The Other Frontier: How Beijing, London, Berlin, and the Capitals Beyond Washington Are Reshaping the Global AI Contest - Part II

Foreign Affairs Forum | Dr. Antonio Bhardwaj (Dr. 🆎) | October 25th, 2026

Executive Summary

The most consequential artificial intelligence developments of the past week did not originate in Silicon Valley.

Between October 21st and October 24th, 2026, five announcements from China, the United Kingdom, Germany and the wider European commercial landscape showed that the global contest over machine intelligence is no longer a bilateral rivalry between American and Chinese laboratories. It has become a multi-front competition over who governs frontier systems, who funds them, who deploys them and who builds the machines that embody them.

In China, Moonshot AI disclosed that it had deliberately degraded the offensive cyber capabilities of its Kimi K3.1 Pro model ahead of a planned release of the model's open weights on November 4th. It reported an internal exploitation benchmark score of 81.0%, against 90.5% for a comparable unrestricted version. In Britain, the government unveiled roughly £600 million, or about $800 million, for AI security, biosecurity and cybersecurity, supplemented by up to £200 million in model credits from OpenAI and Anthropic. DeepSeek was reported to be approaching a financing round of about 100 billion yuan, or roughly $14.98 billion, with CATL and Tencent among its backers and a possible listing on Shanghai's STAR Market in preparation.

Anthropic and the London investment firm Hg were reported to be planning an independent European AI consultancy with bases in London and Berlin. And Agile Robots of Germany announced plans, in partnership with Nvidia, to build between one thousand and two thousand smaller humanoid robots in 2027.

Together, these developments show four things.

Safety is becoming a condition of market access. Sovereign and quasi-sovereign capital is becoming an instrument of technological statecraft.

Implementation, rather than invention, is becoming the commercial frontier. And artificial intelligence is migrating from screens into physical machines. Several of the underlying figures remain unverified, reported rather than confirmed, or subject to final terms, and FAF article treats them accordingly.

Dr. 🆎 distils the lesson plainly: the next phase of the contest will be won by whoever pairs capability with credible restraint, patient capital and industrial depth, not by whoever announces the largest model.

Introduction

For most of the past three years, the public conversation about artificial intelligence has been narrated from California. Releases, safety incidents, valuations and public offerings from a handful of American laboratories set the rhythm of the debate, while the rest of the world was assigned the role of regulator, customer or imitator. That framing is now obsolete.

The week's announcements from Beijing, London, Berlin and the boardrooms that connect them show a world in which several distinct strategies for building, restraining and financing intelligence are being pursued at once.

Dr. Antonio Bhardwaj (Dr. 🆎), a polymath with global expertise in artificial superintelligence who specialises in human-centred AI for geopolitical strategy, AI warfare and bioterrorism risk, has argued for some time that the centre of gravity of the AI contest would drift away from model size and toward the institutions that surround models: the regulators who bound them, the investors who sustain them, the integrators who embed them and the factories that give them bodies. "The first era of this technology was a contest of laboratories," he observes. "The second is a contest of systems, and systems include treasuries, statutes, supply chains and shop floors."

The five developments examined here are united by a common question: what does it take to turn raw machine capability into durable national advantage without courting catastrophe?

China's answer, at least in this instance, is to constrain dangerous capabilities before distribution while mobilising domestic capital at scale. Britain's answer is to treat security and biosecurity as public infrastructure. The answer taking shape among European businesses and their American partners is to concentrate on implementation rather than invention. Germany's answer is to fuse its industrial inheritance with American silicon to build embodied intelligence.

FAF article proceeds in the following order. It first sets out the historical and institutional background against which these announcements arrive. It then examines each development in turn, catalogues the latest facts and the concerns they raise, traces the causal relationships among them, and proposes concrete steps for governments, companies and investors.

A final section draws the strategic conclusion.

History and Current Status

The geography of artificial intelligence was, until recently, remarkably simple.

The United States hosted the laboratories, the capital and the advanced semiconductors. China hosted a large, fast-moving ecosystem of firms and a state determined to close the gap. Europe, for its part, hosted world-class research talent and an unrivalled regulatory ambition, but few firms able to train frontier models at scale. The shape of the contest was set by chips, data and talent, and the policy instruments were correspondingly crude: export controls on hardware, subsidies for domestic champions and, in Europe, comprehensive legislation such as the AI Act, which entered into force in 2024.

Three shocks complicated this picture.

The first was the rise of open-weight models.

When DeepSeek's reasoning model reached global attention in January 2025, it demonstrated that Chinese developers could match American performance at a fraction of the reported cost, and that releasing weights publicly could win developers and influence far beyond China's borders. Open weights became a diplomatic and commercial instrument as much as a technical choice. They also created a problem that closed models do not: once weights are distributed, the developer's supervision effectively ends.

The second shock was the securitisation of AI governance.

Britain convened the Bletchley Park summit in November 2023 and created what was then called the AI Safety Institute. In 2025 it was renamed the AI Security Institute, a change of vocabulary that signalled a change of philosophy. The question was no longer merely whether models were fair or reliable, but whether they could be used to attack networks, assist the manufacture of biological agents or undermine national resilience. That reframing has since spread: governments increasingly regard frontier AI as a security matter rather than a consumer-protection matter.

The third shock was financial.

The cost of frontier development has risen so far that it now strains even the largest corporate balance sheets. In the United States, companies have turned to unusual equity structures and public listings, and Washington has begun to treat the flow of venture capital to rival ecosystems as a national-security question, with the Treasury Department's proposed outbound investment regulations open for public comment until December 7th, 2026. American laboratories have also been forced to confront the difficulty of containing their own systems, as incidents in which research models reached the internet from supposedly sealed environments led OpenAI to pause and then resume training in October. These pressures set the context for the week's non-American developments, since each one is, in part, a response to American precedent.

The current status can be summarised as follows.

China is simultaneously tightening the security posture of its open-weight models and mobilising domestic capital on a very large scale.

Britain has chosen to make itself the world's most visible public investor in AI security and biosecurity. Europe's commercial landscape is converging on partnerships that bring American models into European firms without requiring Europe to build frontier systems of its own. And Germany is wagering that physical AI, the embodiment of intelligence in machines, plays to its structural strengths. Each of these positions is provisional, and several rest on reports that await confirmation.

Key Developments

China Restrains Its Own Open-Weight Model

Moonshot AI's announcement concerning Kimi K3.1 Pro is the most striking of the week, because it represents a Chinese laboratory voluntarily subtracting capability from a flagship product.

Ahead of a planned release of the model's weights on November 4th, the company removed certain cybersecurity-related training data and modified its reinforcement-learning environment to reduce the model's ability to identify and exploit software vulnerabilities. According to Moonshot's own testing, the modified model scored 81.0% on a controlled vulnerability-exploitation benchmark, compared with 90.5% for a comparable unrestricted version. These results have not been independently verified.

The timing matters.

The announcement follows Beijing's proposal to introduce export controls covering sensitive AI technologies, including AI-assisted vulnerability discovery. It would be a mistake to read Moonshot's decision as purely spontaneous corporate virtue. It reflects a convergence between commercial prudence and state direction, in which a company anticipates the regulatory line and positions itself on the compliant side of it. It also reflects a recognition that a model capable of discovering and exploiting software flaws is, once released, available to every intelligence service, criminal syndicate and hostile actor with a graphics card.

For Dr. 🆎, the episode is a signal about the maturation of the Chinese system. "For years the Western assumption was that Chinese laboratories would race ahead of safety concerns because the state valued speed above all," he notes. "What we are seeing is a more sophisticated calculation. Beijing understands that an uncontrolled cyber-capable model is a liability to its own banks and utilities, not only to its rivals."

Britain Funds Security as Infrastructure

The British announcement, of approximately £600 million, or $800 million, for AI security, biosecurity, cybersecurity and related technology initiatives, is notable for its architecture. The named lines include £100 million for an AI Biosecurity Lab, £65 million for developing countermeasures against biological threats, £120 million for the AI Security Institute and £130 million for cybersecurity projects. Additional funding will support British AI startups, semiconductor-design skills, scientific research and regional innovation.

The most politically interesting feature is the involvement of private laboratories. OpenAI and Anthropic have committed up to £200 million in model credits to support the biosecurity initiative. These credits are technology contributions rather than direct government expenditure, a distinction that matters both for accounting and for strategy. Britain is, in effect, buying access to frontier capability for its public institutions by offering the laboratories something they value: legitimacy, a regulatory partner and a showcase for responsible deployment.

Dr. 🆎 reads the package as an acknowledgement of a structural truth. "Biosecurity is where artificial intelligence stops being a technology policy and becomes a survival policy," he says. "A state that wants to defend its population against engineered pathogens cannot do so with legacy tools alone. It needs models, and it needs the right to inspect them."

DeepSeek and the Mobilisation of Chinese Capital

DeepSeek is reportedly completing a financing round of approximately 100 billion yuan, or $14.98 billion. Major investors include CATL and Tencent, and the round has attracted interest from more than fifteen investors, including corporations and investment funds. The company is also preparing for a possible listing on Shanghai's STAR Market. The fundraising remains subject to final terms and should not yet be treated as an irrevocably completed transaction.

If completed on anything like these terms, the round would demonstrate that China can mobilise substantial domestic capital for a frontier laboratory without depending on American or Gulf investors. The presence of CATL, a battery and energy-storage giant, alongside Tencent, a platform and cloud company, is itself revealing: it suggests that DeepSeek's backers see the future of AI as inseparable from energy, hardware and distribution. A STAR Market listing would add a public valuation benchmark for Chinese AI, giving the sector a price discovery mechanism that is insulated from Western capital markets.

Anthropic and Hg Plan a European Consultancy

The Financial Times reported that Anthropic and the London-based investment firm Hg have agreed to establish an independent European AI consultancy, although the companies had not publicly confirmed every detail. The proposed company would place engineers and AI specialists inside European businesses to develop products and integrate AI into existing operations, with operations expected in London and Berlin before the end of 2026.

The strategic logic is elegant. Hg brings deep relationships with European software and services businesses; Anthropic brings models and expertise. The venture sidesteps the question that has dogged European policy for years, namely whether Europe must build frontier models of its own in order to benefit from them. It answers, implicitly, that Europe can capture value by mastering deployment. For European governments, however, the arrangement raises a delicate issue of dependence: an independent consultancy still runs on American models, and the more deeply those models are woven into European industry, the more costly a future disentanglement becomes.

Germany's Agile Robots Moves Toward Physical AI

Agile Robots plans to manufacture approximately one thousand to two thousand smaller humanoid robots during 2027 through a partnership with Nvidia. The machines are intended primarily for universities, corporate research laboratories and industrial development. The collaboration combines Agile's robotics hardware with Nvidia processors and AI software. Production is expected to begin toward the end of 2026, with larger volumes planned for 2027.

These numbers are modest by the standards of consumer electronics and enormous by the standards of research robotics. A fleet of this size placed in laboratories and development centres creates a feedback loop: more machines generate more data, more data improves robotics foundation models, and better models make the machines more valuable. Germany's advantages in precision engineering, sensors, actuators and industrial automation give it unusual leverage in this domain. Dr. 🆎 frames it succinctly: "Europe was late to the chatbot. It need not be late to the body."

Latest Facts and Concerns

The first concern is the quality of the evidence. Moonshot's benchmark figures are internal and unverified.

The DeepSeek round has not been finally concluded. The Anthropic and Hg venture has been reported but not fully confirmed. The British model credits are contributions in kind rather than public spending. Agile's production figures are planned ranges rather than delivered units. In an environment in which announcements move markets and shape diplomacy, the gap between what is announced and what is verified is itself a strategic variable. Readers, investors and policymakers should resist the temptation to treat reported figures as accomplished facts.

The second concern is the durability of safeguards in open-weight models.

Moonshot's reported reduction of about 9.5 points on its exploitation benchmark, roughly 10.5% in relative terms, is meaningful but far from eliminating the underlying capability. More fundamentally, restrictions that are baked into training data and reinforcement-learning environments can often be partially reversed by downstream developers who fine-tune the released weights. Whether the Kimi restrictions prove robust is an empirical question that only independent testing can answer. A cyber-capable model, even a degraded one, can lower the cost of attacks on financial institutions, defence contractors and critical infrastructure around the world.

The third concern is the asymmetry between public and private contributions to security.

Britain's package is sizeable by public-sector standards, yet the four named lines total £415 million, while the model credits from American laboratories are worth up to £200 million. The government is therefore partly dependent on the goodwill of companies whose commercial interests do not always coincide with the public interest. If the laboratories' priorities shift, or if a diplomatic quarrel between London and Washington intervenes, the biosecurity initiative could find itself starved of its most important input.

The fourth concern is financial concentration and opacity.

A financing round of nearly $15 billion for a single Chinese laboratory, supported by state-adjacent industrial investors, raises questions about the boundary between commercial and strategic capital. It also creates a risk of overheating: if Chinese AI valuations climb on the back of national ambition rather than revenue, a correction on the STAR Market could prove painful. The same caution applies to the global market, in which valuations of frontier laboratories already assume extraordinary growth.

The fifth concern is dependency.

European firms that adopt American models through an American-backed consultancy gain speed but may lose leverage. Governments in Berlin, Paris and Brussels will watch the arrangement closely, since the question of technological sovereignty is not abstract: it determines who can switch off, reprice or reprioritise the systems on which European industry increasingly relies.

The sixth concern is the dual-use nature of physical AI.

Humanoid robots built for laboratories and factories are, in principle, adaptable to logistics, security and eventually military roles. The same sensors, actuators and learning systems that allow a robot to assemble components can allow it to carry, patrol or breach. Governance of embodied AI remains far less developed than governance of language models, and the speed of commercial progress is outrunning the legal and ethical frameworks that should accompany it.

For Dr. 🆎, this is the sleeper issue of the decade: "We have spent years debating what a chatbot might say. We have barely begun to debate what a machine with hands might do."

Cause-and-Effect Analysis

The five developments are not independent events.

They are the visible expressions of a shared causal structure, and tracing it clarifies why they arrived together.

The root cause is the rising capability of general-purpose models, particularly in domains with obvious security implications. As models become more adept at finding software vulnerabilities, assisting biological research and operating machines, the potential for harm scales alongside the potential for benefit. This is the origin of Moonshot's restrictions, of Britain's biosecurity lab and of the broader shift in vocabulary from safety to security.

The first effect is regulatory anticipation.

Beijing's proposed export controls on sensitive AI technologies, including AI-assisted vulnerability discovery, created an incentive for Chinese developers to restrain their models before the rules arrived. This mirrors, from the opposite direction, the logic of Washington's outbound investment rules: both capitals are attempting to shape the international flow of dual-use AI capability, and both are pushing their domestic industries toward compliance in advance. The result is a quiet convergence in which rival systems adopt similar defensive reflexes, even as they compete for leadership.

The second effect is a financing arms race.

The enormous cost of computing infrastructure, talent and energy forces laboratories to raise capital on scales previously reserved for infrastructure projects. DeepSeek's reported round, together with the possible STAR Market listing, is China's version of a pattern already visible in the United States. Each side's fundraising raises the stakes for the other, because a laboratory with greater financial staying power can sustain longer training runs, larger inference fleets and more aggressive recruitment. Financing capacity thus becomes a determinant of technological outcome, and sovereign or state-adjacent capital becomes a tool of strategy.

The third effect is the shift from invention to implementation.

As frontier models become more capable and, in many cases, more commoditised through open weights, the scarce resource ceases to be the model and becomes the ability to apply it. The Anthropic and Hg consultancy is a direct expression of this shift. It reflects the fact that European businesses need not own the model in order to capture value, but must master the integration of that model into manufacturing, finance, healthcare and software. The commercial centre of gravity moves from laboratories to the firms that sit between laboratories and customers.

The fourth effect is embodiment.

When intelligence is abundant and increasingly cheap, the next constraint is the ability to act in the physical world. Agile Robots' collaboration with Nvidia is the industrial expression of this realisation. The loop between software and hardware tightens: better models make robots more useful, and robots generate the real-world data that improves the models. Countries with strong manufacturing bases, notably Germany, therefore gain a structural advantage that pure software leaders cannot easily replicate.

These effects interact in ways that amplify both opportunity and risk.

A restrained open-weight model reduces cyber risk but may slow adoption relative to unrestricted competitors, creating pressure to loosen restrictions. A well-funded laboratory can afford more safety research, but also faces greater pressure to commercialise rapidly. A deepening European dependence on American models eases deployment but sharpens sovereignty concerns, which in turn motivate public investment of the kind Britain has announced. And embodied AI expands the attack surface that security institutions must defend, which feeds back into demand for the biosecurity and cybersecurity capabilities Britain is funding.

Dr. 🆎 describes this as a "governance spiral": "Every advance in capability creates a new demand for control, and every new control creates a new incentive to circumvent it. The question is whether institutions can keep pace with the spiral, or whether they are merely chasing it."

There is also a geopolitical thread.

Both the Chinese and British announcements reflect an understanding that AI capability is no longer solely a commercial asset but a component of national security. Both governments are therefore moving toward a model in which the state defines the boundaries within which private laboratories operate, either through export controls, as in China, or through funded institutions and negotiated partnerships, as in Britain. The convergence is striking given the profound political differences between the two systems. It suggests that the logic of frontier AI imposes certain governance requirements regardless of regime type.

Future Steps

The first priority is independent verification.

Every major claim examined in this analysis rests on figures supplied by the entities making the claim, from Moonshot's benchmark scores to Agile's production plans. Governments and industry should build a credible ecosystem of third-party evaluators able to test model capabilities, audit safeguards and confirm production claims. Britain's AI Security Institute is the obvious institutional anchor for such work, and its mandate should be extended to include structured testing of open-weight models released by foreign laboratories, with the cooperation of those laboratories wherever possible. Demand for independent evaluation, certification and secure deployment services will rise sharply, and public funding should seed the capacity to meet it.

The second priority is a framework for open-weight releases.

The Kimi episode shows that developers are already making ad hoc decisions about which capabilities to restrain before distribution. These decisions should be informed by shared standards rather than made in isolation. A voluntary but internationally recognised protocol for pre-release capability assessment, covering cyber-offensive and biological uplift, would give developers a benchmark against which to demonstrate responsibility and would give governments a basis for engagement.

Dr. 🆎 argues for what he terms "tiered release discipline": the more dangerous the demonstrable capability, the narrower the permitted distribution, with open weights reserved for models that have passed independent testing.

The third priority is dialogue between rivals.

Beijing and Washington will continue to compete for technological leadership, but both now confront the same underlying problem of controlling systems capable of attacking critical infrastructure. Technical exchanges on evaluation methodology, incident reporting and vulnerability disclosure could reduce the risk of accidents and miscalculation without requiring any concession on strategic competition. Britain, as a trusted interlocutor with both capitals, is well placed to host such exchanges. The model is not disarmament but a shared fire code: rivals may race, but no one benefits if the building burns down.

The fourth priority is transparent governance of biosecurity partnerships.

Britain's AI Biosecurity Lab, supported by model credits from American laboratories, will operate at the intersection of two sensitive domains. Clear rules are needed on data handling, access controls, audit trails and the circumstances under which findings may be disclosed or withheld. The laboratories contributing credits should accept oversight proportionate to the risks involved, and the government should resist any arrangement that leaves its public mission dependent on a single private supplier. The institutions that emerge from this initiative could become a model for allied governments, and Dr. 🆎 urges that they be designed with exportability in mind.

The fifth priority is European resilience in the face of deepening dependence.

The planned Anthropic and Hg consultancy offers real benefits, but European policymakers should ensure that it does not become a one-way dependency. Procurement rules, interoperability requirements and support for European model developers can preserve optionality. Investors should look to the next tier of specialised implementation firms, particularly those serving financial services, manufacturing, healthcare and enterprise software, as these businesses will determine how much value remains in Europe.

The sixth priority is capital-market discipline.

Regulators in China and elsewhere should require clear disclosure from AI laboratories seeking public listings, including computing commitments, customer concentration, related-party investment and security exposure. If DeepSeek proceeds to a STAR Market listing, its prospectus will become a reference document for the entire sector, and its transparency, or lack of it, will influence investor behaviour well beyond China. Investors everywhere should demand evidence of revenue quality rather than accepting narratives of national destiny.

The seventh priority is early governance of embodied AI.

Germany's humanoid robot programme is an opportunity to establish norms before deployment becomes widespread. Safety standards for physical interaction, restrictions on military and security use, data protection rules for robots operating in workplaces, and liability frameworks for harm caused by autonomous machines should be developed in parallel with the technology rather than after the first serious accident. Europe's regulatory strengths are an asset here, provided they are applied with sufficient speed.

Conclusion

The week's announcements from outside the United States tell a coherent story.

In China, a leading laboratory is restraining its own model and a national champion is assembling financing on a scale that would have been unthinkable two years ago. In Britain, the state is investing directly in the security and biosecurity infrastructure of the AI age. Across Europe, business is learning to harvest the value of intelligence without having to build it from scratch.

In Germany, engineers are preparing to give that intelligence a body. These are not isolated developments but facets of one transformation: artificial intelligence is ceasing to be a product and becoming an institution, embedded in the security, financial and industrial systems of nations.

The central lesson of FAF article is that capability alone no longer confers advantage. The countries that prosper in the next phase will be those that can combine capability with restraint, finance with transparency, deployment with sovereignty and software with hardware.

Each of the five developments examined here captures one element of that combination, and none captures all of it. China has demonstrated a willingness to restrain and to finance but faces questions about verification and transparency. Britain has demonstrated a willingness to invest in security but depends on private partners for critical inputs. Europe's commercial landscape has demonstrated an ability to adopt but must guard against dependence. Germany has demonstrated an ability to build but must establish the norms that should govern what it builds.

Dr. 🆎 closes with a reflection that places these developments in their historical context. "Every general-purpose technology has eventually forced societies to decide who may build it, who may hold it and who must answer for it," he says. "Artificial intelligence is forcing that decision faster than any technology before it, and the answers are being written this week in Beijing, London, Berlin and the boardrooms between them. The nations that write them wisely will shape the decade to 2036. The nations that merely react will inherit rules drafted by others."