COLUMN: Alibaba, Anthropic Muzzled, Huawei — the Chinese Triad Winning the AI Cold War

COLUMN: Alibaba, Anthropic Muzzled, Huawei — the Chinese Triad Winning the AI Cold War

Introduction: The Day Washington Handed the Global Market to Beijing

June 12, 2026: a date that could enter the history books

On June 12, 2026, at 5:21 PM Eastern Time, the US Department of Commerce issued an export control directive ordering Anthropic to suspend access by all foreign nationals to its Claude Fable 5 and Mythos 5 models — including its own non-American employees. The directive was so absolute, so impossible to apply selectively, that Anthropic had no choice but to disable both models for all its global users. The most powerful models in the world were taken offline within hours. This is the first time in history that a government has exercised export controls over commercially deployed AI software.

The international reaction was swift. Alibaba, the Chinese tech giant already accused by Anthropic in a letter to US senators of having illicitly extracted Claude's capabilities through thousands of fraudulent accounts, accelerated the deployment of its own models. Zhipu AI released GLM-5.2, available as a free download, two days after the Anthropic ban. ByteDance increased its Open-Weight distributions on global platforms. The story of the day Washington banned its best tools is being written — with Beijing as the primary beneficiary.

A directive that even its supporters find hard to defend

The AI advisor to the former Trump administration, Dean Ball, called the decision "baffling" — noting the glaring inconsistency between the desire to loosen controls on advanced chips toward China on one side, and the global ban on Anthropic AI model access on the other. More than 76 cybersecurity experts signed an open letter calling the ban "dangerous", arguing that the capabilities that allow finding and exploiting vulnerabilities are the same that allow defending against them. Taking the world's best defensive tool offline does not reduce risk — it advantages anyone who does not play by the rules.

The directive remained in force for more than two weeks. On June 26, the Trump administration partially lifted the ban on Mythos 5, allowing its redeployment to a limited group of approved cyber defenders and infrastructure providers. But Fable 5 remains offline. And the symbolic and geopolitical damage is already done.

Alibaba vs. Anthropic: The AI War by Proxy

The accusation: intellectual property theft at industrial scale

In a letter sent to several US senators and White House officials, Anthropic accused Alibaba of conducting a large-scale campaign to "illicitly access" its Claude model through thousands of fraudulent accounts — in what it described as "the largest known attempt by a Chinese company" to free-ride on the work of leading American labs. According to Bloomberg, published June 24, 2026, this campaign specifically targeted Claude's most advanced capabilities in software engineering and agentic reasoning.

The accusation is not new in nature — since February 2026, Anthropic had already identified a campaign by DeepSeek and two other Chinese labs to illicitly extract Claude's capabilities. In April 2026, the White House publicly warned that Chinese organizations were using "tens of thousands of proxy accounts" to conduct an "industrial-scale" distillation campaign against the most advanced American AI models. This is an artificial intelligence war — fought through millions of clandestine requests.

Chinese proxy networks: an infrastructure for technological espionage

A proxy network controlling more than 20,000 fraudulent accounts in a coordinated distillation attack by Chinese AI research institutions — this figure, revealed by Anthropic in February 2026, illustrates the scale of means China is deploying to close its technological gap. Distillation involves querying a more powerful model thousands of times, then training a smaller model on those responses — a technique that is legal in itself but becomes intellectual property theft when conducted through fraudulent accounts at this scale.

According to a report published by GIGAZINE on June 26, 2026, despite all the restriction measures put in place by Anthropic — phone number verification, credit card checks, billing address checks, and even selfie identity verification for some users — access from China was never fully blocked. China has invested in a gray bypass infrastructure as sophisticated as the models it seeks to exploit.

Zhipu, ByteDance, and the Rise of Chinese Open-Weight AI

The strategic counter-attack: offering what Washington denies

Two days after Anthropic's models were taken offline, Zhipu AI published GLM-5.2 — a model available as a free download, which any company or developer can download, fine-tune, and run on their own servers. According to CNBC (June 26, 2026), this model is "free to download, fine-tune, and run on a company's servers," putting price pressure on American labs precisely at the moment when access to their tools is becoming uncertain. The timing is not accidental.

In May 2026, Chinese Open-Weight models already accounted for 61% of all tokens consumed on OpenRouter, the largest neutral language model routing platform — according to data published by Data Gravity on June 24, 2026. Four of the five most-used models on this platform are Chinese. Meta's Llama, which was the open-weight leader two years ago, does not even appear in the rankings. The silent revolution of Chinese AI is winning the battle for global distribution.

ByteDance and the ecosystem of soft dependency

ByteDance, the giant that owns TikTok, is not merely a social media company. Its AI lab, Doubao, has distributed Open-Weight models that are progressively integrating into software development, data analysis, and content creation workflows worldwide. The strategy differs from Zhipu's — less focused on performance benchmarks, more focused on integration into popular tools already used by international developers.

What Washington seems not to have grasped is that free access to reasonably capable models creates an infrastructure dependency as real as paid access to superior models. If developers, companies, and governments around the world get used to using Chinese models because American models are either too expensive or inaccessible for political reasons, the global AI landscape tilts toward Beijing — without any war or invasion being necessary.

Huawei and the $295 Billion National AI Grid

The infrastructure that will reduce dependence on American chips

While the West debates restrictions on AI model access, China is building something far more fundamental: a national computing infrastructure valued at $295 billion, mandated by the National Development and Reform Commission, with a requirement that 80% of the underlying technology be of domestic origin. This effectively excludes Nvidia and AMD from the largest computing procurement in the world. According to FrontierNews.ai (June 23, 2026), this project aims to connect thousands of data centers into a unified network by 2028.

In May 2026, nine categories of locally developed AI chips — from Huawei, Alibaba, Shanghai Biren Technology, and Moore Threads — were approved through a Chinese government security review for deployment in government and sensitive sectors. Huawei sits at the center of this strategy with its Ascend chips, which, while still lagging behind Nvidia's H100 in raw performance, are improving at a pace that concerns Western analysts.

The narrowing performance gap: the numbers speak

In June 2026, the performance gap between the best American and Chinese models has narrowed considerably. According to the Stanford HAI 2026 AI Index, cited in a report by the Foreign Affairs Forum (June 22, 2026), the Arena score differential between the best American model (Claude Opus 4.6 by Anthropic) and the best Chinese model (Dola-Seed 2.0) is now just 39 points, or 2.7%. Two years ago, the gap was measurable in double-digit percentage points. Technological convergence is happening, with or without Nvidia chips.

The American strategy of chip export controls may have slowed China's catch-up — but has not stopped it. And Open-Weight models like DeepSeek R1 or GLM-5.2 have demonstrated that top-tier performance is achievable with less powerful chips through more efficient training architectures. Algorithm quality can partially offset hardware performance — a lesson Beijing has learned and is applying methodically.

The European Collateral Effect: Digital Sovereignty at Risk

Macron calls out Washington: "strictly nationalistic"

The American decision of June 12 sent shockwaves through Europe. President Macron called the measure "strictly nationalistic" at the G7 and a "wake-up call" — unusually direct words for an American ally. Canadian Prime Minister Mark Carney called on allies to "build and diversify" rather than over-depend on American technology. These are the most visible cracks to date in the technological alliance of democracies.

The reason for this irritation is simple: the American directive did not just cut off potential adversaries' access to Anthropic's models — it cut off access by 200 partner institutions in 15 allied countries that were using those models for their own cybersecurity, intelligence, and defense research. Intelligence-sharing partners, NATO allies, democracies that had integrated these tools into critical systems. And Washington cut off access by letter, without prior consultation.

Europe facing its own technological paradox

The Anthropic episode exposed Europe's structural vulnerability in AI: the European Union has no frontier AI model comparable to those of Google, Anthropic, or OpenAI. European AI initiatives — Mistral AI (France), Aleph Alpha (Germany) — are promising but remain significantly behind. Europe chose to regulate AI before producing it — a sequencing perhaps inverted from what geopolitical competition required.

European defense, cybersecurity, and critical infrastructure companies that had built workflows around Claude found themselves overnight without their primary tool. According to AI Weekly (June 21, 2026), the directive "taught every ally that frontier American AI is a dependency revocable by letter" — a reality Europe must now integrate into its technology strategy. Accelerating funding for European sovereign AI is now a national security question, not merely an industrial preference.

OpenAI in the Crosshairs: The Next Victim of American Policy?

GPT-5.6 and the new restrictions announced

The same Friday June 26, 2026, OpenAI announced it was limiting access to its GPT-5.6 models in response to a government request — confirmation that the Anthropic case is not an isolated incident but the beginning of a broader regime of control over American AI. According to CNBC (June 26, 2026), Anthropic had already disabled access to Fable 5 and Mythos 5 to comply with the government directive, and OpenAI was following a similar logic for its most advanced models.

This simultaneous tightening sends an alarming message to global actors: the two largest American frontier AI providers are now subject to a government control regime that can alter their availability without warning. For companies, governments, and developers worldwide who have integrated these tools into critical processes, this is a supply chain vulnerability of a new kind — not a raw material shortage, but a dependence on software infrastructure controlled by a single jurisdiction.

Anthropic's valuation and the stakes of its IPO

All this turbulence arrives as Anthropic had confidentially filed for an initial public offering at a valuation of $965 billion, according to Fortune (June 13, 2026). The crisis with the US government, the taking offline of its most powerful models, the public confrontation with the Trump administration — all of this significantly complicates the path to public markets. Investors betting on Anthropic's global growth must now factor in the political risk of a company whose flagship products can be disabled by government decree.

Ironically, according to Trump himself in an interview given to Axios on June 19, 2026, it was Amazon — one of Anthropic's primary investors with $8 billion invested — that allegedly alerted the White House to a security vulnerability in Mythos, triggering the directive. A company with an interest in protecting its investment by creating government dependency — or a company that acted in good faith to protect national security? The ambiguity is total.

Europe in the Vise: Neither Washington Nor Beijing Decides for It

European technological dependence: a systemic risk

Europe observes the Sino-American rivalry in artificial intelligence from an uncomfortable position: that of a continent that massively consumes technologies it does not produce. The foundation models used by European companies and institutions are American — OpenAI, Anthropic, Google DeepMind — or potentially Chinese as these players internationalize. This dependence is not only economic: it is strategic. The data processed by these systems, the biases they embed, the rules they apply — all of this is decided in San Francisco or Beijing, not Brussels.

The European AI initiative — with players like Mistral AI in France, Aleph Alpha in Germany, or BLOOM via the BigScience consortium — remains incomparable in scale to the American and Chinese giants. European venture capital investments in AI in 2025 represented less than 15% of American investments and a fraction of Chinese commitments in this sector. Without a radical scale change, Europe risks becoming a captive market of the two great AI powers.

The European AI Act: regulation or competitive handicap?

The European AI Act, which entered into force in August 2024 and has been progressively applicable since 2025, imposes transparency, audit, and compliance obligations on so-called "high-risk" AI systems. It is a pioneering and ambitious regulatory approach. But its compliance costs weigh more heavily on European actors who must absorb them first, while American and Chinese giants can develop and deploy more freely in their domestic markets. The risk is having created a regulatory asymmetry that disadvantages European actors precisely at the moment of most intense competition.

Supporters of the AI Act argue that regulation builds trust and therefore adoption, and that European standards will become global standards — as GDPR did for data protection. Critics reply that AI's dynamic is too fast for regulation to produce its positive effects in time, and that meanwhile Alibaba, ByteDance, and Huawei are not waiting. Both arguments have merit. And Europe does not have the luxury of getting this wrong.

The Geopolitical AI Fracture: A World Splitting Into Two Ecosystems

The growing incompatibility of standards

The Sino-American rivalry in artificial intelligence is not merely producing competition between companies. It is producing a bifurcation of standards — two increasingly incompatible technological ecosystems. American models are trained on Western data, aligned with liberal values, subject to American regulations. Chinese models — notably Alibaba's Qwen, ByteDance variants, and Huawei's Ascend chips — are trained within a different regulatory framework, with censorship and information control rules specific to the Chinese context. Interoperating between the two ecosystems will become increasingly difficult.

For third-party countries — Africa, Latin America, the Middle East, Southeast Asia — this choice of ecosystem is becoming a geopolitical choice. Adopting Chinese open-weight models potentially means integrating surveillance backdoors or informational biases aligned with Beijing. Adopting American models means depending on infrastructure controlled by Washington. Europe is seeking to offer a third path — but it still lacks the computational power and capital needed to make that path credible at global scale.

The chip race: the real bottleneck

Behind the model war lies a more fundamental battle: the race for semiconductors. American export restrictions on advanced chips to China — notably Nvidia's H100 and H200 GPUs — have accelerated Chinese investment in developing alternative chips. Huawei, with its Ascend 910B chip, has demonstrated that a credible alternative, while inferior to American standards, is possible. But the United States maintains a significant technological lead in the most advanced chips, notably through control of the EUV lithography supply chain (ASML).

This technological control is the most powerful strategic weapon the West holds in this digital cold war. But it has its limits: restrictions create incentives for China to develop its own alternatives, and every Chinese technological breakthrough reduces the Western advantage. The question is not whether China can build competitive chips of its own — it is how much time it will take. Estimates range from 3 to 10 years depending on the expert. That timeline creates both a window of opportunity and a pressure for urgency on the Western side.

Conclusion: The Digital Cold War and Its First Western Losers

An incoherent American strategy that feeds its adversary

What emerges from the June 2026 analysis is the portrait of a fundamentally incoherent American strategy in the AI cold war. On one side, chip export controls to slow Chinese development. On the other, a ban on access to the most capable AI models that deprives allies of their tools and reinforces precisely the Chinese argument that Open-Weight models are a more reliable alternative. It is a policy that maximizes friction with allies while minimizing real impact on the adversary it claims to contain.

China does not need to win a frontal battle in AI to win the digital cold war. It only needs to stay in the technological race while Washington makes strategic mistakes. The Open-Weight models of Zhipu, ByteDance, and tomorrow other Chinese labs offer the rest of the world a functional alternative. And the $295 billion infrastructure of Huawei and its domestic equivalents progressively reduces dependence on American chips. The result, by 2028–2030, could be a bipolar technological world — but with a smaller American pole than its current leadership would suggest.

What the West must do — and quickly

The correct response to this situation is not American technological isolation — it is the democratic technological alliance. Allow trusted allies to access the best American tools, invest massively in European sovereign AI, create multilateral governance frameworks that enable cooperation without sacrificing security. Treat allies as allies, not as potentially dangerous customers to monitor.

The era in which the United States could simultaneously dominate AI and share it generously may be over. But the era in which it can dominate AI by limiting access to its own citizens alone is equally illusory. AI talent is global. Data is global. Markets are global. An American technological fortification strategy will leave a void that China — with its free models, its massive infrastructure, and its geopolitical patience — will gladly fill. Washington has a narrow window to correct course. Closing it out of ideology would be a historic mistake.

By Maxime Marquette, columnist

Columnist's transparency note

My positions and my limits

I am Maxime Marquette, and I openly acknowledge that I consider China a strategic threat to democratic values and the sovereignty of free nations. This position is not an irrational bias — it rests on documented facts: the repression in Hong Kong, threats against Taiwan, mass surveillance, forced technology transfer, industrial-scale intellectual property theft as documented in this article. I do not claim neutrality on these issues.

I also acknowledge that American policy in this article is presented critically. This is not anti-Americanism — it is an analysis of strategic errors that I consider factually defensible. I support a solid democratic technological alliance between the United States and its allies. What I criticize is the incoherence of the policy, not the value of the alliance.

What I do not know

I do not know exactly what Anthropic communicated to the American government during the June 15 and 16 meetings. I do not know the true extent of the security vulnerability in Mythos 5 — both parties present different versions. I do not know whether Zhipu's GLM-5.2 is genuinely comparable to Fable 5 on high-complexity tasks, despite the marketing claims. These uncertainties are real and I flag them clearly.

This article was written on June 27, 2026. The situation in the AI cold war evolves daily. Some elements may have changed since this column was written.

Sources

Primary sources

Bloomberg — White House's ban on Anthropic AI access may boost China's open-source models — June 26, 2026

Bloomberg — Anthropic accuses Alibaba of illicitly accessing its AI models — June 24, 2026

CNBC — Zhipu closes in on top US models as Anthropic and OpenAI face restrictions — June 26, 2026

Politico — Trump administration partially lifts Anthropic AI export ban — June 26, 2026

Secondary sources

AI Weekly — Washington's AI kill switch broke the alliance — June 21, 2026

Data Gravity — China's open-weight takeover — June 24, 2026

FrontierNews.ai — China's $295 billion AI grid gamble: why Huawei cannot yet deliver — June 23, 2026

Tech Policy Press — Anthropic's Mythos recall and the White House's missing AI safety playbook — June 22, 2026

GIGAZINE — The reality of Chinese proxies enabling access to Anthropic AI models — June 26, 2026

Foreign Affairs Forum — The narrowing divide: China's AI renaissance and the reconfiguration of global technological power — June 22, 2026