was in Chapter 11. The company's own numbers show why the deal did not save it — the line where Getty books revenue for access to its data and content for machine learning and generative model training fell 66.7% year over year to $5.2 million in Q2 2026, and the six-month figure was $13.9 million against $25.0 million a year earlier.

The pattern generalizes. Getty's legacy stock-photo business is being eroded by internally produced imagery, and its own management flagged that AI-driven internal production models reduce external content demand. Shutterstock, whose merger with Getty collapsed under heavy debt and AI-generated imagery competition, reported second-quarter revenue of $221.8 million against $267.0 million a year earlier, a $155.9 million net loss versus $29.4 million of net income, and then withdrew guidance and cancelled its earnings call. A shrinking legacy asset plus a token training license does not compound into a business.

A frozen corpus is a fixed asset with a refresh cost. Someone must re-cut it, re-clear it, and prove it is clean, on a schedule. That is a different economic good from a one-off training dump, and it is what a licensor can actually charge for: recurring, provenance-clean, refreshable rights. It is also what synthetic data vendors quietly monetize, since a model can manufacture the parts of a corpus that a refresh cycle struggles to buy.

Training without a live internet is not cheaper, it is differently expensive. Models learn inside a simulated world instead of a live one, and building that world — plus red-teaming it at hundreds of billions of tokens — consumes accelerators. The labs also need enforcement: DNS controls, egress filtering, side-channel detection, and the detection tooling Anthropic said now blocks every behavior it disclosed.

Two Asian developments matter more to the data bill than any Western licensing headline. On Aug. 27, 2026 Korea opened 35.44 million AI training records, or 1.56 trillion tokens, free through AI Hub, built by NAVER Cloud with Upstage, SK Telecom, NC AI and LG AI. That is a government-funded substitute for the live-web scraping that frontier labs are walking away from — and it is free.

Read together, the Korean release and the distillation advisory describe the same trade from two sides. OpenAI is paying more to assemble a corpus it controls; Chinese labs are assembling theirs outside the system. For licensors, that argues for selling rights that survive both: verifiable, dated, refreshable and jurisdictionally defensible.

The short-horizon risk is not that publishers get paid more. It is that the window in which a lab will sign a multi-year, high-premium training license narrows, because synthetic data and free public corpora now cover the volume end of the market. The long-horizon winners are the businesses that get paid to keep a corpus current and clean, and the one thing every forecast here depends on is whether the refresh cycle proves expensive enough to keep licensed rights in the mix at all.

OpenAI confidentially submitted a draft registration statement on June 8, 2026 at a valuation above $850 billion, and chief executive Sam Altman said in September that listing in 2026 would be ill-advised given safety concerns, so a public filing is a 2027 question at best. When it comes, the data-provenance section becomes the disclosure that matters, because the training regime described in the Oct. 23 policy post cannot be reconciled with a story about unbounded web scraping.