University of Oxford
CTS Insights

How Superintelligence Could Reshape Global Power

If artificial superintelligence arrives, the states closest to it may not wait until they can trust it. A coalition of middle powers cannot govern that race on its own: any credible regime has to start with Washington and Beijing.

S. Yash Kalash

Senior Fellow · · 11 min read

If Artificial Superintelligence (ASI) is achieved, it will change the balance of power. A system that is superior to human experts in almost every cognitive task could dramatically accelerate the rate of scientific research, military planning, and the design of its own successors. A state that is fearful of a rival about to gain such an advantage and wants to maintain its global hegemony may well go ahead and deploy its own systems before it is willing to trust them. Any credible restraint from a coalition of middle and smaller powers is going to face strategic and tactical challenges, and thus any governance regime is most likely going to have to start with the states that can shift the balance and impose costs on each other.

Why a stability board is not enough

In September 2026, Canadian Prime Minister Mark Carney floated the idea of a Global Technology Stability Board in the model of the Financial Stability Board (FSB). The idea would be for a coalition of willing governments to bring together the companies and the experts to develop common assessments and learn from the incidents1. This idea would most likely be a separate deal from the one that Washington and Beijing would be pursuing on major risks bilaterally2. The need for an alternate deal is clearly evident, but any technology stability board in itself will not create the political deal that would make governance in any way consequential. A government would still have to make up its mind as to whether it wanted to disclose a capability that might give it an edge over a rival and whether it wanted to act on the findings that might in fact slow down its own programme.

The capabilities one government would like another to be restrained from having may well be precisely the capabilities the rival government believes it needs to remain secure.

The FSB coordinates non-binding standards which domestic authorities implement3. When Credit Suisse went under in 2023, Swiss authorities stepped in to back the UBS takeover with public liquidity and guarantees4. Switzerland had the powers and the means to do so, and the rationale for rebuilding confidence was shared by all involved, but in a race for ASI the capabilities one government would like another to be restrained from having may well be precisely the capabilities the rival government believes it needs to remain secure.

The international system has often struggled with gridlock. History proves that a single superpower can easily paralyse international initiatives. For example, Russia vetoed a UNSC draft resolution on its invasion of Ukraine in 20225. The WTO Appellate Body was effectively rendered unable to hear appeals in 2019 following the US blocking appointments, and many other members established an interim arrangement for disputes whilst the gridlock continued6. Ultimately, a technology board of this kind could risk devolving into another powerless coalition, merely regulating those willingly ready to comply whilst the most critical global actors retain complete freedom of action.

Precedents and recent moves

  1. 1963The US and the Soviet Union set up a direct communications link to reduce the risk of miscalculation in a crisis
  2. 2019The WTO Appellate Body can no longer hear appeals after the US blocks appointments
  3. 2022Russia vetoes a UN Security Council draft resolution on its invasion of Ukraine
  4. 2023Swiss authorities back UBS’s takeover of Credit Suisse with public liquidity and guarantees
  5. Feb 2026New START, the last bilateral arms control treaty between the US and Russia, expires
  6. Sep 2026Mark Carney proposes a Global Technology Stability Board; the US and China agree to set up an AI safety channel

Carney's broader argument for middle powers is that they can gain leverage through co-operation on trade and investment.7 But this argument has structural limits; pooling influence works in situations where others need their markets or the inputs they control. Canada's own AI strategy makes it clear that its AI cloud and data centre offerings are largely foreign-owned and foreign-controlled8. While pooling resources can certainly improve the terms on which public agencies buy AI resources, it cannot be used to force a rival to reveal its military capabilities or to delay a release that its own leaders would consider strategically important. A government that is prepared to sacrifice sales abroad in order to protect a military advantage can live with a board's disapproval.

Why Washington and Beijing are essential

The distribution of technical power makes Washington and Beijing essential. Industry was responsible for over 90% of the significant models that Stanford's team looked at in 2025, and the US was the leader in national model development, with China being its main rival9. The decisions that Washington makes on the licensing of advanced chip exports to China will have a direct bearing on the terms of competition10. A nation which is dependent on foreign models for its defence or for its critical services could be in a position where it has to consider whether a diplomatic decision will put access at risk. A number of states could end up with ASI and still be in competition as to who can turn research into industrial or military capability the fastest. Thus, the prospect of an advantage, even a fleeting one, would make the cost of waiting for evidence that a rival's systems are safe exceedingly high.

Frontier AI companies have geopolitical leverage of their own. A developer that gives a military customer early access to a more capable model could prompt a rival government to accelerate its own deployment. Cloud providers can shape which countries have enough computing power to build advanced models by deciding where to invest and whom to serve. An agreement among the major powers should therefore involve these firms in designing the procedures they would have to carry out, subject them to binding obligations, and give independent evaluators confidential access to the evidence they need to validate compliance.

Who else must be at the table

Washington and Beijing will have to negotiate the first set of constraints on frontier development and deployment. India must be part of the wider security bargain from the start. Its demographic weight, economic size, nuclear forces, and military decisions are relevant to crisis stability in Asia. Russia, though not a superpower in the traditional sense, may well be required in the case of any terms relating to nuclear command and escalation. Any state that controls a critical part of the chip supply chain will have to be part of any commitments that are predicated on that input. The mix of states required for each of these agreements will of course evolve as capabilities spread.

Who needs to be at the table

All other governments

Public terms; the right to report incidents, request investigations and receive warnings. Middle powers help build independent assessments.

Frontier AI developers and cloud providers

Help design the procedures, accept binding obligations and give evaluators confidential access to evidence.

The wider security bargain

India from the start; Russia on nuclear command and escalation; states that control critical parts of the chip supply chain.

Washington and Beijing

Negotiate the first constraints on frontier development and deployment.

What a US–China deal could include

Washington and Beijing have already agreed to establish a dialogue and a bilateral channel for AI incidents11. Historical precedent exists for cooperation between rival powers even during times of severe competition. In 1963, the US and the USSR established a direct communications link to reduce the risk of miscalculation in a crisis12. Thus, both sides could see a rationale for doing so on AI even as their rivalry continues.

What a US–China deal could include

  • Humans in the loopOn essential military decisions such as nuclear launches, with independent checks on machine-generated warnings.
  • Hands off command systemsNo deliberate AI-enabled interference with nuclear warning and command systems or other critical infrastructure in peacetime.
  • Incident reportingIncidents defined by what a system does, with duties to notify, preserve records, report findings and isolate a deployment.
  • Release triggersAgreed test results that pause a release or an expansion of permissions until independent testing confirms the risk is contained.
  • Training-run noticeAdvance notice of very large training runs at declared facilities, with agreed inspectors examining selected records.
  • A joint crisis drillA stress test of a major AI incident of unknown origin, run now through the new US–China dialogue.

For example, Washington and Beijing could keep humans in the loop in essential military processes, such as authorising nuclear launches, and require independent corroboration of the machine-generated warnings that would be used in those decisions. They could also agree to refrain from any deliberate AI-enabled interference with each other's nuclear warning and command systems or any other critical infrastructure in peacetime. India, Russia, and the other nuclear powers would all have to negotiate their own equivalent commitments where their decisions would have an impact on the risk.

The two governments could define reportable incidents by what a system does. So, if there is credible evidence that a hosted model is operating outside of its authorised parameters and is in some way threatening critical infrastructure of another country, the host government would be required to notify its counterpart, preserve records and report initial findings by an agreed deadline. If the evidence meets a threshold, the host authority would be required to temporarily isolate the deployment in question and to provide its counterpart with a protected account of what it did. Both governments would require domestic rules which bind the developers and cloud operators.

Some interventions may need to happen earlier. Once a model's weights have been widely copied, then pulling the source off the net leaves the copies out there. A system can also bring in new risks when it gets access to networks or tools that can run code13. A developer may find themselves in a position where they cannot get back control after a serious failure. There should be agreement on what test results would trigger a temporary suspension of release or an expansion of permissions. A domestic regulator should be able to lift the suspension only after independent testing has confirmed that the measures in place will be sufficient to contain the risk which originally triggered the suspension.

Each government could require advance notice of very large training runs at declared facilities and allow mutually agreed inspectors to examine selected records under agreed constraints. Large computing installations provide a physical point of observation, though distributed training can make that more difficult14. Computing thresholds can trigger a review, but the assessment of behaviour is still the key determinant of whether or not a system is subject to intervention. Inspectors could be asked to verify that an assessment had in fact taken place or that an operator had in fact curtailed a deployment. Each party would be able to challenge a specific declaration and request evidence under terms which would safeguard proprietary and classified information. Where access to sensitive computing resources is a negotiated benefit, the licensing authority could of course suspend that benefit following a verified breach, and importantly emergency communications should remain open.

Making a deal last

International treaties are only as durable as the strategic self-interest and trust that underpins them.

International treaties are only as durable as the strategic self-interest and trust that underpins them. The New START (Strategic Arms Reduction Treaty) between the US and Russia officially expired on February 5, 202615. This was the last existing bilateral arms control arrangement between the two nations. The collapse of this treaty is a clear reminder that in international relations there are no binding guarantees. A country which believes that restraint will cost it the decisive lead may be in a position to reject even a very narrow agreement. The negotiation process should start with the risks each side faces irrespective of who is in charge, for example, an incident which could be misinterpreted as a military attack. A deal between great powers could itself serve to reinforce the hierarchy if it were used as a basis to say who else is entitled to access advanced AI. The terms of such a deal should be as public as possible, and governments outside of the deal should be able to report incidents, request an investigation, and receive warnings about risks to their own infrastructure. This is where middle powers can help in the development of independent assessments and can serve as a force for pushing for those kinds of protections.

Stress-testing the worst case

Rather than waiting for a crisis to strike, Washington and Beijing should use their new dialogue to stress-test a worst-case scenario, for example, a major AI incident of completely unknown origin where a renegade developer refuses to cooperate. Running this exact drill would force the two superpowers to confront a terrifying reality check: exactly how many hours does it take to diagnose the anomaly, shut down their own operators, and deliver enough verified data to the other side to avert a catastrophic military escalation? Exposing these operational blind spots now is critical. It is infinitely easier to patch the gaps in global crisis capabilities today, while the lines of communication are open, than in the heat of a future standoff when neither side is willing to blink first for fear of losing their technological lead.

Notes

  1. bloomberg.com/news/articles/2026-09-14/carney-calls-for-global-ai-oversight-body-to-ensure-safe-development ↩
  2. cigionline.org/articles/the-world-needs-an-ai-stability-board ↩
  3. fsb.org/work-of-the-fsb ↩
  4. fsb.org/uploads/P101023.pdf ↩
  5. research.un.org/en/docs/sc/quick/meetings/2022 ↩
  6. wto.org/english/thewto_e/minist_e/mc12_e/briefing_notes_e/bfwtoreform_e.htm ↩
  7. pm.gc.ca/en/news/speeches/2026/01/20/principled-and-pragmatic-canadas-path-prime-minister-carney-addresses ↩
  8. ised-isde.canada.ca/site/ised/en/canadas-national-artificial-intelligence-strategy-ai-all ↩
  9. hai.stanford.edu/ai-index/2026-ai-index-report/research-and-development ↩
  10. bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china ↩
  11. pbs.org/newshour/world/china-and-u-s-agree-to-establish-ai-safety-channel-and-continue-trade-and-military-talks ↩
  12. history.state.gov/historicaldocuments/frus1961-63v05 ↩
  13. internationalaisafetyreport.org/publication/international-ai-safety-report-2026 ↩
  14. arxiv.org/html/2402.08797v1 ↩
  15. icanw.org/new_start_expiration ↩

← All CTS Insights

A centre of the