A Glimmer of the Future: The Open-Source Cold War and What If China Wins AI by Giving It Away?

For most of the artificial-intelligence revolution, we have been keeping score in a fairly obvious way. Who has the smartest model? OpenAI releases something new, Google responds with Gemini, Anthropic advances Claude, DeepSeek surprises everybody from China and another collection of benchmarks appears telling us which machine is marginally better at mathematics, coding, reasoning or understanding some obscure collection of questions designed to measure intelligence. The assumption underneath all of this is perfectly reasonable: the country which produces the world’s most capable artificial intelligence wins the AI race. Mark Zuckerberg has just suggested, perhaps inadvertently, that we may be measuring the wrong race.

Meta has launched a new open-weight AI model called Muse Glimmer, designed to perform agentic tasks while being sufficiently compact to run on a single graphics processor, and Zuckerberg has accompanied it with an unusually geopolitical argument. America, he says, needs to reduce the obstacles facing open AI because Chinese developers are increasingly dominating that part of the ecosystem. His concern is not simply that DeepSeek, Alibaba’s Qwen or other Chinese models might become better than their American equivalents. It is that they might become more widely used. Washington appears to be listening, at least partly. The Trump administration has just told leading AI companies that open-weight models will not be subjected to the voluntary safety-testing regime being discussed for advanced systems, suggesting that American policymakers increasingly see openness not merely as a technical or philosophical question but as part of the strategic competition with China. 

Before going further, there is a slightly annoying but important distinction. We routinely call these systems open-source AI, including in the title of this piece, yet many are more accurately described as open-weight. Their developers release the trained weights which allow somebody else to download, modify, fine-tune and operate the model, without necessarily revealing everything about the training data or the process which created it. That distinction matters enormously to researchers debating what genuine openness means, and China’s claims to openness deserve exactly the same scrutiny as America’s. For geopolitics, however, the crucial characteristic is simpler. If I can download your model and operate it on my own computers, you no longer control my continuing access to it. 

That changes the competition.

Imagine America produces an artificial intelligence which, on some imaginary composite measure of capability, scores 100. China produces one which scores 90. The American model is demonstrably better, so America appears to have won. Now add one additional fact. To use the American model I must access servers controlled by an American corporation, accept its terms, pay whatever price it sets and remain exposed to changes in American regulation, export controls and foreign policy. The Chinese model can be downloaded, modified, translated, fine-tuned on my own information and operated indefinitely inside my own country. Suddenly the ten-point performance gap may not settle the argument. If I am a student wanting the best possible answer to a difficult mathematics problem, perhaps I choose the American model. If I am the government of Indonesia deciding what artificial intelligence will eventually process sensitive information across ten ministries, the calculation becomes considerably more complicated.

This is where my 20% Rule reappears in a slightly different form. The rule is not really about twenty per cent. It is about the difference between technological superiority and strategic sufficiency. Russia does not need Rassvet to become better than Starlink if Rassvet becomes good enough that Russia cannot be disconnected from satellite communications. China does not need immediately to reproduce ASML’s finest lithography equipment if domestic machines become good enough to manufacture the strategically essential chips which Western export restrictions are intended to deny it. China may similarly not need the world’s best artificial intelligence if Chinese models become good enough that governments, companies and developers around the world prefer owning them to renting something marginally better from somebody else.

Capability is one dimension of usefulness. Control is another.

Once you think about AI that way, Alibaba’s Qwen becomes at least as interesting geopolitically as DeepSeek. DeepSeek captured Western attention because it appeared to demonstrate that Chinese researchers could produce extremely capable models more efficiently than many people thought possible. Qwen has been doing something quieter and potentially more consequential: spreading. An April 2026 study of roughly 1,500 major open models concluded that Chinese models overtook American counterparts in open-model adoption during the summer of 2025 and subsequently widened the gap. Another large study examining billions of model downloads found the same broad shift towards Chinese industry, particularly Qwen and DeepSeek. Alibaba said in January that Qwen had already accumulated around 700 million downloads on Hugging Face and spawned more than 180,000 derivative models. Those numbers should be treated with some caution because Alibaba itself supplied part of the underlying claim, but the broader independent evidence that Chinese open models have surged is compelling. 

This is where the history of computing becomes useful because technological dominance has rarely depended simply upon possessing the best individual product. Platforms matter. Standards matter. Developer ecosystems matter. Linux became enormously important not because one company persuaded everybody to purchase Linux subscriptions but because other people could build things upon it. Android became globally dominant without Google manufacturing most Android phones because Samsung, Xiaomi, Oppo and hundreds of other companies could build products around Google’s operating system. The internet itself became overwhelmingly shaped by American technology partly because American standards, software, programming languages, universities and companies became the infrastructure through which everybody else participated in the digital revolution.

China may have learned the lesson.

Giving away something which cost hundreds of millions or eventually billions of dollars to develop sounds economically irrational if the objective is to sell that particular thing. It looks considerably more rational if the objective is to make your technology the foundation upon which everybody else builds. A university downloads Qwen and trains researchers on it. A startup fine-tunes it for accounting. A hospital adapts it for medical administration. A government modifies it for local law. A software company creates tools around it. Thousands of developers discover bugs, improve deployment techniques, create quantised versions capable of running on cheaper hardware and publish instructions for everybody else. The model ceases to be merely a product and starts becoming infrastructure.

America may therefore be trying to sell the world’s best AI while China attempts to become the operating system underneath everybody else’s.

The Global South is where that possibility becomes geopolitical. Imagine you are Kenya, Indonesia, Brazil, Vietnam or Saudi Arabia and believe, quite reasonably, that artificial intelligence will become critical national infrastructure. You can access an extremely capable American proprietary model through somebody else’s cloud. Every query potentially passes through infrastructure outside your country. The provider controls updates, pricing and continued access. American law applies somewhere in the chain and future American governments could conceivably impose restrictions which today’s government has no intention of imposing. Alternatively, you can take a slightly less capable open-weight model, install it inside your own data centre, fine-tune it using your own government documents, teach it your laws and languages, restrict access however you choose and continue running it even if relations with the country which created it deteriorate.

This does not require anti-Americanism. It requires risk management.

Europe itself increasingly talks about sovereign AI for exactly this reason. Governments everywhere are discovering that cloud computing, semiconductor fabrication, satellite communications, payment networks and artificial intelligence are no longer merely commercial services. They are strategic infrastructure, and strategic infrastructure creates vulnerability when somebody else controls it. The lesson Russia learned through Starlink applies surprisingly neatly to AI. The system somebody else cannot switch off can be strategically more valuable than the superior system they can.

Chinese open-weight AI therefore has the potential to become something resembling a digital Belt and Road. China’s original Belt and Road strategy recognised that infrastructure creates relationships. Selling somebody a television gives you a customer; financing the port through which their exports travel gives you influence. Railways, telecommunications systems, electricity grids and ports embed one country’s technology inside another country’s economy. Artificial intelligence could become an even subtler form of infrastructure because there need not be a Chinese-owned data centre or a Chinese engineer anywhere in sight. The Kenyan government could run its model in Nairobi. The Brazilian company could operate its system in São Paulo. The Indonesian university could fine-tune it in Jakarta. China simply supplied the technological foundations.

That may actually make open AI more geopolitically powerful than traditional infrastructure because adoption does not look like dependence. A country using Huawei equipment knows it is using Chinese telecommunications equipment. A country which takes Qwen, fine-tunes it on millions of words of Indonesian, adds local knowledge, changes the safeguards, gives it a local name and deploys it on domestically owned computers may quite reasonably regard the resulting system as Indonesian AI. Chinese influence has become almost invisible because China supplied the foundations rather than the finished building.

Language makes this particularly interesting. English dominates much of the internet and therefore much of the data upon which the first generation of large language models was trained, yet most of humanity does not speak English as a first language. Qwen’s family has deliberately expanded multilingual capability, while open weights allow local developers to adapt models far beyond whatever languages the original laboratory prioritised. A Vietnamese university does not have to wait for an American company to decide that Vietnamese deserves greater investment. It can take an existing model and work on the problem itself. An Arabic government can specialise a model around legal or administrative Arabic. An African startup can optimise one for Swahili without asking Alibaba’s permission every time it wants to change something.

This is where giving AI away begins to resemble soft power. During the original Cold War, America’s influence came from much more than aircraft carriers and nuclear weapons. People wanted American products, attended American universities, listened to American music, learned English, adopted American management practices and built companies using American technologies. Influence was powerful precisely because much of it was voluntary. China does not need a Nigerian programmer to admire the Chinese political system for that programmer to build something using Qwen. Brazil does not need to support Chinese foreign policy to find DeepSeek useful. Indonesia does not need to join a Chinese geopolitical bloc to decide that a locally controlled Chinese-origin model offers greater sovereignty than dependence upon a closed American API.

The technology can spread without the ideology.

That is potentially much more powerful than trying to persuade everybody to choose sides.

BRICS fits naturally into this story, although not because I expect Xi Jinping to announce some grand BRICS artificial-intelligence model. That would almost miss the point. BRICS countries repeatedly discuss technological sovereignty, alternatives to Western-controlled financial infrastructure and the creation of additional choices within the international system. China does not need BRICS to adopt a Chinese AI. It needs Brazilian companies, Indian researchers, South African universities, Emirati data centres and other institutions across the expanded grouping to discover that Chinese-origin open models are useful foundations for their own systems. Nobody signs a treaty. Nobody formally defects from the American technological ecosystem. Another dependency simply disappears.

That is how multipolarity often develops in practice. It does not require countries to replace America with China. It requires them to acquire enough alternatives that neither America nor China possesses complete leverage over them.

There is also a fascinating economic reason China may be unusually comfortable with giving models away. OpenAI and Anthropic need eventually to earn enormous amounts of money from artificial intelligence because providing access to intelligence is central to their business models. Google has a much broader ecosystem but still has powerful incentives to monetise Gemini. Alibaba occupies a rather different position. If Qwen becomes ubiquitous, Alibaba can potentially benefit through cloud computing, enterprise services and the enormous ecosystem surrounding deployment even if the model weights themselves are free. Huawei benefits if Chinese models increasingly run on Huawei accelerators. Chinese software companies benefit from tools built around Chinese architectures. Chinese universities acquire expertise. Chinese data centres acquire workloads.

The relevant economic unit may therefore not be the model.

It may be the ecosystem.

China has used variations of this strategy in physical industries. Solar panels became extraordinarily cheap, destroying margins for many manufacturers while leaving China with overwhelming supply-chain dominance. Ferocious domestic competition has driven down EV prices while creating an industrial ecosystem Western manufacturers now struggle to match. Batteries became cheaper while Chinese companies established enormous positions throughout the value chain. The Chinese state does not necessarily require every participant to earn monopoly profits if the overall result is national industrial capability.

Open AI could become the software version of the same strategy. Let intelligence become cheap. Let models proliferate. Let thousands of companies adapt them. Capture value in the chips, clouds, applications, services and expertise surrounding them.

The irony is that American attempts to slow China’s AI development may have helped make this strategy more attractive. Washington restricted Chinese access to the most advanced Nvidia chips because frontier AI requires extraordinary quantities of computing power. The logic is straightforward and the restrictions have undoubtedly imposed real costs on Chinese developers. Yet technological competition is dynamic. If compute is scarce, efficiency becomes unusually valuable. Chinese laboratories acquired an enormous incentive to find ways of extracting more intelligence from less hardware, while developers focused heavily on quantisation, mixture-of-experts architectures and other techniques which reduce the resources required to deploy models. The recent open-model ecosystem shows rapid growth in exactly these efficiency-oriented approaches. 

DeepSeek’s psychological impact came partly from demonstrating that the relationship between enormous compute expenditure and high capability was not quite as fixed as many Western investors had assumed. Restrictions intended to increase the cost of Chinese AI also increased the reward for anybody who could make AI cheaper. That does not mean American semiconductor restrictions were pointless; preventing China from obtaining the world’s most advanced chips still matters enormously. It means sanctions and export controls generate adaptation.

The weapon creates the incentive to escape the weapon.

This is the same pattern we encountered with Chinese lithography. Restrict access to ASML and domestic lithography becomes strategically priceless. Restrict access to advanced Nvidia chips and efficient Chinese AI becomes more valuable. Restrict Russia’s ability to rely upon Starlink and Rassvet becomes more valuable. Freeze a country’s reserves or threaten its access to dollar payment systems and alternative financial mechanisms become more valuable. Western technological and financial chokepoints are enormously powerful precisely because they work, yet every successful use of a chokepoint increases the incentive for everybody exposed to it to construct another route.

Open-weight AI may be the route around one of the next great chokepoints.

This makes Zuckerberg’s intervention especially interesting because Meta is structurally better positioned than most American companies to respond. Meta earns the overwhelming majority of its money from advertising rather than charging people directly for access to AI models. Zuckerberg can therefore treat advanced models partly as infrastructure which makes the wider Meta ecosystem more useful. He does not necessarily need to charge every developer every time the model thinks. That puts Meta, strangely enough, in a position somewhat analogous to Alibaba. Both companies can potentially benefit enormously from an AI ecosystem even if the underlying intelligence becomes cheaper.

Muse Glimmer is consequently interesting less because it suddenly makes Meta the unquestioned leader in artificial intelligence than because of what Zuckerberg is saying alongside it. He is explicitly framing open AI as part of American technological leadership and warning that excessive restrictions could hand the open ecosystem to Chinese competitors. Meta says Glimmer can perform useful agentic tasks on a single GPU, precisely the kind of local deployment which reduces dependence upon giant centralised cloud systems. 

There is a wonderful ideological reversal buried inside this. For decades America associated itself with technological openness while China was associated with control. The internet was supposed to carry American ideas into closed societies. Open standards, open research and decentralised innovation were part of the mythology, and often the reality, of Silicon Valley.

Artificial intelligence could invert the perception.

Chinese companies increasingly release powerful models which developers around the world can download.

America’s most celebrated frontier systems increasingly sit behind corporate gates.

That does not mean China has suddenly become the global champion of information freedom. Chinese models operate within a domestic political system characterised by extensive censorship and information control, while researchers correctly point out that many supposedly open Chinese models withhold important training information. American open-weight models can likewise fall well short of genuine open-source standards. The paradox remains striking. The country which built the Great Firewall could nevertheless become one of the principal suppliers of AI that foreigners can download and control themselves.

There is, of course, a very large reason America might hesitate before copying the strategy completely. The characteristic which makes open-weight AI geopolitically attractive is precisely the characteristic which makes it potentially dangerous.

Nobody can switch it off.

Once powerful model weights have been released, the original developer loses much of its control. Safeguards can potentially be removed. Governments can modify the system. Militaries can fine-tune it. Intelligence services can operate it privately. Cybercriminals can adapt it. A sanctioned state can download a model today and continue using it tomorrow even if the company which created it subsequently decides that access should be prohibited. A closed AI service can monitor usage, update safeguards and terminate accounts. An open-weight model sitting on somebody else’s computer cannot easily be recalled.

Washington therefore faces a genuinely difficult strategic choice. Regulate the release of increasingly powerful American open-weight models aggressively and Chinese alternatives may become the default global open ecosystem. Allow American companies to release them and accept that some extremely capable technology will inevitably reach actors the United States would prefer not to empower. The Trump administration’s decision not to impose the proposed voluntary safety testing on open-weight systems suggests that, for the moment, Washington is giving considerable weight to the competitive danger of falling behind. 

The policy problem will become much harder as the models improve. Giving away a chatbot capable of drafting emails is one thing. Giving away a model capable of sophisticated cyber operations, autonomous scientific research or controlling increasingly capable robots is something quite different. Our DeepSeek–Unitree love affair suddenly appears again. An open Chinese AI model can be downloaded by somebody else and potentially connected to a physical machine. The geopolitical virtue of decentralisation and the security danger of decentralisation are two sides of exactly the same technology.

This may ultimately produce two competing philosophies of artificial intelligence. One treats frontier AI rather like nuclear technology: extremely powerful, potentially dangerous and therefore concentrated inside a relatively small number of heavily controlled institutions. The other treats AI more like software: something which becomes more valuable when millions of people can inspect, modify, distribute and build upon it. America currently contains powerful advocates of both positions. China, interestingly, can pursue both simultaneously, keeping some of its most advanced systems controlled while flooding the international ecosystem with increasingly capable open models.

The strategic brilliance, if it works, is that China does not actually have to persuade the world that Chinese AI is better.

It only has to make Chinese AI useful enough to become difficult to avoid.

Suppose the world’s best closed American model scores 100 and the best downloadable Chinese model scores 90. A London lawyer seeking the most sophisticated reasoning might happily pay for the 100. A Brazilian government wanting to process confidential tax information on domestic servers may prefer the 90. An Indonesian university wanting to modify a model for Bahasa Indonesia may prefer the 90. An Indian company worried about API costs across hundreds of millions of interactions may prefer the 90. An Iranian institution which knows American access could disappear tomorrow will almost certainly prefer the 90.

Suddenly the ten-point capability gap tells us very little about geopolitical influence.

This is why the artificial-intelligence race may eventually resemble the history of computing more than the history of luxury goods. The finest product does not automatically become the infrastructure. IBM once appeared unassailable. Microsoft understood the importance of platforms. Linux became embedded across global computing precisely because nobody could monopolise it. Android conquered huge parts of mobile computing by allowing thousands of manufacturers and developers to build around it. Technologies become extraordinarily difficult to displace once entire ecosystems grow around them.

AI may follow the same pattern.

The most important model might not ultimately be the smartest model.

It may be the model upon which the largest number of other things are built.

Once universities teach it, companies optimise software for it, governments adapt it, hardware manufacturers design around it and millions of developers become familiar with it, switching becomes expensive even when something technically superior exists. Network effects begin operating around intelligence itself.

This also changes how we should think about American and Chinese technological power. America remains extraordinarily strong at creating frontier technologies and companies capable of capturing enormous profits from them. China has repeatedly demonstrated an ability to take technologies, drive down their costs, scale production and spread them through enormous ecosystems. In solar panels and batteries that process involved physical manufacturing. In artificial intelligence it could involve the manufacture of something stranger: abundant, customisable intelligence.

China’s objective does not need to be destroying OpenAI, Anthropic or Google. Those companies may remain richer and their frontier models may remain better. The strategic objective could simply be ensuring that whenever a university in Africa, a startup in Southeast Asia, a manufacturer in Latin America or a government in the Middle East decides it wants artificial intelligence it actually owns, the easiest answer begins with a Chinese model.

That would be technological influence on an extraordinary scale.

It would also be a very different kind of Cold War. The Soviet Union attempted to compete with the United States through military power, ideology and political alliances. China is increasingly capable of competing through infrastructure and optionality. A port does not ask you to become communist before unloading your ship. A solar panel does not care how you vote. A Qwen model does not require you to support Beijing’s position on Taiwan before you download its weights.

That makes technological ecosystems unusually effective instruments of influence because they can cross political boundaries which alliances cannot.

The Open-Source Cold War may therefore not produce two neat blocs with American AI on one side and Chinese AI on the other. The world is likely to be messier. A country may use OpenAI for some applications, Gemini for others, a locally hosted Qwen derivative for government information and a domestic model built upon Chinese foundations for its own language. Companies will arbitrage price, capability, sovereignty and security. Governments will hedge. Developers will use whatever works.

China does not need exclusivity to win influence in that world.

It needs ubiquity.

That is why a relatively small model called Glimmer may offer a surprisingly useful glimpse of the future. Zuckerberg is effectively warning Washington that artificial-intelligence leadership cannot be measured solely by the intelligence of America’s most powerful proprietary model. Leadership also depends upon what everybody else is building with. America can possess the most impressive AI laboratories on Earth and still discover that much of the world has quietly standardised around models originating somewhere else.

We have spent the last few years assuming that artificial intelligence would concentrate power because training the best models requires extraordinary quantities of chips, electricity, data and money. Perhaps that remains true at the frontier. Open weights introduce a countervailing force because once the intelligence has been created, copies can spread at almost zero marginal cost. The laboratory may cost billions. The next copy costs almost nothing.

That is an extraordinary geopolitical characteristic.

China spent hundreds of billions constructing Belt and Road infrastructure because physical infrastructure is expensive. Artificial intelligence offers something radically different. Build the model once and potentially distribute it to millions of people. Every additional user can make the ecosystem larger without China financing another railway, port or power station.

The Belt and Road moved Chinese infrastructure around the world.

Open AI could move Chinese intelligence around the world.

That does not guarantee Chinese victory. America’s frontier laboratories remain formidable, Meta has clearly recognised the danger, American open models could regain ground and governments may eventually become much more cautious about deploying Chinese-origin systems in sensitive infrastructure. Questions about censorship, hidden biases, cybersecurity, supply chains and trust could all constrain adoption. Chinese companies themselves may also become less open if they discover that giving away increasingly expensive frontier research undermines their commercial interests.

The direction of travel nevertheless deserves far more attention than another benchmark showing that one American model scored three points higher than another Chinese model.

For most of the AI race, Washington has been asking how to ensure that the world’s most powerful artificial intelligence remains American.

China knows there is another way to win.

Build something ALMOST as good. Make it cheap. Let people change it. Let them call it their own….Then give it to everybody!