# COLUMN: Alibaba Launches HappyHorse 1.1 — China Rushes Into the Void Left by the US

> On June 13, 2026, the American government gave Anthropic 90 minutes to block access to its Claude Fable 5 and Claude Mythos 5 models for non-American users. 90 minutes. Not 90 days. Not even 90 hours. Anthropic complied — geofencing its most powerful models to American territory in under an hour and a half. This is, in the history of commercial artificial intelligence regulatio

- Author: Maxime Marquette
- Publisher: MadMax (https://mad-max.co)
- Section: Column
- Published: 2026-06-27
- Canonical URL: https://mad-max.co/en/article/chronique-alibaba-lance-happyhorse-1-1-la-chine-s-engouffre-dans-le-vide-laisse

## Key takeaways

- On June 13, 2026, the US government gave Anthropic 90 minutes to block non-American access to its top models, and Anthropic complied.
- HappyHorse-1.0, from Alibaba's Taotian Future Life Lab, reached first place on the Artificial Analysis Video Arena in April 2026 and climbed to second on HuggingFace after the ban.
- Alibaba is alleged to have run an adversarial distillation operation using 25,000 fake accounts and 28.8 million requests; the column treats this as unconfirmed.
- The Financial Times called the export ban "a gift to China," and the column argues restrictions on software models are less effective than on chips.
- The author argues the ban pushed allies in Europe and Asia to diversify away from American models and weakened trust in US technology.

## Introduction: Washington Bans, Beijing Prospers

### June 13, 2026: A 90-Minute Decision That Changes Everything

On **June 13, 2026**, the American government gave **Anthropic** **90 minutes** to block access to its **Claude Fable 5** and **Claude Mythos 5** models for non-American users. **90 minutes**. Not 90 days. Not even 90 hours. Anthropic complied — geofencing its most powerful models to American territory in under an hour and a half. This is, in the history of **commercial artificial intelligence** regulation, an unprecedented act: the application of export control legislation (**Export Administration Regulations**) to a commercial software service, with a temporal brutality that has no parallel.

The consequences were immediate. Thousands of developers, companies and users around the world — in **Europe**, across **Asia**, throughout **Latin America**, across the **Middle East** — found themselves cut off from their working tools overnight. Without a prepared substitute. Without warning. And into that void suddenly created by American policy, one actor was waiting, ready to move: **Alibaba**. Its **HappyHorse** model, unknown to the broader public just weeks earlier, was about to make a dazzling leap up the global **AI** rankings. This was not luck. This was strategy.

### The Geopolitics of Artificial Intelligence: The New Battleground

We are living through the first great geopolitical battle of the artificial intelligence era. And like all geopolitical battles, it is not fought solely on the technology battlefield — it is fought on the fields of regulation, trade, diplomacy and trust. The American decision to ban the export of **Anthropic**'s models fits within a broader policy of restricting advanced **AI** technologies to non-allied countries — a policy inspired by the semiconductor export controls targeting **China** since **2022**.

The intention is understandable. Preventing potential adversaries from using the most advanced cognitive capabilities of American models for military, surveillance or disinformation applications is a legitimate national security objective. But the execution raises a fundamental problem: in the domain of **AI software**, unlike physical chips, open-source alternatives and models developed outside the **United States** are advancing at a speed that makes partial restriction of access to American models less a security lock than a gift handed directly to the competition.

*There is something profoundly ironic in this decision. The **United States** built its global technological dominance on openness — universities welcoming the world's brightest minds, companies whose products crossed every border. And now, out of fear of **China**, they are closing the doors. The problem is that **China** does not need to walk through their doors. It has learned to build its own.*

## HappyHorse: Anatomy of a Model That Comes From Nowhere

### 15 Billion Parameters and a Unified Architecture

The **HappyHorse-1.0** model, developed by **Alibaba**'s **Taotian Future Life Lab**, did not appear with the American ban. It had been introduced in **April 2026**, passing relatively unnoticed at the time in the constant stream of **AI** model announcements. Its technical specifications were impressive — **15 billion parameters**, a **unified 40-layer Transformer architecture**, designed for multimodal processing: text, image and video in a single coherent model. But what truly propelled **HappyHorse** onto the global stage was its performance on public benchmarks.

On the **Artificial Analysis Video Arena** — one of the most respected community evaluations in **AI** for video capabilities — **HappyHorse-1.0** had reached first place in **April 2026** with an **ELO of 1383** for text-to-video and **1413** for image-to-video. These scores placed the Chinese model above all its American competitors in this specific segment. This was not marginal. It was first place. And the developer community, which uses these arenas as an objective reference, took note.

### The Rise on HuggingFace After the American Ban

Before **June 13, 2026**, **HappyHorse** was a respected but non-dominant model on the **HuggingFace** platform — the global reference for the open-source **AI** community. Within 72 hours of the American ban on **Fable 5** and **Mythos 5**, downloads of **HappyHorse** exploded. European, Asian and Latin American developers, searching for an available and powerful substitute, turned en masse to **Alibaba**'s model. Within days, **HappyHorse** climbed to **second place globally** on **HuggingFace**.

This meteoric rise is not merely a popularity indicator — it is a structural shift in the global **AI** ecosystem. Developers building applications on the foundation of **HappyHorse** are not just temporary users. They write code, create integrations, build teams around this model. Switching base models six months later will be costly. Adoption creates loyalty. And that loyalty, once acquired at significant international scale, is hard to dislodge. The **United States** has inadvertently accelerated the global adoption of a Chinese model.

*I never expected to write a column about an **AI** model called **HappyHorse**. The name invites a smile. The reality it represents, far less so. **Alibaba** has just achieved, through **Washington**'s regulatory clumsiness, what years of marketing would not have produced: a massive, rapid, global adoption of its model by the international developer community. It is a gift American strategists had probably not planned to give.*

## The Adversarial Distillation Affair: 25,000 Fake Accounts

### 28.8 Million Requests and a Methodical Extraction

But there is a darker layer in the **HappyHorse** story. According to information published by several specialist sources, **Alibaba** allegedly conducted a large-scale **adversarial distillation** operation against **Anthropic**'s models. The method: creating tens of thousands of fictitious accounts on **Anthropic**'s platforms to generate millions of requests and use the responses as training data for its own model. Available estimates speak of **25,000 fake accounts** and **28.8 million requests** — a volume that far exceeds standard detection thresholds and suggests an organized and deliberate operation.

This practice — if confirmed — raises serious legal and ethical questions. **Anthropic**'s terms of service explicitly prohibit large-scale data extraction and training competing models on the basis of their responses. Several intellectual property law experts have noted that this practice could constitute a violation of contractual conditions and potentially a form of unfair competition. But jurisprudence in this area remains embryonic — courts have not yet established solid precedents on the question of large-scale **AI** model distillation.

### Model Distillation: Common Practice, Blurry Boundaries

Model distillation — using the outputs of a powerful model to train a smaller or different model — is a widely used technique in the **AI** research community. Many open-source models acknowledge having used data generated by **OpenAI** models or other major players in their training processes. The boundary between legitimate research and unfair commercial exploitation is, under current law, deeply unclear.

What distinguishes the alleged **Alibaba** case is the industrial scale of the operation and the use of fake accounts to circumvent detection measures. This is not a researcher using a few thousand requests for an academic study. It is, according to the allegations, an organized operation with considerable resources to methodically extract the knowledge of a competing model. If these allegations prove founded, they raise a problem that goes beyond commercial law — they raise the question of reciprocity in the global **AI** ecosystem and the ability of Western players to protect their investments in fundamental research.

*Adversarial distillation is the new intellectual property theft — sophisticated, difficult to prove, and perfectly adapted to the digital era. If **Alibaba** genuinely used **28 million requests** to extract **Anthropic**'s knowledge, it is the equivalent of copying the contents of a library while having the courtesy to ask for access first. Legally ambiguous. Morally questionable. Strategically brilliant. Choose the adjective that suits you — I am still trying to reconcile them all.*

## The Boomerang Effect of American Export Policy

### A "Gift to China" According to the Financial Times

The phrasing appeared in the **Financial Times** on **June 15, 2026**, two days after the ban: "**US export ban is a gift to China**." This is not the framing of a partisan commentator — it is the cold assessment of a conservative financial newspaper read by decision-makers worldwide. The reasoning is clear: by cutting access to the most advanced American models for non-American users, **Washington** created a need that non-American models — primarily Chinese ones — are particularly well positioned to fill.

Analysis of the situation reveals a deep strategic paradox. Export control policies on semiconductors — **Nvidia**, **AMD**, **Intel** chips — carry a certain logic: physical chips are difficult to duplicate, and restrictions create real bottlenecks in computing capacity. But **AI** models are software. And in the software domain, the open-source paradigm has fundamentally changed the game: when a powerful model is available open source, it can be studied, reproduced and improved by thousands of independent researchers everywhere in the world.

### DeepSeek and the Chinese Open-Source Ecosystem: The Hydra That Restrictions Feed

**China** has developed, despite American restrictions on semiconductors, a robust ecosystem of open-source **AI** models. **DeepSeek V4** reached performance comparable to **Claude Opus 4.6** according to evaluations published by **Fortune** in **June 2026**. **Qwen**, **Alibaba**'s model, has repeatedly demonstrated competitive capabilities with the best American models on specific tasks. And now **HappyHorse** is taking the lead on video benchmarks.

The Chinese open-source ecosystem is not a marginal phenomenon. It is a nationally supported effort backed by massive investment, a world-class research community — many of whom were trained at American universities before being turned away by the restrictive visa policies of the **Trump** era — and a regulatory framework that, for all its considerable failings on civil liberties, favors massive industrial investment in technologies deemed strategic. Restricting access to American models in this context does not block **China**. It hands it a market.

*I see in this situation the same pattern as in the energy domain: the **West** imposes restrictions on **China**, **China** develops its own capabilities, the restrictions become obsolete, and in the meantime the **West** has lost market share and technological partners. I am not denouncing naivety. I am denouncing strategic inefficiency. Restrictions work when they are total and sustained by everyone. Applied halfway, they only train the competitors one wants to slow down.*

## Betrayed Allies: Europe and Asia in Anger

### Partners Given No Notice, No Alternative

The American government's decision to cut access to **Anthropic**'s models in **90 minutes** did not only affect individual users. It hit European, Asian and Latin American companies full force — companies that had built entire products, services and processes on the foundation of these models. Startups that had raised funding on the strength of **Claude Fable 5**'s capabilities. Multinationals whose development teams were in the middle of critical implementations. Academic institutions in the midst of ongoing research.

**Al Jazeera** reported that the decision had "**further strained**" American alliances in **Asia** and **Europe**. European diplomats expressed their consternation privately, pointing out that their companies were being treated as potential adversaries when they are **NATO** allies and first-tier commercial partners. This brutality — 90 minutes to cut access for major economic partners — generated a shockwave across already-fragile transatlantic technological relations, relationships already weakened by the trade tensions of the Trump era.

### The Paradox of Technological Trust

American export control policy on **AI** rests on an implicit assumption: that allies and partners of the **United States** are potential vectors for transferring technology to adversaries. This is an assumption of generalized mistrust that, applied without nuance, transforms partners into suspects. Is a French company using **Claude Fable 5** a risk to American national security? Is a Japanese startup building translation tools on **Mythos 5** threatening American interests?

The implicit answer of American policy is: possibly. That "possibly" is exactly what fuels its partners' distrust. And that distrust has practical consequences: European and Asian companies that had bet on American models are actively beginning to plan diversification strategies toward non-American models. Not out of anti-American ideology — but out of basic commercial prudence. No one can afford to see their technological infrastructure cut in **90 minutes** by the unilateral decision of a foreign government.

*If I were the Chief Technology Officer of a major European company in **June 2026**, I would have had a single thought after the American ban on **Anthropic**: how do I diversify my dependencies so that no single government can ever cut my tools in **90 minutes** again? That is not anti-Americanism. That is elementary risk management. And that risk management directly benefits Chinese and European competitors of American models.*

## The Technology Race From a Different Angle

### What Does Leading in AI Video Actually Mean?

**HappyHorse**'s dominance on video benchmarks deserves careful contextualization. **AI** benchmark rankings are snapshots — they measure performance on specific tasks at a given moment, and may not reflect the model's actual value for concrete applications. A model can rank first on a benchmark and perform poorly for particular professional use cases. The **AI** research community knows that benchmarks do not tell the whole story.

That said, an **ELO of 1413** in image-to-video is not a statistical detail. It is a measure assessed through comparative human votes — thousands of people preferred **HappyHorse**'s video outputs to those of its competitors. In a domain where visual quality and temporal coherence are measurable by any user who watches a video, this type of benchmark has real value. Creative applications — advertising, entertainment, media — are directly affected. And in these sectors, **HappyHorse**'s dominance could translate into concrete market share.

### The Video Content Industry: A Massive Market at Stake

**AI**-generated video is one of the fastest-growing markets in the technology sector. Market studies project that the global market for **AI**-generated video could exceed **1.2 trillion dollars** by **2030**. The applications are multiple: personalized advertising, large-scale content production, entertainment, training, design prototyping. Being in the lead on this segment at this precise stage of market development is an extraordinarily valuable strategic position.

Hollywood studios, global advertising agencies, streaming platforms — all are integrating video generation tools into their workflows. The decisions being made today about which model to use will have consequences lasting years. If **HappyHorse** and its successors establish themselves as the benchmark for high-end video generation in these industries, **Alibaba** will have achieved a first-magnitude commercial coup — directly facilitated by the American export policy that eliminated its competitors from the global market.

***AI** video will change content production as profoundly as the internet changed music distribution. And in this revolution, **China** is taking a leading position on a crucial segment, while the **United States** closes its doors to its own allies. I am not sure the architects of this policy have thought carefully about the ten-year consequences. I am not even sure they thought about the ten-month consequences.*

## Alibaba's Strategy: Opportunism or Long-Term Vision?

### The Taotian Future Life Lab: A Serious Research Infrastructure

The **Taotian Future Life Lab**, the **Alibaba** entity that developed **HappyHorse**, is not a structure assembled in a hurry to capitalize on the American ban. It is a fundamental research organization that has been working for several years on multimodal architectures. The choice of a **unified 40-layer Transformer architecture** to simultaneously process text, image and video reflects a coherent architectural strategy — building a single foundation model capable of handling different data types with a single latent representation, rather than specialized modules for each content type.

This architectural approach has real advantages in terms of computational efficiency and generalization. It is also, strategically, a response to the dominant paradigm of American models, which have tended to develop separate models for text, image and video. If multimodal unification genuinely produces better overall performance with fewer resources, **Alibaba** will have contributed an important architectural advance — independent of the controversies over its data acquisition methods.

### Version 1.1 and the Accelerating Development Cycle

The release of **HappyHorse 1.1** shortly after the initial version illustrates the development cadence that **Alibaba** is maintaining on this model. In the **AI** ecosystem, the speed of iterations is a competitive advantage as important as the quality of the initial model. Bug fixes, performance improvements, safety adjustments — all of this determines the extent to which developers trust a model for production applications. A stable, improved **version 1.1** sends a signal to developers: the team is committed, responsive, and the model will be maintained and improved.

The open-source availability strategy is also deliberate. By making **HappyHorse** freely downloadable on **HuggingFace**, **Alibaba** forgoes immediate direct revenue in exchange for massive adoption. The goal is ecosystem dominance — building a developer community around the model, generating external contributions and improvements, and creating the long-term dependency that will then be monetized through cloud services, premium APIs and enterprise services. This is exactly the strategy **OpenAI** used with its open-source versions — and it works.

*I find myself admiring **Alibaba**'s strategy while wondering whether I am right to do so. It is brilliant — using the American ban as an accelerator, publishing open source to build the ecosystem, releasing impressive benchmarks at the right moment. But there are also questions about adversarial distillation, about the data used, about transparency. Admiration for the strategy must not obscure scrutiny of the methods. I prefer to hold both together.*

## Anthropic's Response: Between Compliance and Frustration

### A Company Caught Between Two Fires

**Anthropic** is in an extraordinarily difficult position. Founded on principles of **AI** safety and alignment, its stated mission is to build **AI** systems that benefit humanity. This mission is global by definition — it assumes that the benefits of safe **AI** spread as widely as possible. The government injunction to cut international access to its most advanced models enters directly into tension with this mission. But refusing to comply was, legally, a difficult position to defend.

By complying, **Anthropic** lost international clients to its competitors. Companies that had built their products on **Claude** were forced to migrate — and many will not return. Trust is the most precious asset of a technology infrastructure provider, and that trust was damaged — not through **Anthropic**'s fault, but through the American government's decision. This is a form of **regulatory collateral damage** that American decision-makers should have anticipated but evidently underestimated.

### The Industry's Response: Between Mobilization and Resignation

Several major companies in the **AI** sector — **OpenAI**, **Google DeepMind**, and others — expressed privately, according to sources cited by **Bloomberg**, their concerns about the direction American policy is taking. The shared fear: overly broad restrictions could harm the overall American competitive position more than they protect national security. The open model — contributing to global research, attracting the world's best researchers, benefiting from global adoption — is what has allowed the **United States** to dominate **AI** until now. Closing it off risks undermining the foundations of that dominance.

But the political pressure in **Washington** is moving in the opposite direction. The American **Congress** is in protection mode, not openness mode. Legislators with only a superficial understanding of **AI** are making decisions whose consequences will play out over a decade. And in the meantime, the American **AI** industry is navigating growing regulatory uncertainty that complicates its international relationships, product decisions and global recruitment capabilities.

***Anthropic** did not betray its international users. Its government betrayed them. This distinction matters. Private companies are subject to the laws of their country — that is the very foundation of sovereignty. But when those laws serve short-term interests that compromise long-term alliances, one must ask who is truly benefiting. The answers I find are not reassuring.*

## European Digital Sovereignty: Missed Opportunity or One to Seize?

### Europe Between Two Giants: A Space Still to Define

Every time the Sino-American technological rivalry intensifies, the same question arises: where is **Europe** in all of this? The American ban on **Anthropic**'s models and the rise of **HappyHorse** are yet another illustration of a structural European dependency on **AI** models developed outside its borders. **Europe** consumes **AI** services produced either in **San Francisco** or in **Hangzhou** — rarely in **Paris**, **Berlin** or **Amsterdam**.

The European **AI Act** — which came into application in 2024 — established a regulatory framework but did not resolve the problem of industrial sovereignty. **France** has **Mistral AI** — a French gem whose models are respected in the global community. **Germany** has world-class **AI** research laboratories. But no European company yet has the capitalization, data resources and computing power needed to compete head-on with the American and Chinese giants. This gap will not close by itself.

### Investment in Sovereign AI: A Strategic Necessity

The American ban on **Anthropic** should be a wake-up call for European decision-makers. If a unilateral decision by **Washington** can cut millions of Europeans' access to their **AI** working tools in **90 minutes**, then technological dependency is, like energy dependency, a sovereignty risk. The answer cannot simply be to fall back on Chinese models as a substitute — replacing American dependency with Chinese dependency is not a sovereignty policy.

The answer must be to build a proper European capacity — models trained on data that respects European standards, hosted on European cloud infrastructure, subject to European law, technically capable of competing with the world's best. This is expensive. It takes time. It requires coordinated industrial policies at European scale that community institutions still struggle to deploy at the speed of **AI**. But the alternative — being a spectator of the Sino-American **AI** rivalry while alternately depending on both — is an untenable position in the long run.

*I feel like I have been writing the same paragraph for ten years, with different subjects. Energy, semiconductors, social platforms, and now **AI**. Each time: **Europe** is dependent, **Europe** must invest in its sovereignty, **Europe** passes resolutions. Each time, a few years later, we write the same assessment. I have not given up on **Europe**. But I am beginning to wonder how many additional wake-up calls are needed before the resolutions become industrial realities.*

## Implications for Global AI Security

### When American National Security Undermines Global AI Safety

American **AI** export control policy rests on a national security logic that presupposes that keeping the most advanced models out of adversaries' reach will reduce the risks of malicious use. This logic has a certain validity for truly differentiated capabilities — models capable of designing biological or chemical weapons, for example. But for text and video generation models like **Fable 5** and **Mythos 5**, the logic is less clear.

In practice, restricting access pushes developers toward models that are less rigorously evaluated on safety. **Anthropic** is, in the sector, recognized for its particularly serious investment in **AI** safety research — it is literally its reason for being. By forcing non-American users to migrate toward **HappyHorse** or other alternative models, American policy directs them toward systems on which safety research is less advanced and less transparent. This is not an improved safety outcome. It is potentially a degraded safety outcome at global scale.

### The Control Dilemma in an Open-Source World

The fundamental problem is that the **AI** world has evolved irreversibly toward a partially open-source paradigm. When a model is published open source — like **Meta**'s **Llama**, like **Alibaba**'s **HappyHorse**, like **DeepSeek** — its weights are downloadable by anyone in the world. No export policy can contain a model whose parameters are freely available on the internet. The only thing restrictions can contain are proprietary models accessible via API — like those from **Anthropic**.

This reality creates a painful policy paradox. Restricting American proprietary models — the best documented, the best evaluated in terms of safety, the most transparent in their usage policies — favors the adoption of open-source and non-American models, over which no control lever exists. The policy that claims to control **AI** risks actually creates, in reality, the conditions for an acceleration of those risks outside the American perimeter. This paradoxical outcome warrants a serious reassessment of the strategy.

*I am not an **AI** safety engineer. But I understand systems. And when a policy designed to reduce a risk systematically produces the opposite effect — pushing users toward less safe alternatives, favoring the proliferation of uncontrollable open-source models — one must have the courage to say it: the policy is not working. And continuing to apply it out of inertia is obstinacy, not strategy.*

## Possible Futures: Where Is This Race Headed?

### Scenario 1: Global Technological Fragmentation

The most plausible short-term scenario is an accelerated fragmentation of the global **AI** ecosystem. A **splintered AI internet** divided into geopolitical blocs: American models for close allies of the **United States**, Chinese models for **Beijing**'s sphere of influence, European models for those seeking to avoid external dependencies, Indian models for a market of **1.4 billion** people that refuses to choose sides. This fragmentation is technically inefficient — it duplicates research investment, creates incompatibilities and slows global innovation.

But it is perhaps politically inevitable. The trust between major powers necessary to maintain a genuinely global **AI** ecosystem simply does not exist in **2026**. And incidents like the American ban on **Anthropic** accelerate precisely this fragmentation that both parties should theoretically want to avoid. The result will be a world where the interoperability of **AI** systems becomes a geopolitical challenge as complex as the interoperability of telephone standards was in the twentieth century — but with infinitely higher stakes.

### Scenario 2: The Open-Source Race as the New Norm

A second scenario — more optimistic in some respects — is that American restrictions accelerate the adoption of open source as the dominant norm for foundation models. If proprietary American models are only accessible to American users, global developers structurally turn toward open-source models — whether Chinese, European, Indian or otherwise. Open source, by definition, escapes national restrictions: once the weights are published, anyone can download, use and modify them.

In this scenario, **Alibaba** and its Chinese peers who have massively invested in high-performing open-source models would be the big medium-term winners. The **United States** would maintain an edge on the most advanced proprietary models — but for an American-only market. American global **AI** dominance, built on global adoption, would be replaced by a fragmented dominance: technically superior on certain axes but geographically limited. This represents a considerable strategic impoverishment of the **United States**' global position.

*The worst of both scenarios is not fragmentation — it is fragmentation without the **West** having realized what it has lost. A techno-geopolitical fracture managed with clear-headedness can be navigated. A fracture built by accumulation of short-sighted decisions, without being anticipated, is far more dangerous. And for now, I struggle to see a long-term American strategy for global **AI**. What I see is reaction. And in strategy, reaction loses.*

## What All of This Really Means

### Technology as an Extension of Geopolitics

The story of **HappyHorse** and the American ban on **Anthropic** is, at its core, a geopolitical story dressed in technology. It illustrates that in the world of **2026**, technology is no longer a sphere separate from geopolitics — it is its natural extension, the new battleground where the century's power balances are decided. **AI** models are not simply commercial tools. They are infrastructures of cognitive power at national and international scale.

Which language does a model understand best? Which cultural biases does it encode in its responses? What data did it use for training, and with what angles? These questions are not technical — they are political. A world where the dominant **AI** models are developed in **China** is a different world from one where they are developed in the **United States** or in **Europe**. The values, representations and conceptual frameworks encoded in these models will shape the way billions of people process and interpret information. This is an influence power without precedent in human history.

### The West Must Choose: Competition or Retreat

Faced with the rise of **HappyHorse**, the ascent of **DeepSeek**, the Chinese open-source ecosystem in full expansion, the **West** has a fundamental decision to make. It can choose competition — massive investment in **AI** research and development, openness policies that attract the world's best minds, building robust technological alliances with its partners. Or it can choose defensive retreat — growing restrictions, technological walls, progressive isolation — at the risk of losing not only market share but also the legitimacy of a global technological model.

The **West** won the Cold War because it demonstrated that its model was open, creative, economically superior to the closed Soviet model. The temptation, faced with **China**, to reproduce Soviet methods — centralized planning, control of exchanges, closure of technological borders — is understandable but strategically dangerous. The **West** will win the **AI** competition through quality, openness and innovation — or it will not win it. There is no third path.

*I am pro-Western not because the **West** is perfect — it manifestly is not — but because I prefer its imperfect values to the available alternatives. Freedom of expression. The rule of law. Pluralist democracy. These values must also apply to technology: openness, transparency, free competition. When the **West** abandons these principles in the technological competition, it surrenders its most precious comparative advantage. The one that justifies fighting for it.*

## AI Regulation in 2026: When Governance Cannot Keep Up With Innovation

### The Regulatory Vacuum as Fertile Ground for Chinese Competition

The **Alibaba HappyHorse** affair highlights a structural problem of global **AI** governance: state regulation cannot keep pace with the rhythm of technological innovation. American restrictions on access to advanced **AI** models were designed to maintain a competitive edge over **China**. But their practical application creates unanticipated perverse effects: pushing international users toward Chinese alternatives, weakening the position of American developers on global markets, and creating precedents for technological fragmentation that complicate international cooperation on **AI** standards.

This is not a question of incompetent American regulators. It is that the **speed of AI innovation** is fundamentally incompatible with traditional legislative and regulatory cycles, which take months or years to produce rules. A restriction adopted on the basis of existing models in **January 2026** is potentially obsolete by **July 2026**, when new models with different capabilities have emerged. **Alibaba's HappyHorse was able to appear and capture global market share in the time it took Washington to put its restrictions on competing models in place.**

### Toward International Coordination on AI Standards

The solution to the **AI** governance problem is not national — it is international. Minimum technical standards on model safety, training data transparency, documentation of capabilities and limitations, and audit and traceability obligations, applied consistently by the world's leading economies, would create a more equitable and safer playing field than current unilateral restrictions. Organizations like the **OECD**, **ISO** and **ITU** are working on these standards, but their progress is too slow relative to the speed of model deployment.

The **Seoul Declaration on AI Safety** of **2024**, signed by around twenty countries, is a beginning. It should be transformed into binding obligations with verification mechanisms. **France**, which chairs the **AI Action Summit** for **2025-2026**, has attempted to accelerate this process. But without American and Chinese engagement in a common framework — engagement that is difficult to obtain in the current geopolitical climate — the standards will remain statements of good intentions rather than real constraints. **AI governance is the test of democracies' capacity to cooperate even when their interests diverge.**

*AI governance is too important to be left solely to governments and technology companies. Civil society, independent researchers and ethics experts must be integrated into the regulatory process — not as occasional consultants, but as permanent actors. Closed technocracy produces rules that no one understands and no one respects. Inclusive democracy produces rules that carry legitimacy.*

## AI and Economic Competition: Who Will Capture the Value?

### The AI Value Chain: An Unequal Distribution

The rise of models like **Alibaba**'s **HappyHorse** raises a fundamental economic question: who captures the value generated by artificial intelligence? The **AI** value chain is complex, distributed across several actors: providers of computing power (**NVIDIA** and its competitors), developers of foundation models (**OpenAI**, **Anthropic**, **Alibaba**, **Google**), application developers who use these models, and end users who benefit from productivity gains. In this chain, value tends to concentrate at the extremities: compute providers and model developers capture substantial rents, while applications and end users capture primarily efficiency gains.

The rise of **Alibaba** as a provider of competitive foundation models — now globally accessible via HappyHorse — shifts part of this rent toward **China**. For countries that have not developed their own foundation models, dependency on American or Chinese providers creates a strategic vulnerability similar to energy dependency. **Europe**, which has not yet produced a globally competitive foundation model, faces this dilemma: depend on **Washington** (with the risk of geopolitical restrictions) or on **Beijing** (with different security risks). **Digital sovereignty requires European foundation models — and their funding remains insufficient.**

### Global South Countries in the AI Race

The competition between American and Chinese **AI** models for global markets has a dimension often overlooked: countries of the **Global South** — India, Brazil, sub-Saharan Africa, Southeast Asia — are the terrain of this competition, but also the potential beneficiaries of democratized access to **AI** technologies. **Alibaba**'s **HappyHorse**, being accessible where American restrictions apply, offers emerging markets generative **AI** capabilities they would not otherwise have had access to. This is a consideration that Western restriction policies must factor in: in their battle against Chinese technological dominance, they risk alienating Southern countries who will see American restrictions as another episode of techno-colonialism.

The optimal strategy for the West is therefore not to restrict access to **AI** models in emerging markets — it is to offer better models, more accessible, better adapted to local languages and contexts, with transparent terms of use and data sovereignty guarantees. **Winning the AI race in the Global South is achieved through quality and accessibility, not restrictions.** This is a lesson that Washington still seems to struggle to integrate into its technology strategy.

*The American-Chinese competition in **AI** risks turning Global South countries into a battlefield rather than partners. This repetition of the technological Cold War serves neither the **West**'s interests nor those of emerging countries. **AI** should be a lever for global development — not one more geopolitical tool. I know that sounds naive. But someone has to say it.*

## Conclusion: The Void Washington Created Now Belongs to Beijing

### The Lesson That 90 Minutes Taught Us

In **90 minutes**, the American government created a void in global access to advanced **AI** models. Within a few weeks, that void was partially filled by a model developed by a company whose government maintains close ties with the Chinese state. The symmetry is perfect — and damning. The policy designed to protect American technological advantage accelerated the global adoption of a Chinese competitor. This is a strategic failure that deserves to be named clearly, without euphemism.

The **HappyHorse** case is not an anomaly. It is a preview of what policies of **AI** export controls will systematically produce when they fail to distinguish partners from adversaries, when they treat European and Asian allies as vectors of risk, and when they forget that technological dominance is built on global adoption, not selective exclusion. If the policy does not change, the next **HappyHorse** models will bear different names, but the dynamic will be the same: American vacuum, Chinese gain.

### A Column Is Not a Prediction — It Is a Warning

I do not know how this technological competition between the **United States** and **China** will end. I do not know whether **HappyHorse** will stay at the top of the benchmarks in six months, whether the allegations of adversarial distillation will be confirmed and sanctioned, whether American policies will soften or harden. What I know is that decisions made today in government offices in **Washington**, in research laboratories in **Hangzhou** and in startups in **Paris** and **Berlin**, will shape the global technological landscape for decades to come. And that this landscape will be decisive for the capacity of democracies to maintain their advantage against authoritarian systems that, for their part, have a very clear long-term vision.

The lesson of **HappyHorse** is simple: in the technology war, your adversary's strategic mistake is your best ally. **Beijing** knows this. It is waiting patiently for Washington's next **90 minutes**.

*I end this column with the uncomfortable feeling of having documented a defeat that no one has yet officially acknowledged. The **United States** retains considerable technological leads. But leads that are not deployed globally are not victories — they are fortresses. And fortresses, in history, are always eventually bypassed. Sometimes in **90 minutes**. Sometimes in a generation. But always.*

**By Maxime Marquette, columnist**

## Columnist's transparency note

### Editorial stance

I am not a journalist but a **columnist and analyst**. My expertise lies in observing and analysing the geopolitical, economic and strategic dynamics that shape our world. My work is to take political strategies apart, to understand global economic movements, to put the decisions of international actors in context and to offer analytical perspectives on the changes that are redefining our societies.

I do not claim the cold objectivity of traditional journalism, which stops at reporting facts. **I claim analytical clarity**, rigorous interpretation and a deep understanding of the complex issues that concern us all. My role is to give meaning to the facts, to place them in their historical and strategic context, and to offer a critical reading of events.

### Method and sources

This text keeps the basic distinction between **verified facts and interpretive analysis**. The rule never changes: a factual claim is published only when a verifiable source supports it, and the sources this article actually used are listed under **Sources**, never here.

**Primary source families** used by the publication, when the story has them: official statements from governments and international institutions, public statements by political leaders, reports from intergovernmental organisations, dispatches from recognised international news agencies.

**Secondary source families**: specialised publications, internationally recognised news media, analyses from established research institutions, reports from industry organisations.

When an article cites statistical, economic or geopolitical data, it comes from the institutions that produce it (intergovernmental organisations, central banks, national statistics offices), and the exact institution is named under **Sources**.

### Nature of the analysis

The analyses, interpretations and perspectives in the analytical sections of this article are a **critical, contextual synthesis** based on the information available, the trends observed and the expert comments cited in the sources consulted.

My role is to interpret these facts, to set them in the context of today's geopolitical and economic dynamics, and to give them a coherent meaning within the larger story of the changes shaping our time. These analyses reflect expertise built through continuous observation of international affairs and an understanding of the strategic mechanisms that drive global actors.

This article describes a documented state of affairs on its publication date, not a prediction: later developments may shift its outlook. No update is promised in advance; when an article is corrected or completed, the change is dated in the text.

## Sources

### Primary sources

- [Bloomberg — White House ban on Anthropic AI access may boost China's open-source models — June 26, 2026](https://www.bloomberg.com/news/newsletters/2026-06-26/white-house-s-ban-on-anthropic-ai-access-may-boost-china-s-open-source-models)

- [Fortune — US Anthropic ban: a boon for open-source AI including DeepSeek — June 16, 2026](https://fortune.com/2026/06/16/us-anthropic-ban-open-source-ai-deepseek-zai/)

- [The AI Dude — The US export ban on Claude Fable 5 and Mythos 5: explained — June 2026](https://theaidude.net/blog/us-export-ban-on-claude-fable-5-mythos-5-explained)

- [DataNorth — Alibaba launches HappyHorse, its benchmark-leading video model — June 2026](https://datanorth.ai/news/alibaba-launches-happyhorse-video-model)

### Secondary sources

- [Al Jazeera — US export ban on Anthropic's AI models further strains alliances — June 19, 2026](https://www.aljazeera.com/news/2026/6/19/us-export-ban-on-anthropics-ai-models-further-strains-alliances)

- [HappyHorse — Official information on the HappyHorse 1.1 model — 2026](https://happyhorse1.info)

- Foreign Policy — AI export controls: the American strategy against China — June 2026

---

Source: Maxime Marquette, MadMax. When citing, name the author and link https://mad-max.co/en/article/chronique-alibaba-lance-happyhorse-1-1-la-chine-s-engouffre-dans-le-vide-laisse