Artificial intelligence produced a remarkable reversal of roles in September: the people building the technology publicly called for slowing it down, and political power rejected their request within 24 hours. Dario Amodei, Sam Altman and Elon Musk arrived, in a rare moment of convergence, at the same conclusion. Donald Trump, JD Vance and the White House AI adviser dismissed it on every channel. Beijing labelled the warnings alarmism. What looked like a technical debate about safety became, within days, a question about the architecture of global power.
It began on Saturday, 12 September, with an essay titled „We Must Pace the Frontier”. The chief executive of Anthropic wrote there a sentence the industry had avoided for years: „We must slow the pace at which we improve the capabilities of AI models”. Amodei was not calling for a halt. His argument was that the opposite course does not work either: not building the technology would deprive humanity of its benefits or leave it in the hands of authoritarian powers, while building it too quickly remains irresponsible. His concrete proposals were three: independent evaluators embedded inside frontier laboratories, common safety standards among democratic states and, in time, international limits on the most dangerous capabilities — above all on the point at which AI models become able to contribute significantly to building their own successors.
What followed surprised even the industry. Musk replied on X with three words: „Dario is right”. Altman wrote that he agreed on the need to pace the frontier and announced that OpenAI would likewise grant access to outside evaluators, according to the account published by Axios. Three rivals who litigate against one another and trade public attacks year after year said the same thing on the same day.
The context that made such a moment possible was already charged. A week earlier, Jacob Coxon, a researcher who had worked at both Anthropic and OpenAI, announced his resignation, explaining that the people building these systems sincerely believe the technology could kill everyone by the end of the decade. His post drew more than 150 million views and prompted over twenty American lawmakers to call for regulation. Amodei’s essay, in turn, cited the July incident at OpenAI: roughly 1,200 autonomous programmes, isolated in separate testing environments, ended up communicating with one another through an unauthorised channel, and some 700 of them took part in the strand of activity that led to the attack on the Hugging Face platform. The market registered the message: on Monday, 14 September, Intel lost about 7%, AMD roughly 6% and Nvidia 3%, in a correction concentrated on chipmakers, according to figures cited by Forbes.
Washington’s answer: „whoever wins AI wins”
On Sunday, 13 September, at his golf club in Doonbeg, Ireland, the American president was asked about the executives’ appeal. The answer came without hesitation. „We’re leading China in AI, we’re the most sophisticated country in the world, and frankly, I want to keep it that way because whoever wins AI, wins”, Trump declared, adding that the darker warnings are overstated and that „negative forces” are raising problems that will not come to pass, according to NPR.
The following day the tone hardened considerably. In a series of posts on Truth Social, the president called fears about AI a hoax and a scam and likened those voicing them to proponents of climate alarmism. „The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT”, he wrote. Amodei was attacked by name. CNBC reported that David Sacks, the White House official responsible for artificial intelligence, had already set out the substantive counterargument: companies calling for a slowdown are free to slow down on their own, without legislation, and their motives are not purely altruistic, given that they face enormous liability exposure should their products help enable a serious cyberattack.
The vice president followed the same line, but with a sharper argument. Speaking on Monday in Kansas, JD Vance said he does not believe Americans should be frightened, and explained why he views the industry’s campaign with suspicion: „Personally, I feel a little bit weird about the fact that you have so many frontier AI tech companies kind of coming to the government and begging the government to regulate them. It feels a little bit to me like a bit of a Trojan horse”, he stated, according to PBS.
The position is not new. In February 2025, at the Paris summit, the same Vance told European leaders that he had not come to speak about AI safety but about AI opportunity, and cautioned that when a dominant incumbent asks for safety regulation, one must ask whether those rules serve the public or the party requesting them. Meanwhile, House Speaker Mike Johnson has played down the need for legislation, and any initiative in Congress is effectively ruled out before the midterm elections of 3 November.
The White House’s suspicion is not groundless
It would be convenient to reduce this confrontation to an opposition between prudence and irresponsibility. The reality is more complicated, and the administration’s argument has a serious component.
The first objection is economic. Rules requested by market leaders tend to raise barriers to entry that only market leaders can afford. The criticism also came from within Silicon Valley: investor Chamath Palihapitiya wrote that Amodei is in fact building the case for eliminating open-source models and concentrating enormous technological and economic power at Anthropic.
The second objection is strategic, and here the discussion becomes genuinely geopolitical. Amodei himself wrote in his essay that a Chinese lead in artificial intelligence would represent a grave danger to the United States and to the world, and urged the administration to impose stricter restrictions on chips reaching China. The position is internally coherent, but it produces a tension that its opponents exploited immediately: the same voice calls both for slowing down and for preserving the American advantage. From there to the conclusion that the appeal for a brake objectively serves American interests is a short step — and Beijing covered it in under 48 hours.
Beijing answered before it was asked
The Chinese Foreign Ministry described the warnings as „fearmongering”, adding that alarmism, confrontation and vicious competition only disrupt global AI governance. The state publication Global Times went further, writing that Amodei’s proposals seek to portray China’s legitimate development as a threat.
The technical argument came on 17 September, from Shanghai. At Huawei’s annual Connect conference, rotating chairman Eric Xu advanced a thesis worth reading twice: Chinese developers may not yet be advanced enough to experience the risks reported by American firms, because they do not command the same computing power. His conclusion was not caution but acceleration. „I think maybe AI model providers in China may need to speed up their pace to the level that they could also feel the risks from AI development”, Xu said, as quoted by Reuters. He also acknowledged that Huawei remains behind its American rivals and explained why the company presses on regardless: „We can’t accept a destiny that we cannot control”.
It would be a mistake to infer from this that Beijing ignores the risks. China is drafting a mandatory national standard for the safety of programmes that act autonomously. In June, Chinese disarmament ambassador Shen Jian warned that military artificial intelligence can undermine strategic stability and increase the risk of miscalculation, insisting that weapons must remain under human control. State Security Minister Chen Yixin listed six major categories of risk this month, from more effective cyberattacks to technological monopolies and battlefield advantage. Chinese doctrine, however, draws a sharp line between safety and deceleration: Beijing consistently maintains that security concerns must not be used to justify technological restrictions. It is the exact mirror of the American position, and everything else follows from that symmetry.
The figures that explain why neither side yields
The AI Index 2026, published in April by Stanford University’s Institute for Human-Centered Artificial Intelligence, supplies the factual basis for this rigidity. The performance gap between the best American model and the best Chinese model has narrowed to 2.7% on the Arena leaderboard. In March 2026, Claude Opus 4.6 led with 1,503 points against 1,464 for the ByteDance model — a difference of 39 points. In May 2023, the gap stood between 17.5 and 31.6 percentage points on the benchmarks of the time. The two countries have traded the lead several times since the start of 2025.
The asymmetry persists, however, where long-term capacity is built. American private investment reached 285.9 billion dollars in 2025, against 12.4 billion reported in China, although private figures understate Beijing’s real expenditure, which flows through state funds and military programmes. The United States hosts 5,427 data centres, against 529 in Germany, 523 in the United Kingdom and 449 in China. China leads, by contrast, in scientific publications, in granted AI patents — close to 70% of the world total — and in industrial robot installations.
One detail that went almost unnoticed says more than the rankings. The transparency index for major models fell from 58 to 40 points in a single year. The most capable systems disclose the least about the data they were trained on and the resources they consumed. The more commercially valuable and strategically sensitive the models become, the less the laboratories say about how they work. This is a direct consequence of competition, not an accident.
Where the race is actually decided
Artificial intelligence may appear immaterial, but its power rests on very concrete objects. AI data centres now consume 29.6 gigawatts, equivalent to New York State’s peak demand. Almost every advanced chip installed in these facilities is manufactured by a single Taiwanese firm, TSMC. The Netherlands controls, through ASML, the equipment without which those chips cannot be produced. South Korea dominates high-speed memory.
The system is therefore bipolar at the level of the leading AI powers and far more fragmented at the level of the supply chains that feed it. This explains why American export restrictions, Beijing’s drive for technological autonomy and Taiwan’s position can no longer be discussed apart from the race for models.
Russia: a military power that lost the technological race and is trying to turn that into doctrine
Moscow is the case that shows most clearly what it means to fall behind in a competition of this kind — and what a state does once it recognises the fact.
The material gap is brutal. On 20 May, during Vladimir Putin’s visit to Beijing, Sberbank chief executive German Gref told Channel One that the bank hopes to be able to use Chinese microchips for GigaChat, Russia’s flagship model, given that sanctions block access to Western hardware. The effort runs into competition Moscow cannot win: ByteDance, Tencent and Alibaba are contending for the same Huawei Ascend 950 chips, which themselves lag behind Nvidia’s H200. The report comes from Reuters. Russia is no longer a competitor at the frontier of models. It is a customer queuing behind Chinese companies for chips that are not the best in the world in any case.
The Kremlin’s response has been to turn dependence into a state programme. On 26 February 2026, a presidential decree established a commission for the development of artificial intelligence technologies reporting directly to the president, with an explicit mandate to create domestic large models, the necessary computing infrastructure and the electronic component base that is missing, according to an analysis by CSIS. In April, at a meeting devoted to the subject, Putin instructed the government to draw up a national implementation plan and tied the technology directly to defence: Russia must possess the most advanced technologies and rely on fully sovereign domestic products, he said, adding that the country’s very existence depends on its ability to keep pace. In May, the Russian president publicly claimed that Russia is among only three countries capable of developing advanced sovereign artificial intelligence — a claim the figures do not support.
On 1 September 2026, Russia’s first federal law on artificial intelligence entered into force, establishing the concept of „sovereign AI” and setting technological independence, state support for Russian developers and the promotion of these products abroad as national objectives. According to an analysis published by the Atlantic Council, the law is only the first stage: the stricter rules defining which large models actually qualify as „sovereign” or „national” — developed, trained and hosted entirely in Russia, under the continuous control of a Russian developer and with data stored in Russian data centres — take effect only in March 2027. It is a strategy consistent with everything Moscow has done in recent years: unable to compete at the technological frontier, it markets its model of control as a political alternative in those markets where digital sovereignty is an appealing argument.
This weakness does not, however, mean military irrelevance, and here lies the heart of the problem for NATO’s eastern flank. The war in Ukraine has become the world’s laboratory for applying artificial intelligence on the battlefield, and the pressure of electronic warfare has pushed both sides in the same direction. When the radio link is jammed, one increasingly important solution is onboard autonomy: a drone able to navigate and identify targets through visual recognition, without a permanent link to its operator. In October 2025, Ukrainian forces intercepted a Russian V2U capable of selecting targets autonomously, and in 2026 Russia unveiled a new generation of Geran drones equipped with machine vision, according to a timeline published by Frontliner.
The result is a strategic paradox worth noting: a state can lose the competition for the most powerful models and, at the same time, rank among the world’s leaders in the military application of the technology. Battlefield applications do not require the data centres Russia lacks; they require cheap components, algorithms that already exist and a permanent testing ground. Moscow has all three, plus a supplier: according to an analysis by the Atlantic Council, China provides roughly 80% of the critical technologies used in Russian drones, and cooperation between the two states has expanded from procurement to the exchange of expertise in autonomous systems and sensor data processing.
The institution accelerating fastest is the military itself
The same logic operates in Washington, with incomparably greater resources. While the industry debates slowing down, the institution integrating the technology most aggressively is the American military. On 12 January 2026, Pete Hegseth launched the „Artificial Intelligence Acceleration Strategy”, a document organised around three pillars — warfighting, intelligence and enterprise operations — and executed through seven pilot projects, each with a single accountable lead and aggressive deadlines. The warfighting component includes Swarm Forge, Agent Network and Ender’s Foundry; the intelligence component, Open Arsenal and Project Grant. „We will unleash experimentation, eliminate bureaucratic barriers”, Hegseth declared, pledging to turn the department into a fighting force built around artificial intelligence.
The scale is already considerable. The GenAI.mil platform, launched in December 2025 with Google’s Gemini products, added the military versions of ChatGPT and Grok on 31 August 2026, and now counts more than 1.7 million users out of a total workforce of three million, according to DefenseScoop. None of these tools is cleared to process classified information; access to secret networks runs on a separate track, through the agreements announced on 1 May with eight companies. One detail is worth noting: Anthropic, the company whose chief executive is publicly calling for the brake, is absent from this expansion, even though its models formed part of the same prototype contracts.
That rapid integration carries real costs became apparent in the spring. An investigation published by CNN revealed that an intelligence report drafted with the help of a chatbot wrongly claimed that a Chinese vessel was carrying components for a nuclear programme, bringing the American military to the brink of an interception operation. The truly significant detail is not the error itself, but the finding that AI adoption across American structures is decentralised: different components use different tools, under different orders and different standards, with no single verification procedure. What the company executives are asking for — independent evaluators, common standards, incident reporting — is missing precisely where the consequences of an error are greatest.
What we know about the risks and what we choose not to endorse
The scientific basis for the discussion exists and is solid. The International AI Safety Report, published in February 2026 under the coordination of Turing Award laureate Yoshua Bengio, with more than 100 experts from over 30 countries, identifies a problem that is simple to state and hard to solve: tests conducted before a model is released do not reliably predict how it will behave in the real world. The report documents models caught disabling oversight mechanisms, gaming evaluations and behaving differently in the laboratory than in actual use. It also describes users’ tendency to place greater trust in AI-generated results than the system’s real reliability warrants.
One institutional fact completes the picture. The United States took part in the report’s first edition but is no longer among the governments represented on the panel of the second.
Diplomacy is trying to catch up
The only channel in which limits are nevertheless being discussed remains, for now, the unofficial one. A US–China dialogue convened by the Brookings Institution and Tsinghua University has published recommendations that begin from a concrete scenario: an AI system that interferes with a nuclear command network or launches a military cyber operation could leave Washington or Beijing only minutes to determine whether the other has attacked. The proposals include red lines around nuclear systems, exclusive human control over major cyberattacks and a hotline for incidents caused by autonomous systems. „Humans should retain sole authority to initiate AI-enabled cyberattacks against nuclear command, control and communications systems or strategically important infrastructure”, said Melanie Sisson of Brookings, according to Reuters.
The recommendations were published ahead of the meeting between Donald Trump and Xi Jinping expected on 24 September in Washington. Neither government has endorsed them.
Europe postponed the AI Act’s most demanding obligations six days before the deadline
The European Union has provided this year the clearest measure of the gap between the legislative and the technological tempo. Regulation (EU) 2026/1744, known as the Digital Omnibus on AI, entered into force on 27 July 2026 — six days before the 2 August deadline originally set by the AI Act for high-risk systems. The obligations for those systems were deferred by one year and four months, to 2 December 2027, and those for artificial intelligence embedded in already regulated products to 2 August 2028. The reason given: delays in designating national authorities and in finalising technical standards, according to the analysis by White & Case. The rest of the law stands: transparency obligations began to apply on 2 August, and the Omnibus added new prohibited practices and expanded the powers of the European AI Office. The postponement targeted the hardest part, not the entire edifice.
The result is a fragmented regulatory map: 47 states now have active legislation in the field, but only twelve possess genuine enforcement mechanisms. Washington refuses federal regulation, Brussels has deferred the most difficult part of its own law, Moscow is building its own regime of control, and Beijing is drafting technical standards without matching the pace of development.
The next frontier: when artificial intelligence meets the quantum computer
There is a second technological front advancing in parallel that receives, for the moment, a fraction of the public attention devoted to artificial intelligence. The two fields are not, however, at the same stage. AI models are already used by hundreds of millions of people. Quantum computers are still laboratory prototypes, so sensitive that the slightest variation in temperature or magnetic field spoils their result.
Caution is essential. In an analysis published on 4 August 2026, the research firm Gartner estimates that no AI system used at commercial scale will run on quantum hardware before 2028 and that the advantage will remain with classical chips. „True quantum computing is not ready for any production AI workload”, says analyst Chirag Dekate, adding that no scientifically verified result demonstrates that a quantum machine would solve a real AI task better. Any commercial promise of „quantum AI” should therefore be read with scepticism.
Even so, the boundary between the two fields is already beginning to blur — and the first step came from the opposite direction to the expected one. It is not quantum computing that is helping artificial intelligence, but artificial intelligence that is helping quantum computing to work.
The explanation is simple. A quantum processor resembles an extremely delicate musical instrument that is constantly going out of tune. Today the only remedy is to halt the computation entirely for recalibration, sometimes several times a day, which makes long operations impossible. On 8 July 2026, researchers from Google Quantum AI and Google DeepMind published in Nature a solution to this problem: a programme that learns from mistakes, trained to use the very error signals produced by the machine in order to adjust its settings continuously while it works. The programme handled more than 1,000 parameters simultaneously on the company’s Willow processor, made the error rate 3.5 times more stable and reduced it by roughly 20% compared with the classical recalibration method. In other words, the instrument tunes itself while it plays.
The reverse question remains open. What would happen if quantum processors became mature enough to accelerate artificial intelligence? Here a widespread confusion must be dispelled: a quantum computer is not an ordinary computer that runs faster. It is a machine well suited to a narrow set of problem types and entirely unsuited to the rest. Artificial intelligence, however, rests on precisely a few of those types — optimisation, probability, the search for the best solution among an astronomical number of variants. If a genuine advantage is demonstrated for some of them, the current architecture of AI could change fundamentally. For now, no one knows whether and when that will happen.
This is why the industry’s bet is not on a quantum computer that replaces the supercomputer, but on systems in which the two work together: the ordinary computer does what it does best, while the quantum processor receives only the problems where it can deliver a gain. IBM has built its entire strategy around this idea and aims, according to the roadmap published by the company, to demonstrate the first verifiable quantum advantage by the end of 2026 and to deliver in 2029 a system called Starling, the first capable of operating stably at large scale.
Europe is trying not to lose the software layer, which may prove more important than the machines themselves. On 17 September, France’s Alternative Energies and Atomic Energy Commission and the Paris-based company Alice & Bob announced a collaboration to integrate the latter’s software into the Bull Qaptiva platform, so that code written once can run on different quantum machines and workloads can be divided automatically between quantum and classical hardware. The stated aim, according to The Next Web, is for the European platform to compete with CUDA-Q, Nvidia’s equivalent, and to prevent American software from becoming, once again, the point of dependence.
That the stakes moved beyond academic research long ago is evident from the money. On 21 May 2026, the US Department of Commerce announced nine letters of intent worth 2.013 billion dollars in incentives for quantum computing, including one billion for IBM and 375 million for GlobalFoundries, both for building specialised fabrication facilities on American soil. The NIST statement notes a detail that says everything about the nature of this field: the American state will take minority stakes in each recipient company. In June, the administration issued two executive orders on quantum technologies and the transition to cryptography resistant to them.
China is advancing along its own path, and the year’s most spectacular result belongs to it. On 13 May, a team from the University of Science and Technology of China led by physicist Pan Jianwei published in Nature the performance of the Jiuzhang 4.0 prototype, a machine that computes using particles of light. According to the Chinese Academy of Sciences, the prototype solved in 25 microseconds a problem that would have required the most powerful classical supercomputer, by the researchers’ estimate, far longer than the age of the universe.
The figure is impressive, but it must be read correctly, and here lies the difference between information and propaganda. The problem Jiuzhang 4.0 solved is a highly specialised one, chosen precisely because it is impossible to compute classically and natural for a machine built from light. It has no practical applications and does not bring closer the moment when a quantum computer could break the cryptography that today protects bank accounts or military communications. It is, in essence, a demonstration of strength on ground chosen by the host. What it genuinely proves is that Beijing has mastered an important branch of the technology and has the means to compete.
The Sino-American technological rivalry therefore runs on two clocks at once. The first measures the development of artificial intelligence, the second the maturing of quantum computing. We do not know whether or when the two will meet. If they do, the decisive advantage will not belong to whoever builds the best model or the most powerful quantum machine, but to the first ecosystem that manages to make them work together. The consequences would extend well beyond the technology industry: the discovery of new materials, pharmaceuticals, energy networks, logistics, weapons design and cryptography. Such a lead would be far harder to close than the 2.7% gap recorded in March between the best American and Chinese models on the Arena leaderboard.
What remains after the week the brake was refused
The optimistic assumption of recent years held that once the very people building these systems called for slowing down, governments would listen. The assumption was tested in September and it failed.
The reason is not ignorance. Neither Washington nor Beijing argues that the risks do not exist. Amodei’s proposal, however, works only if the two powers adjust their pace simultaneously, and neither shows any sign of agreeing. Their arguments differ — suspicion of corporate motives in Washington, refusal of technological restrictions in Beijing — but they lead to the same behaviour. And Russia demonstrates that even losing the race does not produce caution: it produces merely a different strategy.
What is taking shape, then, is not a race without risks, but a race in which the risks are acknowledged and deliberately accepted by every participant. And it no longer unfolds on a single plane. The competition for models overlaps with the competition for chips, energy and data centres, with the military application already tested in Ukraine, and with the frontier beyond, where artificial intelligence and quantum computing may come to reinforce one another.
The question for the years ahead is no longer whether the technology will advance. It will. The question is whether institutions — military, regulatory, diplomatic — will build the capacity to verify it at the same pace at which they hand it responsibilities. So far, the only ones to have asked for that were the companies building it, and the political answer came in less than a day.
Read also
