The algorithmic war. Fable 5 is the finger, Europe is the Moon
On June 12, 2026, the United States Department of Commerce sent Anthropic a letter that, within the industry's halls, circulated with the speed of news that truly inspires fear: mandatory suspension of the Fable 5 and Mythos models for every foreign citizen, inside or outside American borders. The official justification is national security. Groups linked to Beijing reportedly gained access to Mythos, the version without guardrails designed for cybersecurity applications, by bypassing access control systems. A serious breach, if confirmed, without a doubt. But stopping at this news, as many are doing, obsessed with the names Fable 5 and Mythos as if they were characters in a dystopian series, is like looking at the finger and not the moon.
The real question is not whether China has forced Anthropic's systems. The real question is what that block tells us about a much broader game, where the United States spends twenty-three times more than China for an advantage that has shrunk to 2.7%, where India is quietly investing 1.25 billion dollars, and where Europe risks paying the bill for a war in which it does not participate.
What is happening in the world of artificial intelligence is not a two-way conflict between Washington and Beijing. It is a multipolar chessboard where every move reveals who has technological sovereignty and who, instead, is subject to technological sovereignty. And we Europeans, at this moment, belong to the second category.
Pulling the plug
The letter from the Department of Commerce did not come out of nowhere. Anthropic had already implemented a first tightening in September 2025, blocking access to organizations controlled over 50% by Chinese entities, a move that at the time was described as the first time in recent history that an American artificial intelligence company actively limited sales to China. A significant precedent, but one that evidently was not enough.
The June 2026 block is of an entirely different scope: it concerns every foreign citizen, regardless of nationality, wherever they are. It is not a surgical measure against Beijing. It is a de facto nationalization of access to Anthropic's flagship models. The public motivation speaks of unauthorized access to Mythos, a system designed to operate without the usual safety filters in specialized contexts of cyber vulnerability analysis. Such a system, in the wrong hands, would indeed be a powerful tool.
What is not yet clear is how the breach occurred. The jailbreak hypothesis—that is, access obtained by bypassing protections with prompt engineering techniques—has circulated but has not been officially confirmed. What is known is that Anthropic had already turned to a federal court in San Francisco in March 2026 to contest certain government decisions on its models, signaling a pre-existing tension between the company and regulatory authorities. It is also known, according to some journalistic reconstructions, that the tip-off that accelerated the block might have come from circles close to Amazon, which is Anthropic's main institutional investor, with billions committed to the operation. A detail that transforms this affair from a simple security issue to an episode of corporate geopolitics.
The question to which no one has yet answered satisfactorily is this: was the total block really necessary, or was it also a signal? A message addressed not only to China, but to all the governments of the world, about who controls access to the most advanced models?
Beijing says no to Meta
If the block on Anthropic is the news of the day, the Manus case is the news of the quarter, and it tells the same story from a mirror angle. In April 2026, the Chinese National Development and Reform Commission ordered the block of Meta's acquisition of Manus, an artificial intelligence startup that had presented itself to the world as the first fully autonomous AI platform. The value of the operation exceeded two billion dollars. Beijing said no.
It is the first time that China has so explicitly used the foreign investment control measures introduced at the end of 2020, a tool that had remained largely inert for years and is now being wielded decisively. The official motivation is that the acquisition would have involved an unacceptable transfer of advanced technologies to the United States. In other words: Manus's models, developed by Chinese researchers, will not leave China. It doesn't matter how large Zuckerberg's check is.
TechCrunch reported that Meta is moving to unwind the deal after Beijing's demand, a silent surrender that says a lot about the real balance of power. But the Manus case is not isolated: China is building a series of barriers symmetrical to American ones. Researchers and executives from major tech companies must now obtain government approval before traveling abroad. The largest Chinese artificial intelligence companies—Moonshot AI, StepFun, ByteDance—must report to the government before accepting investment from American funds. And the so-called "Singapore washing"—that is, the gimmick of formally moving legal headquarters to a neutral country to bypass restrictions—has been explicitly closed.
Chinese Ambassador to the USA Xie Feng had said with a certain clarity as early as June 2023: "We are not the first to instigate, but we will not back down in the face of provocations." That phrase, at the time, seemed like diplomacy. Today it seems like a calendar.
The wall in between
To understand the Manus case, it is necessary to place it within a broader framework of regulatory policy that both superpowers have been building, brick by brick, for years. The Biden administration had already given the US Treasury the power to ban mergers, private equity operations, and venture capital investments in Chinese companies active in artificial intelligence, quantum computing, and semiconductors. In January 2025, it introduced the so-called AI Diffusion Rule, a regulation that imposed restrictions on the export of Nvidia chips to about 120 countries, with a total block for China, Russia, Iran, and North Korea.
The two moves—the American one to restrict incoming investment in Chinese tech, the Chinese one to block outgoing talent and technology—produce the same result: a wall. They call it "decoupling," but the word is misleading because it suggests an orderly and consensual separation. What is happening is more similar to a separation under a conflict regime, where both sides build fences while declaring they only want to protect themselves.
The practical result is that the world of artificial intelligence is fragmenting into separate ecosystems. Not yet entirely watertight, as demonstrated by the fact that some researchers continue to move, but increasingly differentiated in architectures, training data, safety standards, and interfaces. It is the Balkanization of the internet applied to language models, and it has been going on for years silently enough to go unnoticed by most.
Paper chips, silicon chips
There is a front in this war where the stakes are even more concrete, because it concerns the physical hardware on which artificial intelligence runs. American sanctions on advanced semiconductors have cut off China's access to the most powerful Nvidia chips, those needed to train frontier models. It was, in Washington's intentions, the definitive choke point: without the right GPUs, China cannot play in the same game.
China has responded on three fronts.
The first is the "chiplet" technique: instead of producing a single high-performance monolithic chip—difficult without the Dutch lithographic machines from ASML, which are also subject to export restrictions—several less advanced chips are assembled together, obtaining a system that overall reaches the necessary performance. It is an ingenious solution, although not without limits in terms of energy efficiency and latency.
The second front is SMIC, the Semiconductor Manufacturing International Corporation, which has demonstrated the ability to produce 7-nanometer process chips reaching about 60% of the performance of Nvidia's H100, a result that Chinese regulators evaluated as sufficient for many practical applications of artificial intelligence. So much so that in September 2025, Chinese internet authorities ordered tech companies to stop purchasing semiconductors from Nvidia, replacing them with domestic alternatives produced by Huawei, Cambricon, Alibaba, and Baidu. A move that, in Beijing's narrative, is a response to American sanctions, but which in practice accelerates the creation of a completely separate hardware ecosystem.
The third front is the most speculative but also the most interesting: analog computing. Researchers at Peking University presented a chip based on RRAM technology—resistive random-access memory—which in tests reportedly showed theoretical performance up to a thousand times higher than Nvidia H100 digital GPUs for certain categories of calculation. The news circulated in international media with tones between enthusiasm and skepticism. Analog chips have real limits: they are difficult to program, sensitive to noise, and still far from the versatility of GPUs. But they indicate a research direction that Washington, focused on the dominant digital paradigm, is not necessarily following with the same intensity.
The point is not whether China has already equaled or surpassed the United States in hardware. The point is that the certainty with which Washington counted on the chip embargo as an impassable barrier has proven, at least in part, an illusion.
Only 2.7%
We come to the numbers, because in this story numbers matter more than declarations. The AI Index Report 2026 from Stanford HAI, published in April, certifies something that until three years ago would have seemed like science fiction: the performance gap between the best American and Chinese language models has dropped to 2.7%.
To give a sense of how much the landscape has changed: in May 2023, when GPT-4 dominated the charts, the gap in Arena scores—the main evaluation benchmark for language models based on comparative human judgments—exceeded 1,300 points. In March 2026, Claude Opus 4.6 by Anthropic leads the global ranking with a score of 1,503, while Dola-Seed 2.0 by ByteDance follows at 1,464. A thirty-nine point difference. 2.7%.
The United States has produced 50 flagship models against 30 Chinese ones, and in 2025 American private investment in AI reached 285.9 billion dollars, twenty-three times the Chinese 12.4 billion. And yet the gap has shrunk almost to zero. How is this possible?
Stanford offers some answers that go beyond the simple "the Chinese copy everything" narrative. Chinese publications in the AI field account for 20.6% of global scientific citations, against 12.6% American. In industrial robotics, the ratio is nine to one in favor of China: 295,000 annual installations against 34,200. And there is a datum regarding talent, perhaps the most unsettling for Washington: the flow of AI researchers to the United States has decreased by 89% since 2017, with an 80% collapse in the last year alone. The DeepSeek researchers analyzed in the report were almost all trained in China, with about 25% having studied in the USA before returning. Stanford speaks of a "one-way knowledge transfer."
Two independent experts, Rory Green of TS Lombard and Dominic Gorecky, argue that Chinese leadership in applied AI compared to the USA and Europe is growing, not diminishing. The distinction is important: China is not necessarily winning the race for frontier language models—those that aim to replicate or surpass human cognitive abilities in a generalized way—but it is building a clear advantage in industrial AI, robotics, and manufacturing optimization. Sectors where the economic impact is measurable and immediate, not a future promise.
2.7% is a number that should make anyone building geopolitical, economic, or industrial strategies on the assumption that American technological supremacy in AI is a given think twice.
The "invisible" giant
There is a third actor in this story that Western media tend to ignore, focused on the USA-China duel as if it were a box-office movie with only two protagonists. That actor is India, and it is doing some rather remarkable things.
The Indian government has launched the IndiaAI Mission with a budget of 1.25 billion dollars, divided across research, sector applications, and training. To this is added a 1.1 billion dollar public-private co-investment program in deep tech and artificial intelligence. These are not figures comparable to the USA, but they are real figures, not announcements.
The most concrete point is the computing infrastructure. India has launched what government documents call the "50,000 GPU ambition": in the first two rounds of public tenders, it has already obtained commitments for 34,000 GPUs, with 18,000 already operational and the rest expected in the following two or three months. In December 2025, the government declared 38,000 units already active. GPUs are made available to startups, researchers, and universities at subsidized prices—65 rupees per hour (about €0.59)—a threshold designed to allow even small academic realities access to computing infrastructures that would otherwise be inaccessible.
The Indian approach is what its architects define as "techno-legal": an explicit integration between law and technology in the construction of a digital sovereignty that depends on neither Washington nor Beijing. For New Delhi, AI investments are not just an industrial matter; they are a tool for strategic autonomy in a world that is fragmenting.
The fourth India AI Impact Summit, held in February 2026, was the first hosted in the Global South and presented to the world a structured vision of how a country with 1.44 billion people, rapid growth in mobile connectivity, and an extraordinary pool of tech talent intends to compete in the new artificial intelligence economy. Italy was present with Minister Urso, a detail that says something about European interest in an area often overlooked in reasoning on AI geopolitics.
India will not win the race for frontier language models in the short term. But it is building a sovereign infrastructure and an innovation ecosystem that could produce significant effects over a five- or ten-year span. Ignoring it, as most Western public debate does, is an error of perspective.
The European bill
We come to the uncomfortable part. The part that concerns us.
Christine Lagarde, president of the European Central Bank, is not known for apocalyptic tones. Yet between November 2025 and February 2026, she multiplied warnings about AI: Europe is lagging behind China and the USA, Europe has already missed the chance to be a pioneer, Europe risks endangering its future by remaining a spectator. And then, with the strategic caution of someone who wants to open a door without knocking it down: "Europe can win on AI through application, not innovation." A statement that is simultaneously an opening and a surrender.
The numbers in the JRC 2025 report are merciless: the European Union contributes 7% of global generative AI activities. China is at 60%, the United States at 12%. And on the investment front, European artificial intelligence startups have raised less than a tenth of what American ones have: less than five billion euros in 2023 against over fifty billion in the USA.
But there is an aspect of the European position that goes beyond investment delays and concerns the very geometry of technological dependence. When in January 2025 the Biden administration introduced restrictions on the export of AI chips, the measure split Europe cleanly: 17 member countries found themselves subject to limitations, including Poland, Romania, Bulgaria, Hungary, and the Czech Republic. The other 10, including Italy, France, Germany, Ireland, and the Netherlands, fall into the restricted group of 18 nations considered close allies, which can purchase AI chips without limitations.
The division is geographically eloquent: it roughly separates Western Europe from Eastern Europe, the countries of the "historical" former NATO block from those that entered the Union more recently. It is Washington that decides which European countries have access to strategic technology and to what extent. The European Commission reacted with a joint statement from Vice-President Virkkunen and Commissioner Šefčovič: "We believe it is in the economic and security interest of the United States that the EU can purchase AI chips without limitations." A phrase that, in its diplomatic courtesy, contains a brutal truth: Europe is asking for permission.
The European Parliament used harsher words, describing American restrictions as "a direct challenge to Europe's economic resilience and technological sovereignty." But challenging and reacting are two different things. The European response arrived mainly in the form of regulation: AI Regulation V2, finalized in January 2026, introduces a classification of systems into "high-risk," "medium-risk," and "low-risk," a mandatory independent audit for the riskiest models, and a 10 billion euro European AI Trust Fund to support startups in Northern Europe. A system-wide response, certainly, but one that arrives years late compared to market development and does not address the structural problem: Europe does not have frontier language models of its own, it does not have a competitive domestic chip ecosystem, and it did not invest as it should have when the time was right.
The CEO of Sipearl, the French company producing the European Rhea processor, summarized the problem with a lucidity that leaves no room for surface-level optimism: American restrictions on chips are "another wake-up call" for a Europe that must reduce its dependence on US suppliers. But reducing that dependence requires investment, time, and a capacity for industrial coordination that the continent has rarely demonstrated.
The point is not that Europe is doing everything wrong. The regulatory approach is serious, the culture of personal data protection is a real asset in a world where trust in algorithms is becoming a competitive factor, and some European industrial realities in applied AI—from healthcare to automotive, from precision agriculture to energy networks—are genuinely competitive. But there is a substantial difference between competing in application niches and participating in the construction of global cognitive infrastructure. And today, on that second level, Europe is absent.
Pragmatic fragmentation
Where does all this lead? Not towards a new Cold War in the classic sense of the term, with two rigid and opposing blocks glaring at each other through a digital curtain. The world that is emerging is more complicated and, in a sense, more unstable.
The category that best describes 2026 is that of "pragmatic fragmentation": an international order composed of multiple actors that aggregate and disaggregate for concrete interests, access to critical resources, control of technological standards, and energy sovereignty, rather than for rigid ideological blocks. China blocks Manus but trades with Europe. The USA limits chips but maintains relations with Gulf countries that invest in Silicon Valley startups. India courts both Washington and Moscow without marrying either. Lines are blurred, alliances are situational.
In this scenario, the decisive variables will be few but crucial. The first is the degree of transatlantic coordination: if the USA and Europe manage to build common standards on model safety, data governance, and investment control, the West maintains a significant negotiating position. If they continue to proceed separately, with Brussels regulating and Washington acting, fragmentation benefits anyone who knows how to play in the cracks.
The second is the speed with which China resolves the hardware knot. If the American chip embargo produces lasting effects on the pace of development of Chinese frontier models, the 2.7% could stabilize or widen again. If domestic alternatives—analog computing, chiplets, Huawei GPUs—prove sufficient to keep pace, that gap could close entirely in the next two years.
The third variable is India. A country with India's talent pool, demographic growth, and infrastructural investments that decides to play its own game—neither with Washington nor with Beijing but as a third pole—could change the sector's geometry in ways that are difficult to anticipate today.
For Europe, the uncomfortable but necessary question is this: is technological sovereignty still possible, or has the moment already passed? It is not a rhetorical question. There are angles from which the answer is "still possible, but time is running out," and angles from which the answer is "we are already irremediably late." The truth is probably in the middle, which is another way of saying it depends on the choices that will be made in the coming years, or perhaps months.
The block on Anthropic models is not the news. The news is that we live in a world where a government can order a private company to turn off access to its tools for every foreign citizen on the planet, and that company, after attempting to resist legally, will probably obey. The news is that China can block a two-billion-dollar acquisition by Meta with an administrative order, and Meta complies without making a fuss. The news is that India is silently building what Europe discusses in committee.
The news, in the end, is that algorithmic war is not a journalistic hyperbole. It is the structure of the world we are inhabiting. And it is worth understanding which side of the border we are on.