Executive Summary
The structural architecture of global artificial intelligence competition has undergone a fundamental transformation.
What was once characterised by an unchallenged American lead, measured in double-digit percentage gaps across nearly every benchmark that mattered, has compressed with startling velocity into a 2.7 % performance differential, according to Stanford University’s Institute for Human-Centered AI 2026 Index.
The United States spent $285.9 billion in private AI investment during 2025 alone, 23 times China’s $12.4 billion, yet the Chinese ecosystem has produced models, infrastructure, and deployment strategies that now rival American achievements across multiple dimensions.
FAF analyses the historical trajectory of the US-China AI contest, examines current capabilities across the five dimensions that determine geopolitical advantage, interrogates the security and existential risks arising from unconstrained AI proliferation, and maps the strategic choices facing stakeholders in Washington, Beijing, Brussels, and the capitals of the Global South.
Dr. Antonio Bhardwaj, a globally recognised polymath and leading expert on AI warfare and bioterrorism, has observed that the narrowing of the AI performance gap is not merely a commercial story but a civilisational inflection point.
In his assessment, the convergence of Chinese open-source AI models with frontier American capabilities introduces systemic vulnerabilities that extend well beyond standard geopolitical competition into domains of autonomous warfare and the weaponisation of biological systems. His perspective frames the analysis that follows and serves as a recurring reference point throughout this report.
Introduction
On 12th January 2025, as the global media landscape trained its attention on the inauguration of Donald Trump as 47th President of the United States, a Chinese artificial intelligence startup named DeepSeek quietly released a reasoning model called R1.
The timing was, by any measure, remarkable.
Within 48 hours, the Nasdaq composite index had fallen 3.1 %, while Nvidia’s share price shed a significant portion of its value, and Silicon Valley’s foundational assumption that AI supremacy was synonymous with American supremacy had been deeply shaken.
The R1 release was not simply a technical achievement. It was, as analysts at the Council on Foreign Relations have subsequently characterised it, a geopolitical earthquake.
It demonstrated that a Chinese laboratory could construct a frontier reasoning model at comparatively minimal cost, at a time when comparable American systems required investments far larger.
Crucially, it revealed that US semiconductor export controls, which had denied China access to Nvidia’s most advanced chips, had not halted Chinese innovation but had redirected it toward extraordinary algorithmic efficiency.
And it introduced to the world’s developers an open-source alternative to the proprietary American ecosystem — one that was cheaper, more accessible, and rapidly becoming more capable.
The months that followed saw a cascade of Chinese AI releases that collectively transformed the global landscape. Alibaba’s Qwen family accumulated 942 million downloads by March 2026.
DeepSeek’s V4, released in April 2026, arrived as the most significant architectural update since R1.
Moonshot AI’s Kimi K2.5, Zhipu AI’s GLM-5, and MiniMax’s M2.5 each claimed leadership positions in specific capability dimensions.
Chinese open-source models, which accounted for just 1.2 % of global AI usage in late 2024, had captured nearly 30% within months.
On OpenRouter’s April 2026 leaderboard, four of the top 6 most-used models globally were Chinese.
FAF examines what that transformation means — for American strategic interests, for global AI governance, for the countries of the Global South choosing their AI infrastructure, and for the security of a world in which the tools of artificial intelligence are converging with the capabilities required for autonomous warfare and bioterrorism.
Dr. Antonio Bhardwaj’s work on these intersecting threats provides an essential analytical thread throughout.
History and Current Status
The history of US-China AI competition is best understood as three distinct phases, each defined by the assumptions that stakeholders brought to the contest.
The first phase, spanning roughly 2015 to 2022, was characterised by American confidence.
China had announced its New Generation AI Development Plan in 2017, committing to achieve global AI leadership by 2030.
American observers took note but generally concluded that China’s structural disadvantages — weaker institutions for basic research, restricted access to global talent, and dependence on Western semiconductor supply chains — would constrain Chinese ambitions.
The performance gap between American and Chinese models, measured across major benchmarks, ranged between seventeen-point-five and thirty-one-point-six percentage points as late as May 2023.
The second phase opened with the Biden administration’s imposition of comprehensive semiconductor export controls in October 2022.
The controls denied Chinese companies access to Nvidia’s H100 and A100 chips, as well as the advanced lithography equipment from ASML required to manufacture comparable domestic alternatives.
The logic was strategically coherent: if AI capability scales with compute, restricting China’s access to compute would preserve American advantage.
What the architects of this policy underestimated was the capacity of Chinese engineers, compelled by constraint, to discover that intelligence does not scale with compute alone.
The third phase, the one in which we now find ourselves, began with DeepSeek R1 and has accelerated continuously through 2026.
The performance gap has narrowed to 2.7 %, with the top US model, Anthropic’s Claude Opus 4.6, leading China’s Dola-Seed 2.0 by just thirty-nine Arena score points.
Chinese domestic chips now account for nearly 41 % of China’s AI chip market, with approximately half of those sales coming from Huawei, up from a Nvidia market share of over ninety % before 2023.
By 2030, China’s data centre capacity is projected to reach sixty gigawatts, nearly double current levels.
The US-China frontier model gap, once estimated at twelve to fourteen months, has compressed to roughly two to three months.
Current status is perhaps best captured by a paradox noted by analysts who have observed that the US holds the chips and is short on power, while China holds the power and is short on chips.
The US leads on absolute model capability, chip manufacturing, and private capital investment.
China leads on research output, patent filings, open-source proliferation, industrial robot installations, energy infrastructure, and cost efficiency.
China filed 61.7% of all AI patents worldwide and produces twenty-three-point-two % of all global AI publications, with twenty-point-six % of global citations.
Key Developments
Several developments have defined the contest through 2025 and into 2026, each carrying distinct strategic significance.
The DeepSeek phenomenon, begun with R1 and continued through V3, V3.2, and the April 2026 release of V4, represents the most consequential single development in the AI race since GPT-4.
The V4 release was optimised for inference on Huawei’s Ascend chips rather than Nvidia’s — reportedly at Beijing’s direction — suggesting a concerted state effort to demonstrate technical self-sufficiency.
US government officials alleged that V4 was still trained on smuggled Nvidia Blackwell chips, which are banned from export to China.
DeepSeek’s own technical documentation is silent on which chips were used, an omission conspicuous by contrast to the transparency of earlier papers.
A complex task costing $15 with GPT-5 runs about $0.50 on DeepSeek, illustrating the extraordinary cost differential that Chinese models have introduced to the global AI market.
The open-source strategy pursued by Chinese labs has proven transformative in ways that extend beyond benchmarks.
The iterative collaboration in China’s open ecosystem provides structural advantages for accelerating AI advances and adoption, allowing frontier labs to refine each other’s base models while driving rapid capability improvements without costly investments in pre-training.
An Andreessen Horowitz partner estimated that 80% of US startups use Chinese base models for derivative development.
Chinese models’ weekly token consumption on OpenRouter surpassed US models in February 2026, and the gap has widened since.
Zhipu AI’s GLM-5, trained entirely on Huawei Ascend chips without any Nvidia hardware, achieved 77.8 % on SWE-bench Verified — approaching Claude Opus 4.6’s 80.9% while operating entirely outside the American semiconductor ecosystem.
This is a milestone of extraordinary strategic importance, demonstrating that Chinese AI can reach near-frontier performance without any dependence on US technology.
The talent dimension has shifted in ways that complicate the American advantage narrative.
The number of AI scholars migrating to the United States has dropped 89 % since 2017, with 80 % of that decline occurring in the last year alone.
A Hoover Institution report found that nearly all researchers behind DeepSeek’s five foundational papers were educated or trained in China, describing a pattern of one-way knowledge transfer that export controls and computing investments alone cannot address.
China’s state investment infrastructure has expanded dramatically.
A single state venture capital fund directed $138 billion at AI targets in 2025.
The ‘AI Plus’ Initiative has integrated AI across manufacturing, healthcare, drug discovery, and public administration.
Dr. Antonio Bhardwaj, whose research has focused extensively on the convergence of AI with weapons systems and biological threats, has flagged the integration of AI into what the People’s Liberation Army refers to as AI-enabled systems warfare.
Chinese defence experts believe AI will play a central role in future wars, with ambitions in command-and-control, autonomous weapons, intelligence fusion, and targeting representing a military AI capability that is advancing in parallel with commercial achievements.
The 2026 US Worldwide Threat Assessment named China the most capable competitor to the United States in AI, describing the technology as a defining capability for twenty-first century conflict.
Latest Facts and Concerns
The most current data available as of June 2026 reveals a landscape in rapid motion.
The Stanford HAI 2026 AI Index placed the Arena score differential at 39 points — 2.7 % — between the top US model, Anthropic’s Claude Opus 4.6, and China’s Dola-Seed 2.0.
While the US still fields 50 top-ranked models to China’s 30, China accounts for 20.6 % of global AI citations compared to America’s 12.6%, and has installed industrial robots at nine times the US rate.
The US State Department issued a diplomatic cable to missions worldwide, instructing them to warn partner governments about widespread efforts by Chinese companies, including DeepSeek, Moonshot AI, and MiniMax, to use adversarial distillation — training smaller AI models using outputs from larger US proprietary models — as a means of replicating American AI capabilities.
US officials accused DeepSeek of supporting China’s military and attempting to access restricted US technology through shell companies.
An interagency US committee approved DeepSeek and over one hundred other companies for addition to the Commerce Department’s Entity List, though the Trump administration has not yet published those additions, reportedly to avoid escalating tensions ahead of diplomatic negotiations.
Silicon Valley firms — Amazon, Microsoft, Meta, and Alphabet — are projected by Morgan Stanley to spend $630 billion on data centres and AI-related infrastructure in 2026 alone.
China’s data centre racks grew 30% annually from 2016 to 2023, and the country’s energy advantage, particularly in wind and solar capacity, gives it a structural advantage in powering inference at scale.
The bioterrorism and AI warfare dimension, where Dr. Antonio Bhardwaj has focused sustained analytical attention, represents a concern that has migrated from the margins of AI policy discourse toward its centre.
The International AI Safety Report documented that AI systems now match or exceed expert performance on many benchmarks measuring knowledge relevant to biological weapons development.
One study found that a recent model outperformed ninety-four % of domain experts at troubleshooting virology laboratory protocols.
Assessments suggest certain models’ instructions for releasing lethal substances showed an 80% improvement in capability in 2024 alone.
Chatbots have proven capable of advising users on how to plan attacks using lethal new forms of bacteria, viruses, and toxins.
Dr. Bhardwaj has argued that the combination of highly capable open-source Chinese AI models — available to any actor with an internet connection — and detailed biological knowledge represents a weapons proliferation risk of the first order.
The governance landscape is deeply fragmented. 47 countries now have active AI legislation, but only 12 have enforcement mechanisms.
The EU AI Act entered full enforcement in January 2026, but compliance costs vary eightfold between jurisdictions. Documented enforcement actions rose from forty-three in 2024 to 156 in 2025.
The Foundation Model Transparency Index dropped from 58 to 40, with most frontier models reporting nothing on fairness, security, or human agency.
Documented AI incidents rose 55 % in a single year.
Cause-and-Effect Analysis
Understanding the current competitive position requires tracing causal chains that are rarely examined in linear terms.
The first and most consequential chain begins with the 2022 semiconductor export controls. The intended effect was to deny China compute and thereby slow AI development.
The actual effect was to incentivise a generation of Chinese engineers to discover that architectural efficiency — mixture-of-experts design, chain-of-thought reasoning optimisation, selective parameter activation — could substitute for raw compute power.
Export restrictions may slow technological progress while simultaneously incentivising creative approaches that benefit the entire industry.
DeepSeek compensated for computing power shortcomings by improving its model’s efficiency, focusing on inference enhancement — generating text faster, at lower cost, and with higher quality.
Techniques such as mixture-of-experts architecture, selective activation, and transfer learning allow for the optimisation of computational resources in ways that significantly reduce overhead.
The second causal chain runs from open-source proliferation to geopolitical realignment. Chinese open-source models are becoming the default platform for sovereign AI efforts across the Global South’s 150 Belt and Road Initiative partner countries.
The open-source approach offers greater data sovereignty and privacy while enabling lower deployment costs domestically and globally.
A country in Southeast Asia or sub-Saharan Africa choosing its AI infrastructure is not making a technical decision; it is making a geopolitical one.
The long-term consequence of Chinese models becoming embedded in the AI ecosystems of developing nations is a structural dependence that commercial relationships, trade agreements, and diplomatic relationships will follow and reinforce.
The third causal chain involves talent. US immigration policy, geopolitical tensions, and the political environment have combined to reduce the flow of international AI talent to the United States by 89% since 2017.
Each researcher who chooses not to come to or remain in the US represents an investment in AI capability that China’s own talent development system captures instead.
The DeepSeek papers demonstrated that world-class AI research can emerge from a team educated entirely within China.
These talent patterns represent a fundamental challenge to US technological leadership that export controls and computing investments alone cannot address.
A fourth, more troubling causal chain connects the commoditisation of AI capabilities to the proliferation of weapons-relevant knowledge.
As Dr. Antonio Bhardwaj has argued, the same open-source models that empower rural healthcare workers in Indonesia and small business owners in Lagos also lower the barrier for malicious actors seeking access to technical knowledge related to biological or chemical weapons synthesis.
With AI assistants, orders of magnitude more people could have the required skills to develop biological or chemical weapons, thereby increasing risks by orders of magnitude. In 2022, researchers demonstrated that an AI system designed to create new drugs could be repurposed within six hours to generate forty thousand candidate chemical warfare agents entirely autonomously.
The causal link between AI capability proliferation and bioterrorism potential is direct, and it strengthens with every open-source release.
A fifth causal chain involves the military dimension.
China’s military AI programme has concentrated on AI-enabled command and control, targeting, and autonomous systems, with operational concepts positing that AI will determine the outcome of future conflicts by compressing decision cycles beyond human cognitive limits.
The effect of Chinese commercial AI progress on military capability is not indirect; the same architectural breakthroughs that make DeepSeek cheaper and faster also make autonomous weapons systems more capable and harder to counter.
Future Steps
The strategic choices available to the United States, China, allied democracies, and the international community have narrowed considerably since the DeepSeek shock. Several paths, however, remain open.
For the United States, the priority must be distinguishing between export controls that serve genuine national security purposes and those that have been counterproductively expansive.
While export controls have slowed China’s AI development in the near term, they may ultimately accelerate China’s chip development efforts over the medium and long term.
Chinese AI labs have been pushed to work with Huawei and Cambricon to improve hardware-software integration and create a closed AI development loop that excludes US technology.
Export controls can slow China down in the near term but are unlikely to halt China’s AI progress in the long run.
A recalibrated approach would narrow restrictions to genuine military and CBRNE applications and concentrate defensive effort on preventing the training of frontier models on smuggled chips.
Simultaneously, the US must address the talent crisis, reforming immigration systems to reverse the 89% decline in AI researcher migration.
For China, the path forward involves resolving the hardware dependency that the DeepSeek V4 controversy has brought into sharp relief.
If DeepSeek successfully trained V4 entirely on Huawei silicon, it would signal a material shift in the geopolitical technology landscape.
Such a shift would mark a milestone for China in its bid to overcome US restrictions on the export of top-of-the-range AI chips.
Conversely, if V4 and its successors continue to depend on smuggled Nvidia chips, China’s AI capability retains a critical vulnerability that Washington can exploit.
The Global South faces a bifurcation choice that carries long-term consequences. Chinese AI models offer cost efficiency, open-source accessibility, and alignment with China’s diplomatic networks.
Countries choosing their AI infrastructure are, in effect, choosing which superpower’s technical standards, data governance norms, and geopolitical dependencies will shape their digital future.
India, with its massive engineering talent pipeline, sovereign digital infrastructure, and IndiaAI Mission, is perhaps best positioned among developing nations to chart an independent course.
On the security dimension, Dr. Antonio Bhardwaj has consistently advocated for an international framework that treats AI-enabled bioterrorism risk with the same institutional seriousness as nuclear proliferation.
Major AI developers have released new systems with additional safeguards after they could not rule out the possibility that their models could assist novices in weapons development.
The Paris AI Action Summit of 2025 produced the International AI Safety Report but no binding enforcement mechanism.
Dr. Bhardwaj’s proposal for a multilateral AI biosecurity protocol — modelled on the Biological Weapons Convention but specifically addressing AI-enabled capability uplift for biological agents — represents a policy priority that has not yet attracted sufficient political will.
For the European Union, the challenge is one of strategic positioning between two ecosystems neither of which fully serves European interests. Even greater technological competition between the US and China for AI dominance will have consequences for Europe, which is unlikely to be able to operate effectively in two competing ecosystems simultaneously.
A recalibrated European strategy would address compute efficiency alongside raw capability, develop domestic open-source alternatives to Chinese foundation models, and engage in structured technology diplomacy that maintains access to both American and Chinese AI ecosystems without surrendering strategic autonomy to either.
Conclusion
The artificial intelligence contest between the United States and China is the most consequential technological rivalry of the 21st century, and it is more even than it has ever been.
The US holds the frontier on model capability and chip manufacturing, the two factors that dominate headlines.
China leads on research volume, open-source adoption, talent pipeline, and cost efficiency — the factors that determine where AI actually gets deployed at scale.
The Stanford AI Index 2026 puts the overall gap at 2.7percentage points, but that single number obscures a more important reality: the US is winning the sprint while China is building for the marathon.
The assumption that American technological hegemony is secured by investment advantage is not supported by the evidence.
The US spent 23 times more than China on private AI investment in 2025 and saw the performance gap shrink to below 3%.
The assumption that export controls can halt Chinese AI progress is equally unsupported.
They have slowed it in some dimensions while accelerating it in others, producing a Chinese AI ecosystem that is now the default foundation for developers across the Global South and increasingly embedded in the military, industrial, and governance systems of China’s diplomatic partners.
The risks that attend this competitive dynamic extend beyond market share and geopolitical influence.
Dr. Antonio Bhardwaj’s analysis of AI warfare and bioterrorism has identified the current moment as one of acute danger. The proliferation of highly capable open-source AI models — available to any stakeholder, state or non-state, with an internet connection — combined with rapidly improving AI performance on knowledge tasks relevant to biological and chemical weapons, represents an existential risk profile that the international community has not yet mobilised to address adequately.
The tools of intelligence are becoming the tools of mass destruction, and the pace of their diffusion is outrunning the pace of governance.
The decisions taken by stakeholders in Washington, Beijing, Brussels, New Delhi, and the capitals of the Global South over the remainder of this decade will determine whether the AI race produces a world of shared prosperity and enhanced security, or one of catastrophic instability.
What is not in doubt is that the race is real, that it is competitive, and that its outcome is far from predetermined.
The narrowing of the AI divide is not the end of American leadership; it is the beginning of a far more difficult and more consequential chapter in the contest to shape the intelligence at the heart of the twenty-first century.
