The exclusivity clause always comes late

Posted on: 18 August 2026

No supplier asks for exclusivity while it is the only supplier. The clause appears at a particular point in the life of a dominant position, and it is never the beginning. At the beginning there is nothing to be exclusive against, and the constraint would simply be a cost imposed on a partner who had no alternative in any case. It appears once the alternative exists, works acceptably and costs less. At that point the incumbent moves the contest away from the ground it is losing, which is the product, and onto the ground where it still holds instruments, which is the contract. Anyone who has worked in a sector where distribution matters more than the thing distributed recognises the sequence without needing to be told.

Which brings us to a draft letter published by Reuters on 14 August. Prepared by the State Department for the thirty-five signatories of the AI Opportunity Statement signed in June, it contains an incompatibility clause: membership, the text says, cannot be held alongside participation in duplicative initiatives whose expectations conflict with American ones. China is nowhere named. An official, speaking anonymously, put it less delicately, saying that a country cannot have it both ways. The letter has not been sent, the draft carries no date, the State Department declines to comment on what it calls purportedly leaked internal documents and Reuters could not establish whether the text will be amended before dispatch.

The reason the sequence matters here is that the recording industry demonstrated its endpoint at considerable expense. Between 1999 and 2003 the majors won more or less every action they brought. Napster closed, Audiogalaxy closed, Kazaa pursued across three jurisdictions, thousands of individual proceedings against American users, copy protection on compact discs that rendered the discs unplayable on a substantial share of installed equipment. Every single action succeeded. Global trade revenues, on the IFPI wholesale series, fell from roughly 22 billion dollars in 1999 to 13.1 billion in 2014, which remains the low point of the series, and the curve of that descent records no inflection at any of the legal victories. The courts were not ineffective. They were extremely effective, on a layer of the architecture that had stopped being the decisive one.

This is where the analogy becomes a good story that happens to be false, so it is worth breaking the parallel before somebody else does it for us. Consider Intel. Between 2002 and 2007, on the European Commission's reconstruction, the company granted Dell, HP, Lenovo and NEC rebates conditional on near-exclusive purchasing of x86 processors, in a window that coincides precisely with the only period in which AMD was genuinely competitive. The Opteron had arrived in 2003 with 64-bit architecture and an integrated memory controller, and by the fourth quarter of 2006 AMD held 25.3 per cent of the x86 market. It then declined for a decade, to around eleven per cent by 2016, and returned to that level only in 2021 with Ryzen. The 1.06 billion euro fine came in 2009, when the contest was already settled, and in 2024 the Court of Justice annulled the rebate findings altogether. Here exclusivity held for fifteen years.

So the clause is not in itself an omen of defeat. It changes sign according to the nature of the good on which it is imposed, and the difference is not subtle. An MP3 is not an excludable good. Kazakh lithium is. Lithography machines are, high-end chips are, and on these layers an exclusivity clause bites, because physical flow can be counted, traced and interrupted, and whoever is left outside does not build alternative processing capacity within a quarter. Pax Silica exists for precisely this: securing supply chains in critical minerals, semiconductors and models, pushing allies and partners towards joint projects and export controls. Across two thirds of that perimeter the constraint works, for the same reason it worked for Intel.

Across the remaining third it does not work at all, because the weights of an open model are a file that can be downloaded. Alibaba passed a billion cumulative downloads on Hugging Face by March 2026, overtaking Llama as the most downloaded model family in the world. In February, Moonshot's latest open release was approaching leading proprietary systems on some benchmarks at roughly a seventh of the price, on the MIT Technology Review's analysis. No letter from the State Department reaches any of this. No clause about exclusive coalition membership prevents a developer in Almaty or Jakarta or Nairobi from downloading an Apache 2.0 licensed model and running it on hardware already sitting in the building.

The letter treats the two layers as though they were one. It asks for coalition loyalty in order to protect a position that on the excludable layer remains solid and on the non-excludable layer has been thinning for eighteen months. This makes perfect bureaucratic sense, since legislation happens where instruments exist rather than where the problem sits, and that is a constant which implicates the American administration no more than it implicates any apparatus holding levers that do not reach. The consequences, though, are not symmetrical. The advantage in minerals is measured in years; the advantage in models is measured in months and evaporates the moment somebody releases weights that are good enough. Two clocks running at different speeds, and the strategy that treats them together ends up spending the relational capital of the first to defend the second.

Relational capital is the part that gets consumed. A country signs an alignment on chips and minerals because it needs Western processing capacity and knows the need will last. When that same signature starts carrying obligations on a layer where it needs nobody, because the file downloads free to anyone with a connection, the total cost of membership rises while the benefit stays where it was. Nobody renegotiates at that moment. They simply do the arithmetic the next time round, which is the time that counts, and it is also why the letter, if sent in the leaked form, will most likely produce signatures rather than defections. Damage of this kind does not show up at the point of sale. It shows up at renewal.

Then there is the question of who receives the envelope. Thirty-five recipients, of whom exactly one is publicly known to belong to both coalitions: Kazakhstan, which is also a significant potential source of critical minerals. The others include Japan, Australia and South Korea, for whom the clause is a constraint at zero cost, since it formalises a choice already made and not under reconsideration. For Kazakhstan it is an ultimatum on which arithmetic can be done, and that arithmetic includes the fact that the Chinese side does not send letters of this kind, because its offer does not require loyalty in order to function. The mechanism is a club good. Access to shared investment projects holds value for as long as the cost of membership stays below the perceived benefit, so an exclusivity clause does not alter the benefit; it alters the cost, and it alters it asymmetrically, at almost nothing for the central member and substantially for the marginal one. The marginal member is usually the one the club was enlarged to reach.

The weak point in all of this should be stated rather than left to be found. If the draft is routine, if the State Department inserts incompatibility clauses into every technological framework it builds, then it signals nothing and I am reading intent into a template. The draft carries no date, which prevents it being tied with any confidence to the competitive sequence of the past eighteen months. There is one indicator, though, that requires no classified documents. Should Kazakhstan remain inside both coalitions over the next twelve months with no material consequence, the clause was noise and this reading belongs in the bin. Should it leave one of them, a great deal depends on which.

There is a detail from the recording industry that has stayed with me. In the autumn of 2004 Sony Music Japan and Avex, which two years earlier had been among the first anywhere to introduce copy-protected compact discs, began removing the protection. The stated reason was growing public awareness of copyright and the success of legal action. In the same weeks iTunes and MSN Music were preparing to open in Japan. Toshiba-EMI announced it would continue with protection, because that was what group global strategy required. It was the last to give up, in December 2006.