I have read more commercial leases than a person outside real estate law ought to, most of them under conditions that did not allow the luxury of reading slowly, and every one of them taught me the same thing: the interesting sentence is never the one about rent. Rent is what everyone negotiates, because rent is what everyone understands. The interesting sentence sits two-thirds of the way down, past the section anybody reads aloud at signing, and it governs what happens when somebody wants out. Termination for convenience. Assignment and subletting. Holdover, and what the rate becomes if you are still in the space on the first of the month after the term ends. Those clauses decide whether the document is a relationship or a trap, and they are drafted by whichever party expects to still be in the room when the other one wants to leave.

So when the State Department drafts a letter about coalition membership and the operative word turns out to be duplicative, my first question is not about the politics. It is about who wrote the exit language, and whether anybody wrote it at all.

The letter has not been sent. Michael Martina reported it for Reuters on 14 August 2026, circulating in draft, addressed to the thirty-five signatories of the AI Opportunity Statement;a wider circle than the roughly twenty-five states that have actually signed the Pax Silica Declaration. Pax Silica is the American-led coalition the State Department stood up on 12 December 2025 with nine founding members, built around a shared standard for artificial intelligence infrastructure. The draft letter expels nobody. It announces that membership is henceforth reviewable, and it warns that signature of the Declaration “is not merely a membership subscription, but a commitment,” one that “cannot be held alongside membership in duplicative initiatives whose expectations conflict with our own.”

Membership subscription. Somebody at State reached for the vocabulary of a gym contract, heard it, and evangelized it to commitment. I find that revealing in a way the drafter likely did not intend, because a subscription is a thing you cancel, and the correction exists precisely because the reader had to understand that this one cannot be. The drafter needed a diplomatic word for irreversibility and could not find one that would survive being read aloud, so he wrote commitment and moved on. Duplicative was the compromise. In ordinary administrative English the word describes a redundant filing cabinet. Here it carries a loyalty test with a euphemism sitting on top of it.

One month earlier, on 16 July 2026, twenty-nine states met in Shanghai and signed the founding agreement of the World Artificial Intelligence Cooperation Organization, which anyone who has had to write about it twice now calls WAICO. Foreign Minister Wang Yi signed for Beijing. UN Secretary-General António Guterres attended, which mattered rather more than the signature did. The agreement invokes the UN Charter, stands up working groups, and pledges coordinated standards development for artificial intelligence deployment (World Artificial Intelligence Cooperation Organization, 2026).

It also contains an exit. Article 17, headed Withdrawal: a member state may leave by written notification to the depositary, the withdrawal taking effect one year after the date of receipt. Termination for convenience, twelve months’ notice, no cause required, no penalty stated. This is the exact clause the American letter conspicuously does not contain.

Five months before either of those, in New Delhi, a summit produced a declaration whose first pillar is democratizing AI resources, and which takes note of a companion document, the Charter for the Democratic Diffusion of AI, as a voluntary and non-binding framework (Government of India, 2026). When the endorsement list was published on 21 February it ran to eighty-eight names. Three more, Bangladesh, Costa Rica, Guatemala, signed on afterward, bringing it to ninety-one, where it has stayed. On that original list of eighty-eight, China appears fourteenth, Russia sixty-fourth, the United States eighty-fifth. Nobody had to be talked into it. Endorsement of the Charter, the text says, “does not imply legal, financial, operational, or governance obligations,” which is to say you can sign it standing next to the people against whom you are actively building an export control regime, and eighty-eight parties did exactly that on the same afternoon.

Then there is Kazakhstan, which signed the Pax Silica Declaration on 25 June 2026, becoming the first Central Asian state to do so, and the WAICO founding agreement three weeks later. It remains the only state known to have signed both.

The draft letter exists because of Kazakhstan’s status. Somebody in Washington read the July signature list, found a name that was already on the June one, and started typing. When a state accepts two invitations and discovers that one of them reads the other as defection, the governmental machinery does not answer with a declaration. It answers with a diplomatic condition, and a condition is a lease clause wearing a flag.

---

The Sentence Nobody Quotes

The resemblance between 2026 and 1947 is the least interesting thing on offer, and people reaching for it are usually reaching for a mood. Cold War is a mood. It supplies dread without supplying a mechanism, which is why it gets deployed most often by people who have not read the statutes. What actually transfers across the seventy-nine years is smaller and considerably harder: a specific instrument, assembled between the Truman Doctrine and the Mutual Defense Assistance Control Act of 1951, since rebuilt on a foundation that nobody can unplug.

Harry Truman, addressing a joint session of Congress on 12 March 1947, declared that at the present moment in world history nearly every nation must choose between alternative ways of life. That sentence gets quoted regularly. The sentence immediately after it gets quoted almost never: The choice is too often not a free one. The doctrine announces a binary and then, in the following breath, admits that the binary is being imposed by the people announcing it. I have wondered for years whether the speechwriter wrote it that was as honesty or as inoculation, and I have never been able to decide, which is probably the point.

George Marshallwhile speaking at Harvard on 5 June 1947 (not as commencement speaker, as one of twelve honorary degree recipients addressing about fifteen thousand people in the Yard that afternoon) put the disavowal in its permanent form: Our policy is directed not against any country or doctrine but against hunger, poverty, desperation and chaos. Section 103(b) of the Mutual Defense Assistance Control Act, signed 26 October 1951 and named for Laurie Battle, the Alabama Democrat who carried it in the House, followed Marshall by fifty-two months.

Under Secretary Jacob Helberg, the architect of Pax Silica and the man who ran its December launch, published an essay on 23 June 2026 arguing that the initiative was built not as a fortress but as a coalition of capabilities. The draft letter stating clear exclusionary language followed him by fifty-one days.

Fifty-two months, then fifty-one days. Everything about the mechanism is unchanged except the interval between the reassurance and the instrument, which has compressed by a factor of roughly thirty. Whether that reflects institutional efficiency or institutional impatience I leave to somebody with better access than mine.

The consequential feature of the American instrument was never the embargo. Embargoes are loud, dated, comparatively easy to legislate around, and they generate their own opposition because everybody can see them. Section 103(b) terminates assistance automatically: all military, economic, or financial aid to any nation “shall, upon the recommendation of the Administrator, be terminated forthwith” if that nation, sixty days after the Administrator’s determination, knowingly ships controlled items eastward. And then, in the same subsection, the exception:

Provided, That the President … may direct the continuance of such assistance to a country which permits shipments of items other than arms, ammunition, implements of war, and atomic energy materials when unusual circumstances indicate that the cessation of aid would clearly be detrimental to the security of the United States.

Followed by the paperwork, which is where the loophole actually lives. The President must report any such determination immediately, with reasons, to the Appropriations and Armed Services Committees of both chambers, the Senate Foreign Relations Committee, and the House Foreign Affairs Committee (four designations, six actual committees) and then, “at least once each quarter,” review every determination previously made and report his conclusions again to the same six. Those quarterly reports “shall contain an analysis of the trade with the Soviet bloc of countries for which determinations have been made.”

Read that last clause slowly. The waiver did not merely permit an exception; it generated, four times a year, a standing analytical product on the trade behavior of every country living inside one. The exception is a subscription to surveillance, paid in the currency of continued aid.

Automatic in the first clause, discretionary in the second, permanently documented in the third. Call it a valve if the plumbing analogy helps; a valve is operated rather than closed, which means the job is never finished, which means the country on the other end of the pipe is never finished either. Every quarter, somebody in an office decides again, and writes down what you have been buying.

The draft letter of August 2026 is written in that grammar, by people who may or may not know they inherited it.

Two Bandungs

Before going further I have to clear something away, because both Washington and Beijing have spent two years invoking a precedent that did not happen in the form anyone remembers it.

The sharpest statement of this is sixty years old. G. H. Jansen, in Afro-Asia and Non-Alignment (1966), wrote that two conferences were held at Bandung in April 1955:

One was the real conference, about which not very much is known, about which people care even less, and which has faded away like a bad dream. The other was a quite different conference, a crystallization of what people wanted to believe had happened which, as a myth, took on reality in the Bandung Principles and, later, in the Bandung Spirit.

Robert Vitalis put that passage at the head of his 2013 essay on Bandung’s fables, and called Jansen’s book a still-unsurpassed postmortem, which after six decades it more or less is. The scholarship since has not been kinder to the myth. Lorenz Lüthi’s work on the movement describes states acting on national interest and calling the aggregate a bloc; Jürgen Dinkel’s history of the Non-Aligned Movement traces an organization that never acquired a permanent secretariat or a constitution, and that ran on summits and communiqués because summits and communiqués were what it could afford to produce.

If the Bandung of memory is a retrospective construction, then what is worth learning from Bandung is not its spirit. It is its defeat, and the defeat has a public-law citation, which the spirit never did.

The forcing instrument of 1947 arrived dressed as an invitation. The Paris conference of July 1947, convened to organize European participation in the Marshall Plan, functioned as a test with an answer key. On 4 July the British and French ambassadors extended the invitation in Prague, and the American chargé Laurence Steinhardt cabled Washington that Masaryk accepted it immediately, adding that the Czechs “are extremely anxious to participate in Marshall plan and will do so unless forbidden by Soviets.” The delegation left for Moscow the following day. Participation was publicly announced on 7 July.

On the night of 9 July, Stalin received them. Gottwald’s own telegram home (which Steinhardt obtained and transmitted the next evening )records two audiences, the second at eleven o’clock, with Masaryk, Drtina, Heidrich and Ambassador Horák present on the Czechoslovak side and Molotov on the Soviet. From the Soviet record of that conversation, Stalin’s instruction:

You need to rescind your decision, you need to refuse to participate in this conference, and the sooner you do it — the better.

And the loyalty framing that rode alongside it, as it always does:

For us, this issue is the issue of friendship between the Soviet Union and the Czechoslovak Republic. Objectively, you are helping — whether you want it or not — but you are helping to isolate the Soviet Union.

Gottwald’s telegram asked Prague to withdraw the acceptance and to time the communication so that it would reach Moscow officially by Thursday afternoon. Steinhardt’s cable of 10 July records the cabinet’s compliance in a clause I have not been able to stop thinking about: after a lengthy debate a decision was reached to comply, and no vote was taken.

Jan Masaryk (foreign minister, son of the republic’s founder) is supposed to have said on returning that he had left as a minister of a sovereign state and come back as Stalin’s lackey. The line is reported rather than documented; it reaches us through Ladislav Feierabend’s memoirs, recording what Masaryk said to Ivan Herben, with a competing version from his secretary in Claire Sterling’s account of the death. Whether he used those exact words is unrecoverable. That he was in a position to say them is not.

He went out of a window at the Foreign Ministry on 10 March 1948, two weeks after the communist takeover. Four inquiries have now examined it. The 1948 verdict was suicide; a 2003–04 forensic reinvestigation concluded murder; the file was reopened in 2019 and closed again on 5 March 2021 with the finding that the assistance of another person in the fall can be neither confirmed nor refuted. Seventy-eight years, and the state that owns the building cannot say how its foreign minister left it.

Kazakhstan in August 2026 is Czechoslovakia in July 1947 with the polarity reversed and the coercion softened into a draft nobody has sent, but many people have read. The structure is identical: two invitations, one of which reads the other as defection, and a state made to unsign. No declaration of war is required. A classification will do, and the classification is usually assigned before the state in question has done anything at all.

Andrei Zhdanov, presenting the founding report at Szklarska Poręba in late September 1947, split the world into two camps; and then, describing the anti-imperialist one, wrote down which countries belonged to it whether or not they had said so. Indonesia and Vietnam, he said, are associated with it. And it “has the sympathy of India, Egypt and Syria.”

Has the sympathy of. Not might be persuaded, not remains undecided. The uncommitted were not being courted; they were being entered into a ledger as already counted. The conference resolution was adopted on 27 September and released through Pravda on 5 October, binding nine parties to a position that had allocated three non-participating states without asking them. Neutrality was never denied as a position. It was simply never recognized as one, because both camps read the unaligned as inventory. Inventory requires accounting.

The accounting is what the current literature on the AI divide is missing, and the Battle Act is where it sits; which is why I quoted 103(b) at length above rather than paraphrasing it. It was also never a European arrangement, a point the Marshall Plan framing has eaten. On 19 December 1951 the State Department handed the Embassy of Thailand an aide-mémoire applying the identical conditionality to Southeast Asia. It sets out the Administrator’s determination of 25 November: Category A, twenty-one item listings covering arms, ammunition, implements of war and atomic energy materials; Category B, two hundred sixty-four listings covering petroleum, transportation materials, and items of primary strategic significance. It names the date the sixty-day clock ran out — 24 January 1952 — twice. And it asks, before that date, whether the Government of Thailand will continue the embargo.

Two hundred eighty-five item listings and a deadline, delivered to a government whose alliance structure looked nothing like NATO’s and which had no illusion it was being invited into a partnership of equals.

Now read the Bandung text against that. The conference met from 18 to 24 April 1955 with twenty-nine countries present, and its Final Communiqué produced ten principles, the sixth of which has been quoted ever since. Subsection (a): Abstention from the use of arrangements of collective defence to serve the particular interests of any of the big powers.

On one side of the table, a moral abstention. On the other, an automatic termination clause with a reporting schedule attached and two hundred eighty-five items on the list. The movement did not fail because its members were insincere or because its historians were romantic. It failed because a principle cannot answer a statute, and the participant list proves the point better than any argument I could construct. Pakistan, Thailand and the Philippines sat at Bandung as SEATO members, the Manila Pact being eight months old. Turkey sat there as a NATO member of three years’ standing, and also as one half of the Baghdad Pact, which at that moment consisted of Turkey, Iraq and the United Kingdom. Pakistan would join in September, five months after the conference, and the Americans never joined at all. The Democratic Republic of Vietnam sat there as participant twenty-two, with the State of Vietnam at twenty-three.

Even the State Department’s own account concedes the pressure. Its history of the conference records that leaders of developing countries “banded together to avoid being forced to take sides in the Cold War contest.” Banded together. Forced.

Power and Cooling

Laurie Battle spent six weeks in the autumn of 1952 touring eight Western European countries to see how the statute bearing his name was working in practice. On 29 September he wrote to Averell Harriman, then Director for Mutual Security, to report what he had found.

The European members of COCOM — the Coordinating Committee for Multilateral Export Controls, the informal Paris body run out of an annex to the American embassy on the Rue de la Boétie, through which the Western allies harmonized what they would and would not ship east, had refused to agree to embargo four items on his list. Four. They had agreed to consult before any export, and nothing had actually shipped since the Act took effect, but they would not agree to the prohibition itself. Meanwhile the United States was unilaterally embargoing twenty-eight items its allies remained free to sell. The Consultative Group sitting above COCOM had met three times that year, and Battle complained that COCOM itself “concentrates primarily on low level discussions of details.”

His summary judgment on the administration of his own act, in his own words: Therefore, it seems to me that this entire situation is indefensible.

Michael Mastanduno’s Economic Containment (1992) is the book-length version of Battle’s letter. American leverage over allied export policy in this period was genuine and it was bounded, and the boundary was negotiated item by item. Ian Jackson has a chapter that could not be more precisely on point if it had been commissioned: “Troubled Partners: Anglo-American Relations and the Battle Act in 1952.” Bloc discipline in that year got haggled over line by line, because the leverage was fiscal and quarterly. Aid could be cut. Licenses could be revoked. Both are reversible inside a budget cycle, and everyone at the table knew it, including the man whose name was on the statute.

Vili Lehdonvirta, Bóxī Wú and Zoe Hawkins mapped where machine learning compute physically sits and found a world in three parts: a Compute North hosting capacity relevant to training, a Compute South whose capacity is relevant mainly to inference, and a Compute Desert hosting no public cloud AI compute at all. Thirty-nine countries have a public cloud region of any kind. Thirty of those host regions equipped with GPUs. The North is seventeen countries; the South is thirteen. The Desert contains every lower-middle-income and low-income country on earth, by the IMF’s classification.

One figure from that paper has stayed with me longer than the headline taxonomy. Latin America hosts five GPU-enabled cloud regions across the entire continent, and not one of them features anything more powerful than the V100; a card Nvidia announced in 2017. A hemisphere is running on eight-year-old silicon. On 20 February 2026 the State Department announced a pilot concierge service, designed to leverage the Department’s presence across approximately 270 diplomatic posts in order to streamline acquisition of American-made artificial intelligence products for Pax Silica signatories – power, cooling systems, software and hardware.

Power and cooling. Somewhere between the Battle Act and last February, the practiced mechanism stopped licensing shipments of specific goods, and started siting buildings, and it did so without anybody announcing the change. The same office simply found itself procuring transformers instead of reviewing manifests.

Section 103(b): shall be terminated forthwith. The Charter for the Democratic Diffusion of AI, addressing the identical subject of access to compute: consider, where relevant, innovative financing assistance; explore mechanisms to facilitate accessible and affordable computing resources; promote high-quality datasets; encourage openness; endeavor to enable benefit-sharing; seek to support human resource development.

Consider, explore, promote, encourage, endeavor, seek. Six verbs, not one of which has ever moved a transformer across a trade boundary. They govern aspirations. The other one governs allocations, and it does so in the present indicative, which is the frame you use when the thing has already happened.

The Charter’s implementation vehicle is MAITRI (Multistakeholder AI for Trusted and Resilient Infrastructure) a platform to be built and maintained by the Government of India as a digital public good, enabling accessibility of existing compute capacity in the public and private sectors and facilitating price transparency and discoverability. It is a directory with a price column. That is a real piece of infrastructure built by people working under real constraints, and I do not want to be glib about it. It is also, in the situation actually confronting a state in the desert, about as decisive as a well-maintained phone book during a blackout.

WAICO answers concentration by a different route, with open weights. Hugging Face’s spring 2026 analysis found that Chinese models had come to account for “the plurality or 41% of downloads” over the preceding year, which is a substantial achievement and solves a substantial problem. Adaptation cost is real, and open weights lower it. What open weights leave untouched is siting cost, because a model you have downloaded still has to run somewhere, and the somewhere is exactly what the compute map describes. While the policy world spent three years arguing about chips, the chokepoint moved to the grid.

There is a thing about concrete that took me an embarrassing amount of time to learn, out of a construction engineering text I picked up for reasons that had nothing to do with any of this. Concrete does not dry. Drying is what paint does. Concrete cures, which is a chemical reaction. Portland cement and water combining into calcium silicate hydrate, and the reaction does not stop at the twenty-eight-day mark that the specification sheets treat as completion. Twenty-eight days is a convention. Somebody picked it because it is a reasonable interval at which to test compressive strength and issue a certificate, and because construction schedules need a number. The reaction itself continues at a decreasing rate for years. A slab poured in 1974 is measurably stronger in 1994 than it was on the day the inspector signed off. It is still, slowly, becoming more of what it was poured to be.

I mention it here because the twenty-eight-day number is precisely the kind of thing that gets mistaken for an ending, and I will come back to it.

In 1951 a country could ignore the statute, absorb the fiscal hit, and wait for the next appropriations cycle, because the thing being withheld was money, and there was forever a chance of the money returning. A country outside the compute map cannot wait, because the thing being withheld is a building with a substation attached to it, and buildings do not arrive on a fiscal calendar.

A Planet of Subscale Clones

Helberg’s June essay argues that United Nations machinery is pressing toward a world in which every country commands “an irreducible minimum of artificial intelligence,” and that forty governments each building a sovereign stack produces, in his phrase, “subscale clones, each heroically reconstructing last year’s breakthrough while the breakthrough itself moves on without them.” He is describing a real capacity constraint, not a rhetorical one, and I think he is right about it. Forty national data centers running last year’s models at a tenth of the efficiency is not sovereignty.

The objection from outside the coalition converges on the same conclusion from an entirely different direction. Zachary Karabell, writing in Foreign Policy in 2020 and arguing that Cold War analogies are a misuse of history, pointed out that trade between the United States and the Soviet Union was “so negligible as to be all but nonexistent” — which makes drawing a 1947 line through a relationship of extraordinary economic interdependence an act of composition rather than observation. Joseph Nye’s version is structural: the United States is in a three-dimensional game with China, unipolar at the military level, multipolar at the economic level, and chaotically distributed at the transnational level, and “the Cold War metaphor is too closely locked into the traditional two-dimensional chess model.”

The technical version of the objection holds that the chokepoint is eroding under its own weight. Sujai Shivakumar, Charles Wessner and Thomas Howell at CSIS state the boundary condition flatly: export controls are relevant “only so long as the United States and its allies possess chip technologies that China wants and needs.” Jeffrey Ding’s Technology and the Rise of Great Powers (2024) makes the longer argument that a state’s capacity to diffuse general-purpose technology across its economy, rather than its access to the frontier, is what determines long-run economic power. This is a thesis that, if accurate, makes the entire compute-chokepoint strategy a bet on the wrong variable.

The behavioral version holds that states are handling this without anybody’s help. Cheng-Chwee Kuik defines hedging as “insurance-seeking behavior under situations of high uncertainty and high stakes, where a rational state avoids taking sides and pursues opposite measures vis-à-vis competing powers to have a fallback position.” Hunter Marston documents Southeast Asian states doing precisely that – deepening ties with both superpowers rather than siding with one or the other. And Chatham House, examining whether middle powers can build sovereign AI at all, concludes that compute scarcity is the central bottleneck and that American gatekeeping over hardware will limit any such effort.

Kazakhstan is the strategy, not the scandal. Hedging is working. No coalition of the aggrieved is needed, and the 1947 analogy has been imported from a period whose defining condition (mutual economic irrelevance)is simply absent from the present one.

Who Pays

Coercion in an interdependent system does not run along a border; it runs through the network. States with jurisdiction over hub nodes can employ what they call the chokepoint effect (the capacity “to limit or penalize use of hubs by third parties”) and because hubs offer extraordinary efficiency benefits and are extremely difficult to circumvent, actors denied access “can suffer very substantial consequences.” Gregory Allen, who has done more than anyone to document the compute penalty, rejects the erosion thesis outright in its popular version: of the two claims commentators drew from DeepSeek’s efficiency gains (that Chinese labs had caught up, and that huge numbers of cutting-edge American chips are therefore not required ) “the second is wrong.”

The DeepSeek numbers themselves deserve more care than they usually get. Liang Wenfeng told a Chinese interviewer in July 2024 that his company’s gap with the international frontier ran roughly twofold in model structure and training dynamics, and twofold again in data efficiency, compounding to about four times the compute for the same result. He was describing architecture and data, not export controls; the reading that attaches that multiplier to the H800-versus-H100 penalty is Allen’s gloss, and a defensible one, but it is a gloss. Liang’s own statement about the chokepoint sits elsewhere in the same interview and is much flatter: Our challenge has never been money but the embargo on high-end chips. A leaked investor transcript from July 2026 puts the current gap at fourfold plus two years, and has him asking for two hundred thousand Huawei 950 chips and being offered sixteen thousand.

Fourfold plus two years. Sixteen thousand against two hundred thousand.

Mark-Alexandre Doumba, the Republic of Gabon’s Minister of Digital Economy and Innovation, writing at Brookings in February 2026, inverts the risk ordering that Western analysis takes as given. “The greatest risk,” he writes, “is not missing the AI revolution, but joining it too early.” Skip the sequence and the outcome is specific: “Without sequencing, Africa risks becoming the world’s raw data mine — exporting information, importing algorithms, and capturing little of the value. The echo with history is sobering: once raw minerals, now raw data.” Doumba is not reaching for a parallel. He is a sitting minister describing the shipping arrangements his own government already has.

Hedging, it turns out, works for states that hold something to hedge with. Kazakhstan can sign both instruments because it sits on minerals the American coalition cannot expel itself from needing: the more interesting half of Marston’s work is not that Southeast Asian states hedge, but that some of them stop. His cases are Cambodia and Myanmar, where autocratic rulers abandoned hedging for alignment with Beijing; not because the external calculation shifted, but because internal threat made regime security the binding constraint. Hedging survives until the domestic situation makes a patron worth more than a fallback position.

For the Desert there is no fallback position to turn to. Insurance presupposes an alternative, and there is no alternative supplier of a substation.

Eighty-eight endorsements of the New Delhi Declaration at announcement, ninety-one now, with the United States and China both on the list from the first day. Twenty countries and four institutions for the Charter; a list containing neither India nor either superpower, and six of whose twenty members are Pax Silica signatories. Endorsement carrying, by its own terms, no legal, financial, operational, or governance obligations. And on the margins of that same summit, India signed the Pax Silica Declaration along with a bilateral Joint Statement on the U.S.-India AI Opportunity Partnership, witnessed by Ashwini Vaishnaw and Michael Kratsios.

The convening power produced a voluntary charter that it declined to endorse, and signed a binding one that it did. Bandung wrote Principle 6(a) and went home to its alliances. Seventy-one years apart, and the form holds without a dent in it.

Strategic Choices

Cairo in 1964 turned non-alignment away from moral positioning and toward economic and technological demands, and its Programme for Peace and International Co-operation was transmitted to the United Nations and circulated as a document of the General Assembly, A/5763. I went looking for a Belgrade equivalent from 1961 and could not find one in the UN’s catalogue, which is suggestive rather than conclusive; the pre-1965 records are imperfect and I would not build much on an absence.

UNCTAD’s first session ran from March to June 1964, and on 15 June, at its conclusion, seventy-seven developing countries issued a Joint Declaration and became the Group of 77. That line runs forward to General Assembly Resolution 3201, adopted at the sixth special session on 1 May 1974, establishing the New International Economic Order and already demanding, at paragraph 4(p), access to modern science and technology “in forms and in accordance with procedures which are suited to their economies.” It runs on to the Code of Conduct on the Transfer of Technology, negotiated under UNCTAD auspices from 1976 to 1985 and then abandoned.

UNCTAD’s own retrospective is admirably blunt about why: the abandonment “was due to the then continuing disagreement between developing and developed country models of technology transfer regulation,” and in any case “much of this debate has now been overtaken by the orientation of the TRIPS agreement.”

Susan Sell documents what did the overtaking. TRIPS, she writes, “is a ‘can do’ story about twelve men (the members of the IPC) who made IP rules that now bind most of the globe” (the Intellectual Property Committee, an American trade association of pharmaceutical, entertainment and software executives). Twelve, by her count, against the seventy-seven who had signed the Joint Declaration. That arithmetic is mine rather than hers, and the two coalitions were never in the same room in that configuration, so take it as a framing and not a finding. But the direction of the result is not in dispute.

Across 1955, 1974, 1985, and 2026 the pattern is identical and, as far as I can tell, unbroken: the party holding a chokepoint writes an instrument with a termination clause and a waiver, and the party without one writes a declaration of principles and calls it a framework. Nils Gilman, introducing a scholarly dossier on the NIEO’s collapse, supplies the tell — “the failure of the NIEO was the result of a deliberate and concerted strategy on the part of leaders in the north, compounded by strategic choices on the part of the south.”

Strategic choices. The non-binding form is the choice that keeps losing, and it keeps getting chosen anyway, because it is the only form available to a party with nothing to terminate. You cannot write a termination clause into an agreement when you hold nothing the other side needs badly enough to fear losing. What reads as idealism in the drafting is very often an accurate reading of one’s own position.

Which brings me back to WAICO’s Article 17, and to why I left it sitting there.

A withdrawal clause is a generous thing to put in a founding document. Twelve months’ notice, no cause, no penalty – that is better exit language than most commercial tenants ever negotiate. It costs Beijing nothing whatsoever, because WAICO holds no chokepoint. There are no allocations to withhold from a departing member, no substations to leave unbuilt, no concierge service to stop returning calls. The organization coordinates standards and convenes working groups.

Pax Silica’s Declaration has no such clause, and now there is a letter explaining that membership is reviewable. That asymmetry is the whole finding. The document that tells you how to leave is the one that cannot hold you. The document that holds you does not discuss leaving at all. Generosity in the exit language is a reliable indicator of weakness in the instrument, which is a thing every commercial tenant learns eventually, usually late.

The demand that countries choose is old. What is new is that the choice, once made, gets poured and energized and will still be standing when the government that made it has been out of office for a decade. Non-alignment was thinkable in 1961 because alignment was reversible: a treaty can be denounced, an aid package can lapse, an ambassador can be recalled before lunch and be home by Thursday. Power purchase agreements run twenty years. Siting decisions run longer than that. Jurisdictional rules about where model weights may lawfully rest attach to buildings, buildings attach to land, and a concierge service operating out of roughly 270 embassies exists for the specific purpose of making sure the buildings go up in the right countries.

So where does the choice actually live?

Not with the delegation that signs, plainly, because delegations are replaced, generally within a term. Not with the government that ratifies, for the same reason. The slab is not replaced. Twenty-eight days is when the inspector issues the certificate, and the reaction keeps going long after everyone involved has stopped thinking about it. Whatever gets poured this year (in Astana, in Abu Dhabi, wherever the concierge service places its next call) will be stronger in 2046 than on the day the ribbon is cut, and the country hosting it will have had four or five governments by then, none of whom were in the room, all of whom will be living inside a decision made by people who are, by that point, dead, retired, or teaching.

The letter has not been sent, but someone in Washington set it down, then leaked it, instead of sending it. The Battle Act’s waiver clause sat in the United States Code for twenty-eight years after Battle himself had called its administration indefensible, requiring only a determination and a quarterly report to six committees that had, by the end, almost certainly stopped reading them. Nobody repealed it. There was never a reason to. Statutes do not decay; they wait, and the waiting is free, and one morning somebody in an office needs a mechanism and finds one already drafted, already constitutional, already outlived by every Congress that declined to look at it.

I would like to be able to tell you there is a clause on the other side of the ledger that does the same work in the other direction. I have read the Charter twice looking for one.

Just like commercial leases and gym memberships, the clause just doesn’t seem to be there.

References

Allen, G. C. (2025, March 7). DeepSeek, Huawei, export controls, and the future of the U.S.-China AI race. Center for Strategic and International Studies.

Cominform. (1947, October 5). Declaration of the Conference of Representatives of Nine Communist Parties. (Adopted Szklarska Poręba, 27 September 1947.)

Dinkel, J. (2019). The Non-Aligned Movement: Genesis, organization and politics (1927–1992). Brill.

Ding, J. (2024). Technology and the rise of great powers: How diffusion shapes economic competition. Princeton University Press.

Doumba, M.-A. (2026, February 10). Why Africa should sequence, not rush into AI. Brookings Institution.

Farrell, H., & Newman, A. L. (2019). Weaponized interdependence: How global economic networks shape state coercion. International Security, 44(1), 42–79.

Feierabend, L. K. (1996). Politické vzpomínky III. Atlantis.

Gilman, N. (2015). The new international economic order: A reintroduction. Humanity, 6(1), 1–16.

Government of India. (2026, February 19). AI Impact Summit Declaration and Charter for the Democratic Diffusion of AI.

Helberg, J. (2026, June 23). The digital sovereignty trap. U.S. Department of State.

Hugging Face. (2026, March 17). State of open source on Hugging Face: Spring 2026.

Jackson, I. (2001). The economic Cold War: America, Britain and East-West trade, 1948–63. Palgrave Macmillan.

Jansen, G. H. (1966). Afro-Asia and non-alignment. Faber and Faber. (Quoted in Vitalis, 2013, p. 261.)

Karabell, Z. (2020, August 14). There’s no Cold War with China. Foreign Policy.

Kuik, C.-C. (2021). Getting hedging right: A small-state perspective. China International Strategy Review, 3(2), 300–315.

Lehdonvirta, V., Wú, B., & Hawkins, Z. (2024). Compute North vs. Compute South: The uneven possibilities of compute-based AI governance around the globe. In Proceedings of the Seventh AAAI/ACM Conference on AI, Ethics, and Society (pp. 828–838). AAAI Press.

Lukes, I. (2012). On the edge of the Cold War: American diplomats and spies in postwar Prague. Oxford University Press.

Lüthi, L. M. (2014). The non-aligned: Apart from and still within the Cold War. In N. Mišković, H. Fischer-Tiné, & N. Boškovska (Eds.), The Non-Aligned Movement and the Cold War: Delhi–Bandung–Belgrade (pp. 97–113). Routledge.

Marshall, G. C. (1947, June 5). Remarks at Harvard University alumni exercises. George C. Marshall Foundation.

Marston, H. S. (2023). Abandoning hedging: Reconsidering Southeast Asian alignment choices. Contemporary Southeast Asia, 45(1), 55–81.

Martina, M. (2026, August 14). Exclusive: US to tell partners they must pick sides in AI race with China. Reuters.

Mastanduno, M. (1992). Economic containment: CoCom and the politics of East-West trade. Cornell University Press.

Mutual Defense Assistance Control Act of 1951, Pub. L. No. 82-213, 65 Stat. 644.

Nye, J. S., Jr. (2022). How not to deal with a rising China: A US perspective. International Affairs, 98(5), 1635–1651.

Parrish, S. D., & Narinsky, M. M. (1994). New evidence on the Soviet rejection of the Marshall Plan, 1947: Two reports (CWIHP Working Paper No. 9). Woodrow Wilson International Center for Scholars.

Record of I. V. Stalin’s conversation with the Czechoslovak government delegation. (1947, July 9). Wilson Center Digital Archive.

Sell, S. K. (2003). Private power, public law: The globalization of intellectual property rights. Cambridge University Press.

Shivakumar, S., Wessner, C., & Howell, T. (2025, April 14). The limits of chip export controls in meeting the China challenge. Center for Strategic and International Studies.

Truman, H. S. (1947, March 12). Special message to the Congress on Greece and Turkey: The Truman Doctrine. Harry S. Truman Library.

United Nations Conference on Trade and Development. (2001). Transfer of technology (UNCTAD/ITE/IIT/28), pp. 21–22.

United Nations General Assembly. (1974, May 1). Declaration on the establishment of a new international economic order (A/RES/3201 (S-VI)).

U.S. Department of State. (1951, December 19). Aide-mémoire to the Embassy of Thailand. In Foreign relations of the United States, 1951, Asia and the Pacific, Vol. VI, Part 2 (Doc. 101, pp. 1644–1647).

U.S. Department of State. (1952, September 29). Representative Laurie C. Battle to the Director of Mutual Security (Harriman). In Foreign relations of the United States, 1952–1954, General: Economic and political matters, Vol. I, Part 2 (Doc. 26, pp. 896–900).

U.S. Department of State, Office of the Historian. (n.d.). Bandung Conference (Asian-African Conference), 1955. Milestones in the History of U.S. Foreign Relations, 1953–1960.

U.S. Department of State. (2026, February 20). Joint statement on the U.S.-India AI Opportunity Partnership.

U.S. Department of State. (2026). Pax Silica.

Varela Sandoval, F. J., Wilkinson, I., Krasodomski, A., & Wilkinson, R. (2026). How middle powers can weather US and Chinese AI dominance: The case for ‘sovereign AI’ strategies. Chatham House.

Vitalis, R. (2013). The midnight ride of Kwame Nkrumah and other fables of Bandung (Ban-doong). Humanity, 4(2), 261–288.

World Artificial Intelligence Cooperation Organization. (2026, July 16). Agreement on the establishment of the World Artificial Intelligence Cooperation Organization. Shanghai.

Zhdanov, A. A. (1947). The international situation: Report delivered at the Informatory Conference of Representatives of a Number of Communist Parties. Foreign Languages Publishing House.