OpenAI’s GPT-5 is out; DeepSeek’s R2 is still pending. The gap is about more than clashing launch calendars. It signals a shift in how power is exercised in the artificial intelligence (AI) economy. The contest is no longer model versus model; it is about who controls the plumbing of the digital world – chips and data centers – and, increasingly, the rulebook that governs them.
Most American tech giants run “black-box” AI: closed weights behind APIs, sold as a service. The advantages – data flywheels, safety control, customer lock-in – are clear. Short of cutting-edge GPUs, Beijing is pushing the other way, toward open-weight software and rapid deployment. Chinese stacks are spreading across the developing world via a Digital Silk Road of cloud deals, telemedicine pilots, and smart-city projects. Yet the chip chokepoint remains. After failed runs on Huawei’s Ascend chips, DeepSeek reverted to Nvidia for training and used Huawei mainly for inference – an awkward compromise that underscores U.S. leverage.
Interdependence persists. In August, Washington struck a novel bargain with Nvidia and AMD: limited exports of China-compliant accelerators may resume, provided the firms remit 15 percent of China sales to the U.S. government. Think of it as constrained monetization – a toll to use the bridge without fully reopening the road. Chinese demand helps subsidize U.S. innovation, while frontier models remain largely ring-fenced.
A more subtle wrinkle is perception risk as a policy tool. Beijing has alleged “back-door” risks in China-bound H20 chips, claiming that devices could be tracked or remotely disabled. China’s government summoned Nvidia to explain; the firm denies any such capability. Whether it’s true or not matters less than the signal. Expect administrative bans for government and defense procurement and a harder push toward self-sufficiency – a closed-loop AI stack from software to silicon and back. The tactic both tightens pressure in negotiations with Washington and locks in China’s strategic pivot from dependence to domestic control.
China’s approach blends open weights where that catalyzes localization, along with tight controls in finance, mobility, and public services where the state demands it.
The Global AI Competition
In the global AI race, the field is stratified. At the top sit the “AI superpowers” – the United States and China – with end-to-end capability across technology, infrastructure, research, and commercial applications. Beneath them are “middle powers” such as Britain, France, and South Korea. These nations are strong in research and deployment but short of the full stack for frontier foundational models. At the base are “follower states,” whose structural constraints keep them from the technological frontier.
Two forces are propelling the race. First, as a dual-use technology, AI throws off headline-grabbing breakthroughs – DeepSeek’s R1 among them – that trigger “Sputnik moments,” stoke status anxiety, and unleash reactive investment and an arms race. Second, governance has become a front in its own right. Both superpowers and middle powers now shape rules through standards work, summitry, and specialized expertise.
Washington is pursuing frontier-first chokepoints to preserve technological primacy, restricting flows of advanced hardware, software, and know-how to China while bankrolling domestic champions and tightening supply-chain alliances. The aim is to control foundational inputs and assemble a tech bloc that is less exposed to Beijing.
Beijing’s counter-narrative casts China as a bridge-builder. Its AI Global Governance Action Plan stresses development, security, and “open cooperation.” At home, the state layers research funding, tech-transfer schemes, and regulatory sandboxes to accelerate deployment. Ecosystems in the Yangtze River Delta, the Greater Bay Area, and the Beijing–Tianjin–Hebei corridor churn out patents and market-ready applications. Abroad, a Digital Silk Road stitches Chinese AI into developing economies, embedding Chinese platforms and standards.
Hardware, Infrastructure, and People
Controls on frontier accelerators have slowed China’s training pipelines, but U.S. export controls let lesser chips through, with restrictions. China’s answer is parallelization: domestic accelerators plus a burst of data-center building that, by official tallies, now puts it second globally in installed compute.
The capex boom has gone global. Hyperscalers are pouring money into concrete, power, and cooling; governments re-label industrial policy as climate-friendly “digital infrastructure.” The United States worries about the grid; China about utilization. Both press on.
In terms of talent, the United States hosts most elite researchers and the highest-paid specialists. At the same time, Chinese nationals make up roughly half of the global AI research workforce, and China continues to produce more AI-relevant Ph.D.s than the United States.
Beijing is trying to keep its talent onshore with higher salaries and “mission-driven” projects. The state-led push creates a self-reinforcing national-project dynamic. It is using capital injections to de-risk private and SOE investment, while introducing offtake agreements that oblige state entities and tech giants to buy Chinese chips even if initially costlier or less capable. Yet a prolonged GPU shortage in China would risk pulling star researchers to Singapore, Toronto, or Abu Dhabi – and, where visas permit, Silicon Valley – reminding everyone that the geography of chips is fast becoming the geography of jobs.
The takeaway is that chips alone won’t win the AI race; ecosystems will. The U.S. leads at the frontier, but China’s strength is the system around its models: a vast base of early adopters, deep engineering benches, and large, centrally managed data-center build-outs that are cost-effective and increasingly low-carbon – and are increasingly being exported to the Global South. This mirrors China’s play in EVs and solar, where scale plus industrial ecosystems, not single products, has tipped the balance.
The Operating Systems of Governance
If algorithms and chips are the hardware, governance is the operating system. The United States favors a flexible patchwork of executive orders, agency guidance, and voluntary safety pledges, reinforced by federal procurement policies.
Brussels has legislated a comprehensive, risk-based regime: the AI Act entered into force on August 1, 2024, with obligations phasing in through 2025–26, overseen by a new AI Office and AI Board.
China runs a centralized model that yokes industrial acceleration to political control: strict on data and platforms, experimental in provincial sandboxes.
Japan and Australia speak the language of trust and human-centric design, edging toward interoperability.
These operating systems travel – through standards bodies, procurement clauses, cloud regions, and technical assistance – often faster than treaties.
The United States bets that innovation will outrun risk; the EU that rules will manufacture trust; China that supervision can deliver both development and control. For parts of the Global South, China’s speed, price, and state backing are appealing; for others, Europe’s assurances or the United States’ innovation pull will prove more persuasive.
Countries export rules as much as technology. The United States’ procurement-led guidance, the EU’s AI Act, and China’s centralized controls spread through contracts, standards bodies, and cloud footprints. How states align with – or hedge between – these rulebooks will set market access, compliance costs, and the portability of AI across borders. There is a window for middle powers and followers to shape interoperable baselines.
The Price of Fracture
Industrial policy is inherently political. Washington’s monetized containment and Beijing’s “back-door” claims (denied by Nvidia) function less as purely commercial or technical judgments than as tools for shaping negotiations, signaling resolve, and, if necessary, accelerating technological decoupling.
Restrictions on frontier AI chips are biting. The U.S. doctrine of monetized containment levies a toll at the chokepoint, complicates multinational supply chains, and normalizes fiscal instruments in tech policy. The measures have already slowed parts of China’s advanced-model training and, over time, will push Beijing toward self-reliance. This will split the landscape into two partly incompatible ecosystems. This plays out at a technological inflection point: the ICT era is plateauing while AI emerges as the next general-purpose technology. Its economic gains are not yet fully realized, but scale and dense adoption networks may give China an advantage in diffusion even if it lags at the frontier.
Fragmentation threatens the digital commons, especially for follower states. Shared safety standards, open research, and coordinated crisis-response mechanisms are at risk. Policing cross-border harms – from autonomous weapons to deepfakes – will grow harder, widening the gap between leaders and laggards. Many followers will opt for prestige pilots or regional pooling; others will hedge between rival stacks. Hedging can buy leverage, but without investment in evaluation capacity, safety assurance, and portability, it merely swaps one dependency for another.
The Middle-Power Playbook
Digital decoupling is incomplete but unmistakable. In the Indo-Pacific – home to some of the world’s most dynamic markets – choices made now will shape trade, security, and governance for decades. Middle powers should lead by shaping interoperability, not by choosing sides.
First, middle powers should set a portable safety floor. They can use procurement and standards diplomacy to require a minimal, testable post-deployment baseline – including independent red-teaming, incident reporting, and update logs – recognized across trusted labs.
Second, middle powers should build a compute commons involving pooled, sovereign cloud capacity for universities, SMEs, and public-interest projects, tied to open evaluation suites. This would spread capability without locking into a single bloc.
Finally, there’s a need to ensure the flow of talent and data corridors. Middle powers can fast-track visas and mutual recognition for assurance/safety roles and pilot narrow, audited data-sharing channels (health, climate, disaster response) even as broader flows tighten.
The separation at the AI frontier is real; the new deal for Nvidia to sell chips to China merely prices a trickle through the pipes. Power will accrue to those who control standards, infrastructure, and diffusion. In this competition, middle powers still have agency. They must use it, or live with other people’s defaults.