Compute, chips and the responsibility gap: artificial intelligence as geopolitics

The Convention on Certain Conventional Weapons closed its Group of Governmental Experts on lethal autonomous weapons systems in the early hours of 5 September 2026, in Geneva, after three years of work. Seventy-six states have now said they are ready to negotiate a treaty, seventy of them parties to the Convention itself, which is a majority of its membership. What they were handed was a final report that Human Rights Watch, reporting on 7 September, called significantly watered down: it keeps a characterisation of autonomous weapons and an affirmation that human control and judgement are required for compliance with international law, and loses the references to design, to development, to ethical considerations, and to technical requirements such as predictability and reliability. Russia and the United States pressed for the deletions.

That is the state of the law. The state of the power is elsewhere — in fabrication plants in Hsinchu, in substations in northern Virginia, in refineries in Guangxi that most readers of this newspaper could not name and could not do without. The reason the Geneva process keeps failing is legible in that industrial geography rather than in the diplomacy.

Thomas Hobbes and Immanuel Kant are the two figures worth thinking with here, and not because they can be waved at. Hobbes has a precise account of why sovereigns who fear each other arm anyway, and an equally precise account of why they can survive doing so. Kant has a proposal for what states may bind themselves to when no one can compel them, and a firm refusal of the thing he is usually credited with proposing.

The competition is for inputs, and the inputs are physical#

The phrase "AI race" invites a picture of laboratories, and the picture is misleading. What states have spent four years fighting over is the material base: the machines that etch transistors, the electricity that moves through them, and the elements that make the wafers work at all.

Start with power. The International Energy Agency's Energy and AI, published in April 2025, put global data-centre electricity consumption at around 415 terawatt-hours in 2024, or roughly 1.5 per cent of world consumption, and projected that it would more than double to about 945 TWh by 2030. The agency named AI as the most important single driver of that growth while declining, sensibly, to put a precise share on it. Nine hundred and forty-five terawatt-hours has to be generated, transmitted and paid for by somebody, in a particular jurisdiction, under a particular regulator, which is why the AI question has become an energy-policy question in half the OECD.

Then fabrication. The figure one hears repeated at conferences is that Taiwan Semiconductor Manufacturing Company makes over ninety per cent of the world's advanced chips, and it is worth pausing on it, because it is the sort of claim that hardens into background knowledge without being checked. Scotten Jones of TechInsights took it apart in March 2024 and called it ill-defined and grossly inaccurate. On a consistent definition of advanced logic — seven nanometres and below for TSMC, with the roughly equivalent Samsung and Intel nodes counted in — TSMC holds about 64 per cent of installed capacity worldwide. The ninety per cent figure survives only if "advanced" is narrowed to three-nanometre logic, where TSMC is for the moment the only genuine producer. Sixty-four per cent remains an extraordinary concentration of a strategic capability in one company on one island, and the difference between the two numbers matters when officials are pricing risk into a policy.

Minerals are where the leverage runs the other way. China supplies about 98 per cent of the world's primary gallium and roughly 68 per cent of its germanium, according to a CSIS analysis published on 17 July 2025 by Aidan Powers-Riggs and colleagues. Beijing began requiring export licences for both in July 2023; in December 2024 it moved to what amounted to an embargo on gallium exports to the United States, asserting extraterritorial reach over transshipment; on 2 January 2025 it added the extraction technologies themselves to the control list, down to the specific ion-exchange methods for pulling gallium out of alumina mother liquor. The Rotterdam price reached 687 dollars a kilogram in May 2025, more than 150 per cent above pre-control levels. Gallium is not an exotic input. It is in the radio-frequency and power electronics that sit around the compute, and in a good deal of defence hardware besides. The supply chain has choke points, and different powers hold them at different points along its length.

Hobbes's gladiators now have a supply chain#

Hobbes's remark about sovereigns in Leviathan (1651) is quoted so often that its second half has almost disappeared. Chapter XIII has kings and persons of sovereign authority "in continual jealousies, and in the state and posture of gladiators; having their weapons pointing, and their eyes fixed on one another; that is, their forts, garrisons, and guns upon the frontiers of their kingdoms; and continual spies upon their neighbours; which is a posture of war." The sentence that follows is the one that does the work:

But because they uphold thereby the industry of their subjects; there does not follow from it that misery which accompanies the liberty of particular men.

Hobbes is saying that the international state of nature is bearable in a way the interpersonal one is not, because armed sovereigns, unlike armed individuals, protect the productive life going on behind them. He therefore does not argue for a world sovereign; he argues that states can live in the posture of war indefinitely without the catastrophe that follows when persons do. Anyone who invokes Hobbes to demand a global AI authority has him precisely backwards.

What Hobbes supplies is the mechanism, which later international-relations theory named the security dilemma. Robert Jervis set out the canonical version in "Cooperation Under the Security Dilemma" (World Politics, vol. 30, no. 2, January 1978, pp. 167–214), specifying when the dilemma bites hardest: it eases when defensive capability is stronger than offensive, and when the two can be told apart from the outside.

Applied to AI, both conditions fail, and they fail for reasons intrinsic to the technology rather than to anyone's bad faith. A frontier model trained for translation and code is the same artefact whether it ends up summarising cables for a foreign ministry or generating spear-phishing at scale; there is no offensive variant with a different silhouette. And capability is not observable. A missile silo can be counted from orbit; a training run cannot, and the weights that result can be copied in an afternoon. Cold War arms control rested on the fact that the objects being limited were large, immobile and photographable. Nothing about a data centre tells an adversary what was trained in it.

A third asymmetry the nuclear analogy obscures: nuclear capability was built by states for states, while the compute that matters is built by companies, sold on commercial terms and rented by the hour. Hobbes's sovereign is defined by holding the means of violence. The present sovereign rents them.

Export controls: the strongest case, and the strongest reply#

The policy that follows from the Hobbesian reading is denial, and the United States has been pursuing it since Jake Sullivan, then national security adviser, spoke at the Special Competitive Studies Project summit on 16 September 2022. Sullivan announced the end of a doctrine: "we have to revisit the longstanding premise of maintaining 'relative' advantages over competitors in certain key technologies", abandoning "a 'sliding scale' approach that said we need to stay only a couple of generations ahead". Instead: "Given the foundational nature of certain technologies, such as advanced logic and memory chips, we must maintain as large of a lead as possible." Three weeks later the Bureau of Industry and Security issued the rule of 7 October 2022, and the architecture has been built and rebuilt ever since.

The strongest intellectual case for it is not economic but anticipatory. Dario Amodei made it in January 2025 in his essay "On DeepSeek and Export Controls", arguing that the world of 2026–27 will be either unipolar or bipolar in frontier compute, that in the unipolar case a temporary lead "could be parlayed into a durable advantage", and that in the bipolar case China "could direct more talent, capital, and focus to military applications". His conclusion is blunt: well-enforced export controls are "the most important determinant of whether we end up in a unipolar or bipolar world". The empirical case is made by Chris McGuire of the Council on Foreign Relations, writing on 15 December 2025: American chips are roughly five times more powerful than Huawei's best, a gap he projects widening to seventeen times by 2027, and Huawei's output amounts to about 5 per cent of Nvidia's aggregate AI compute in 2025, falling to 2 per cent by 2027, constrained above all by high-bandwidth memory after the December 2024 controls. "Huawei is not a threat that justifies loosening controls," McGuire writes; "it is evidence that the controls are working."

The reply is serious. Jensen Huang said flatly on 21 May 2025 that "the export control was a failure", on the reasoning that denial handed Chinese firms "the spirit, the energy and the government support to accelerate their development". One discounts a chief executive's verdict on a policy that costs him a market, but the structural version of his point is made by analysts with no such interest. Ansgar Baums, writing for the Stimson Center's Red Cell project, argues that "choke point" lacks a definition firm enough to bear policy weight, that restricting access to a technology "first and foremost incentivizes the replacement of it", and that the US Chamber of Commerce has costed total loss of the Chinese market at 83 billion dollars a year and 124,000 jobs.

The more damaging objection is one neither side quite presses: the regime has not been stable enough to function as denial. The Framework for Artificial Intelligence Diffusion was published in the Federal Register on 15 January 2025, establishing a three-tier world; it was rescinded on 13 May 2025. In November 2025 the Commerce Department authorised up to 35,000 Nvidia Blackwell GB300 processors each for G42 in the United Arab Emirates and HUMAIN in Saudi Arabia. On 13 January 2026 BIS moved the H200 and AMD's MI325X from presumptive denial to case-by-case review, on conditions including third-party testing in the United States. A fence moved four times in three years teaches adversaries and allies the same lesson, which is to plan around it.

Kant proposed a federation, not a world state, and the difference is the whole point#

Toward Perpetual Peace (1795) is routinely cited as the ancestor of world government, and Kant says the opposite. The Second Definitive Article holds that the law of nations shall be founded on a federation of free states — a foedus pacificum, distinguished both from a mere treaty ending one war and, emphatically, from an international state. It is, in his phrase, "a league of nations, but it couldn't be an international state, a state consisting of nations". Sovereignty is retained. The Third Preliminary Article demands that standing armies be abolished in time, precisely because they constitute a continuing threat to others and drive the spiral of preparation that makes peace "more of a burden than a short war".

Read as an analytical tool rather than an aspiration, this yields a sharper question than "should there be a global AI authority". Kant's question is what sovereigns can bind themselves not to do while remaining sovereign, and his answer concerns capabilities whose existence is itself destabilising. The Third Preliminary Article, applied to compute, is the interesting one: not a demand to disarm, but a demand to stop holding a permanent posture whose only justification is what others might hold.

The empirical test of the Kantian hypothesis has been running for twelve months, and it is producing federations rather than a federation. The UN General Assembly launched the Global Dialogue on AI Governance on 25 September 2025, together with an Independent International Scientific Panel on AI, open to all member states, with the first full convening held in Geneva in July 2026. Laura Caroli and Matt Mande of CSIS read the launch as a map of divergence: the United States and China both seek AI leadership, they wrote in October 2025, but "the United States favoring de facto dominance through diffusion of its technology and China opting to shape AI's trajectory through global institutions".

Then, on 16 July 2026, at the World Artificial Intelligence Conference in Shanghai, twenty-nine states signed the agreement establishing the World Artificial Intelligence Cooperation Organization — an intergovernmental body headquartered in Shanghai and sitting outside the United Nations system altogether, with Russia, Kazakhstan, Pakistan, Indonesia, Brazil, Venezuela, Ethiopia and Cameroon among the founders. Guterres attended. Neither the United States nor the European Union is a member. Caixin, reporting on 17 July, described it as underwriting Beijing's ambition to set global AI standards among developing countries; the proposal had come from Premier Li Qiang at the previous year's conference.

What is emerging, then, is Kantian in form and not in substance: voluntary associations of states that retain full sovereignty, but organised around rival patrons rather than around a shared commitment to republican constitutions, which for Kant was the First Definitive Article and the condition of the rest. A federation of the unlike is a bloc.

The awkward party at the table is not a state#

Michael Horowitz, of the University of Pennsylvania and formerly of the US Department of Defense, published Artificial Intelligence, the Future of War and International Politics with the Centre for International Governance Innovation on 19 February 2026. Its executive summary contains the sentence that most complicates both classical frameworks: "It is companies, not governments, that are developing the bleeding edge of AI capabilities." Horowitz argues that geopolitical tension is undermining military-domain cooperation precisely when it is most needed, and proposes an international agreement on autonomous weapons systems, a rule that humans retain the decision to use nuclear weapons, and a new UN-based institution for rules of the road. The second already has precedent: American and Chinese leaders have affirmed the need to keep nuclear-use decisions in human hands.

The same problem shows up from the disarmament side. Yasmin Afina's UNIDIR study The Global Prism of Military AI Governance, published in Geneva in February 2026 from the previous year's regional consultations, reports real progress on national frameworks and then identifies the binding constraint: "the absence of an authoritative framework for applying responsible AI principles and international law considerations within the private sector".

This is where both poles strain. Hobbes's commonwealth is constituted by the monopolisation of force; a sovereign that leases its decisive capability from a listed company is a sovereign Hobbes did not describe. Kant's federation binds states, and states cannot bind what they do not control. Security Council Report reached a parallel verdict on the institution supposed to hold the ring: in its research report of 27 March 2026, Is the Security Council Wired for Purpose?, it found Council engagement with these technologies "limited, fragmented, and politically contested", even as their consequences run through counter-terrorism, protection of civilians, sanctions enforcement and peacekeeping alike. The longer argument about whether arms-control precedent transfers at all is in our earlier piece on the nuclear model and an AI treaty, and the jurisprudential version in Scott Shapiro on law as code.

What the law actually says about a killing with no author#

The philosophical term of art is older than the current debate. Andreas Matthias coined "the responsibility gap" in Ethics and Information Technology (vol. 6, no. 3, 2004), arguing that when a machine's behaviour is determined by what it learned rather than by what its makers specified, the traditional conditions for holding the manufacturer or the operator responsible cease to be satisfied, and nobody is left who meets them.

International humanitarian law has no gap in this sense; it has three doctrines that presuppose a human. Individual criminal responsibility requires intent or knowledge. Command responsibility, codified in Article 28 of the Rome Statute, attaches where a commander knew or should have known and failed to prevent or punish. State responsibility attaches to conduct attributable to the state. The doctrines are intact; what autonomy threatens is their application. If a commander could not reasonably have predicted the system's behaviour in the circumstances of use, the mental element fails, and with it the case.

The ICRC's position paper Autonomous Weapon Systems and International Humanitarian Law: Selected Issues, issued in October 2025 and updated that December, insists that humans making targeting determinations "remain accountable for these determinations, which cannot be delegated to machine processes". It is candid, in its introduction, that the accountability question itself lies outside its scope — that the ability to hold users of such systems to account for violations is a separate matter requiring separate treatment. Coming from the custodian of the Geneva Conventions, that is the closest thing in the official literature to an admission that the gap is real.

The negotiating record shows how the gap becomes a drafting problem. The Group of Governmental Experts met from 2 to 6 March 2026 and again from 31 August to 4 September, working from a rolling text circulated by the chair the previous December. WILPF's CCW Report, in Laura Varella's despatch of 11 March, records that the United States delegation objected to the term "human control" and proposed "good faith human judgement and care" in its place, a formulation the great majority of delegations rejected as too weak to protect civilians. Support for opening negotiations rose during that week from about forty states to over seventy.

The General Assembly adopted resolution 80/57 on 1 December 2025, which keeps the file inside the Convention, directs High Contracting Parties to complete "the set of elements for an instrument being developed" under the existing mandate, and retains the item for the eighty-first session. It prohibits nothing. On 25 August 2026, Guterres and Mirjana Spoljaric Egger, president of the ICRC, renewed the joint call they first made in 2023 — whose original deadline, a legally binding instrument by 2026, has now expired unmet. "We are now dangerously close to crossing a moral red line," they said: "the autonomous targeting of humans by machines." They ask for prohibitions on unpredictable systems and on anti-personnel autonomous weapons.

Not everyone accepts the premise. Kenneth Anderson and Matthew Waxman argued in 2013, in Law and Ethics for Autonomous Weapon Systems: Why a Ban Won't Work and How the Laws of War Can, that a prohibition trades speculative risks for the real if less visible risk of forgoing automation that would make force more precise and less harmful to civilians nearby, and that the better route is the gradual evolution of codes of conduct from existing principles. Thirteen years on, their prediction about the difficulty of a ban has held up considerably better than their prediction about what the alternative would deliver.

The Kantian formula about treating humanity always as an end and never merely as a means is usually brought in here as decoration. It does better work stated exactly. The objection to autonomous targeting is not that machines will err, which they will, though probably no more often than soldiers have in every war on record. It is that the laws of armed conflict are a system of answerability: they exist so that a person can be required to give an account of a killing to a tribunal, and a person who could not have known cannot give one. A death nobody can be called to answer for is a death outside the law, and the law was the point. Compare the argument about Kantian universality in the categorical imperative and the machine-philosophical reading in Yuk Hui's Kantian machine.

The CCW Review Conference meets in November. Seventy-six states want a mandate to negotiate; the Convention works by consensus, which means two do not have to. Hobbes would have found that unsurprising and not especially alarming, since the industry of subjects continues. Kant, who thought a federation of free states could form only among republics that had already accepted constraint at home, would have wanted to know what kind of states these are before predicting anything at all.