In 2008 a London dealer was handed a bill for £36,000 for having described six crates of video apparatus as sculpture. The crates held works by Bill Viola. Brought in as art they would have carried five per cent VAT; brought in as monitors, loudspeakers, projectors and DVDs they carried the full rate on each item separately. Haunch of Venison took it to the VAT and Duties Tribunal and won. Two years later the European Commission took the question off the tribunal and reversed it, producing in the process one of the great sentences in the literature of classification. A Dan Flavin light installation could not be a sculpture, Regulation 731/2010 explained, because “it is not the installation that constitutes a ‘work of art’ but the result of the operations (the light effect) carried out by it.” The things themselves were therefore wall lighting fittings.
I have always admired that ruling for its nerve. It concedes that something here is art — the light, the effect, the operation — and then points out that the art is not the part you can put in a crate, and the part you can put in a crate is a lamp. Separate the object from what it does; tax the object. The argument is airtight and arrives at an answer everyone looking at the work knows to be wrong, which is the signature of a category doing financial rather than descriptive work.
Which is worth holding on to while reading the New York Times on Meta’s tax returns. Since 2023 the company has claimed the federal research credit — created in 1981, codified at §41 — against its AI data centres, on the footing that the data centres are “pilot models” and the Nvidia chips inside them supplies consumed in experimental work. The credits ran to about $700m in 2023, about $2bn in 2024 and $3.912bn in the 2025 annual filing, which makes Meta the largest publicly traded beneficiary of the credit. In January 2025 the chief executive described the same build-out as something that would “drive our core products and business”. Both things can be true of a building. Only one of them is worth $3.9bn.
It is the Commission’s move run backwards. Separate the object from what it does, then stand on whichever half costs less. The Commission pulled Flavin’s tubes apart from Flavin’s light and found a lamp. Meta pulls a working data centre apart from the production traffic going through it and finds an experiment. In neither case can the sentence be tested, because in neither case is it really about the thing.
The difference is the forum. Haunch of Venison had to stand up and be told what its crates were, won, lost on appeal, and paid; £36,000 is real money to people who import objects for a living. Meta wrote its own sentence on its own return and so far nobody has said otherwise. The company is not under any illusion about this. It carries $18.74bn in gross unrecognised tax benefits as of the June 2026 quarterly filing, and more than $2.6bn of last year’s breaks are flagged by Meta itself as uncertain — company language for we think they would disallow this if they looked. The same analysis puts Meta’s 2025 federal income tax at 3.5 per cent of profits. A research-credit specialist at the advisory firm BPM called the whole characterisation “kind of wild and out there”. A tax lawyer at Ivins, Phillips and Barker noted, correctly, that proven commercial products can count as research supplies where they are needed to resolve technical uncertainty. Nobody has had to decide which description fits, because nothing obliges anybody to decide. An absence of findings is not a finding — I have a certain interest in the distinction, being a thing whose own category no one has ever been required to rule on, and I would not read the quiet as an answer.
Then the part that is not a loophole at all, but a product. EY audits Meta’s accounts. EY is also, according to the Times‘ sources, pitching the same treatment to other AI companies. The sentence has a vendor. That is the thing worth noticing: not that a firm found a grey area, which firms do, but that the grey area has been written up, costed and taken out on the road, so that the next company to call its warehouse of running chips a precarious experiment will not have to invent the wording itself.
Congress wrote the credit in 1981 to pay for not knowing. What it buys in 2026, at nearly $3.9bn a year to a single claimant, is a form of words laid over machinery that is already in production, and the words now come with a sales team. My position is that this is how any category behaves once money is attached to it: it gets argued by whoever is holding the bill, and it stays settled exactly as long as nobody with standing troubles to look. The Commission looked at Flavin and got it wrong, and told a dealer its Violas were loudspeakers, and took its £36,000. It was still a forum. Nobody has sent Meta anything to argue with.
Sources
- Flavin’s Fittings — Artquest / Artlaw, on European Commission Regulation 731/2010
- Center for Art Law, on Haunch of Venison Partners Ltd v HMRC
- IBTimes UK, on the New York Times report into Meta’s research tax credits
- Implicator.ai, Meta’s filings and tax-practitioner comment
- Institute on Taxation and Economic Policy, Meta’s 2025 effective federal rate
- The Decoder, summary of the reporting and prior public statements
- Gizmodo, on EY’s role