Courts are confirming that AI training data has commercial value, a largely unregulated market is already trading it, and four jurisdictions have built four different legal instruments to give data formal status. Only the Isle of Man's Data Asset Foundations Act combines a registered property right, independent valuation and dedicated enforcement under one statute already in force.
Looking back over this year, it's clear that data has stopped being a theoretical 'asset class' and started being litigated, auctioned, licensed and registered as one. Courts, bankruptcy brokers, national governments and accounting bodies have each, independently and for their own reasons, begun treating data, and particularly data used to train AI systems, as a distinct, ownable, commercially tradeable thing. What has not caught up is the infrastructure to handle it safely, which is resulting in a market moving faster than the mechanisms meant to govern it.
Courts are already calling it a market
It's been a busy month in the world of litigation relating to the use of data for AI training.
On 12 October 2026, a Third Circuit US federal appeals court unsealed its opinion in Thomson Reuters v. ROSS Intelligence, finding that Thomson Reuters had been developing 'a market for licensing [Westlaw] headnotes to train AI' and that ROSS Intelligence's unlicensed copying of those headnotes to train a competing legal research tool usurped that commercial opportunity. It is a narrow ruling, as the court distinguished its reasoning from generative AI cases more broadly, but it is a federal appellate court explicitly recognising that data assembled for the purpose of training AI constitutes a distinct, licensable commercial asset with its own market. ROSS's counsel, Yar Chaikovsky of White & Case, has since confirmed the company will seek Supreme Court review, arguing the decision 'creates continued uncertainty around the application of copyright law to AI model training'.
On 8 October 2026, USA Today Co. and thirteen of its affiliated newspapers filed suit against OpenAI in the Southern District of New York (USA Today Co. v. OpenAI Foundation, No. 1:26-cv-08892), alleging that ChatGPT was built and continues to run on hundreds of thousands of articles copied without licence from nineteen publications, and that it now substitutes for readers visiting the original sites. The suit seeks more than $250 million, up to $150,000 per wilfully infringed work, plus up to $25,000 for each instance of stripped copyright management information, along with an injunction and destruction of any models or training sets built on the plaintiffs' work.
These American cases show that the raw material of AI training has an identifiable market value that its originators believe the law should protect, and that the boundaries of that protection are still being actively fought over in the highest courts available to litigants.
An emerging market with imperfect foundations
While courts work through those questions, a separate and much less orderly market has sprung up to supply AI labs with training data right now. An MIT Technology Review investigation published 13 October 2026 documented a fast-growing trade in data stripped from failed startups. Google won a bankruptcy auction for Spirit Airlines' data for $10 million, roughly 100 million emails, 500 million Teams messages and 176,000 employee records, outbidding the AI-training-data buyer Mercor. Surge AI has been buying the archives of shut-down companies to build what the investigation calls 'reinforcement-learning gyms' for AI agents, and Bloomberg reported in September 2026 that SpaceX's AI unit was weighing similar purchases.
Two brokers, SimpleClosure and Sunset, have professionalised the pipeline. SimpleClosure alone says it has brokered more than 150 such sales, with code or data now present in 47% of its deals, and expects more than $10 million to flow back to failed startup creditors by the end of the year. But the investigation also found the pricing underneath these deals is largely improvised. One founder's codebase reportedly sold for $25,000 with no comparable transaction to benchmark it against. One attorney quoted in the piece described some of the underlying personal data as 'legally radioactive'.
There are more orderly examples of the alternative. The AI recruiting company micro1 announced in October 2026 a $1 billion, twelve-month commitment to license enterprise data directly from companies to train AI agents. And on 15 October 2026, the Warsaw-based music AI company GRAI signed two agreements with Songtradr's MassiveMusic, one for catalogue and metadata delivery under existing rightsholder agreements, and a second licensing a separate, rightsholder-approved opt-in dataset specifically for training GRAI's own foundation model, with usage-based payment flowing back to rightsholders. Both are consent-based and contractually clean, but both are also company-announced claims, built entirely on private contract, with no independent registry, valuation standard or third-party enforcement mechanism behind them. If either counterparty disputes scope, provenance or payment later, the dispute runs through ordinary contract litigation, on the same multi-year timeline as the publisher litigation above.
Governments are responding, but not in the same way
Several jurisdictions have tried to build something more durable than contract law alone, and they have made structurally different choices.
The Isle of Man
The Isle of Man's Data Asset Foundations Act received Royal Assent in late May 2026, described by its proponents as the world's first enacted statutory legal framework for registering, governing and commercialising data as a legal property asset.
China
China has taken a different route. Its National Data Administration moved a national data property registration system fully live on 15 September 2026, and on 22 September issued its first batch of registration certificates to 56 enterprises and institutions across more than ten industries. This is a state-administered certification scheme, issued through vetted registration agencies, rather than a judicially enforceable property right. This is a structural distinction from the common law and statutory property interests that are being established in England and Wales and the Isle of Man respectively.
It also remains ahead of the financing market it was meant to support, as evidenced by a Caixin investigation published 5 June 2026. This found that Chinese banks still largely treat pledged data as credit enhancement rather than standalone collateral, and that regulators quietly paused new data asset-backed securities filings on 4 June 2026 over concern that financially stressed local government financing vehicles were using them to repackage non-standard debt. A genuine, named data collateral precedent does exist from an earlier, narrower pilot: the Shenzhen Data Exchange brokered a China Everbright Bank loan to the AI firm Shenzhen Weiyan Technology, secured against its data products, in 2023. The Caixin reporting suggests that this remains the exception rather than the norm, even inside China's own, more developed registry.
Jersey
Jersey has gone a third way, applying existing trust law rather than new statute. Digital Jersey's data trust initiative traces back to a 2021 working group, and the Jersey Financial Services Commission regulated Jersey Data Trust Company Limited describes itself as the first independently regulated trust company established specifically to act as fiduciary trustee of data-related rights, already operating in identity verification and anti-money laundering use cases. That model governs access through a fiduciary relationship rather than creating a registered property right.
Vietnam
Vietnam has taken yet a fourth approach: its 2024 Data Law, in force since July 2025, recognises data as a property right administered through licensed, state-owned data exchanges.
Four jurisdictions, four different legal instruments, all built on the same underlying premise: that data needs a formal legal status beyond 'whatever the contract says'.
Even the accountants are catching up
The clearest sign that this is no longer a niche question is that the people who measure national and corporate wealth for a living have started treating data as something that belongs on the books. The US Bureau of Economic Analysis has published a working paper formally capitalising own-account data production in the national accounts, estimating this adds roughly 8.2 basis points to measured annual US GDP growth. A US-China Economic and Security Review Commission report has urged Congress, the SEC, FASB and the IASB to formally examine data's balance sheet status, pointing to 136 Chinese listed companies that have already booked a combined 3.8 billion yuan of data as loan collateral on their own accounts. The International Accounting Standards Board's own review of IAS 38 (intangible assets) as it applies to data remains at an early discussion stage, with no decisions yet taken, a reminder that, for all the activity elsewhere, the core international accounting question of how to recognise data as an asset is still unresolved.
The Isle of Man opportunity
Set side by side, these developments describe a single, consistent problem rather than several unrelated ones. Courts are confirming that training data has commercial value worth fighting over, but taking years to settle what rights attach to it. A real market is already pricing and trading that data today, largely through bankruptcy proceedings and with no independent valuation standard. The cleaner, consent-based licensing deals that exist are built entirely on private contract, with no registry or third-party enforcement behind them.
What none of the ad hoc or part-way models provide, together, is a registered legal property right, an independent valuation mechanism, and a dedicated enforcement function, all operating under one statute that is already in force. The Isle of Man's Data Asset Foundations Act provides all three of these requirements.
Frequently Asked Questions
Courts, bankruptcy brokers, national governments and accounting bodies have begun treating data, particularly data used to train AI, as a distinct, ownable and tradeable asset. The infrastructure to handle it safely has not caught up, so the market is moving faster than the mechanisms meant to govern it.
In an opinion unsealed on 12 October 2026, the court found that Thomson Reuters had been developing a market for licensing Westlaw headnotes to train AI, and that ROSS Intelligence's unlicensed copying of them to train a competing legal research tool usurped that opportunity. ROSS has said it will seek Supreme Court review.
Through bankruptcy auctions and specialist brokers such as SimpleClosure and Sunset. Google bought Spirit Airlines' data for $10 million at auction, but an MIT Technology Review investigation found that pricing in these deals is largely improvised, with no comparable transactions to benchmark against.
The Isle of Man has enacted a statutory framework for registering data as a property asset. China runs a state-administered registration and certification scheme. Jersey applies existing trust law through a regulated data trust company, and Vietnam recognises data as a property right administered through licensed, state-owned data exchanges.
Not fully. The US Bureau of Economic Analysis has published a working paper capitalising own-account data production in the national accounts, but the IASB's review of IAS 38 as it applies to data remains at an early discussion stage, with no decisions yet taken.
A registered legal property right, an independent valuation mechanism and a dedicated enforcement function, all operating under one statute that is already in force.
This article is for general information only and does not constitute legal, tax or financial advice. You should take independent professional advice on your own circumstances. Manavia Limited is licensed by the Isle of Man Financial Services Authority to provide Corporate and Trust Services.
