DeepSeek's Shock: Has China Levelled the AI Playing Field - and What Happens Next?
In January 2025, a relatively unknown Chinese AI startup named DeepSeek released an open-source reasoning model that matched or exceeded OpenAI's best offerings at approximately one-tenth the reported training cost. Nvidia lost over $600 billion in market capitalisation in a single trading session - the largest single-day market cap destruction in stock market history. The US Congress held emergency briefings. Sam Altman called it 'impressive.' The tech press called it a 'Sputnik moment' for American AI dominance. A year later, with multiple successors to DeepSeek now competing globally, the fundamental question has sharpened: has China genuinely caught up in the AI race - and does it matter who wins?
What DeepSeek Actually Did
DeepSeek's breakthrough was not raw performance - it was efficiency. American AI labs had been operating on the assumption that frontier AI required ever-larger compute clusters, ever-more-expensive chips, and training runs costing hundreds of millions of dollars. DeepSeek challenged all three assumptions simultaneously. Using a technique called Mixture of Experts - which activates only a relevant subset of the model's parameters for each query rather than the full model - combined with aggressive quantisation and novel reinforcement learning approaches, DeepSeek achieved comparable performance while using significantly fewer resources. Crucially, it did so while operating under severe US export controls that denied Chinese labs access to Nvidia's most powerful H100 and A100 chips. The implications were explosive: if you could build frontier AI without the most expensive chips, the entire US strategy of using chip export controls to maintain AI supremacy needed urgent reassessment.
The Export Control Strategy: Did It Fail?
The Biden administration's October 2022 semiconductor export controls - and their subsequent expansions - were premised on a straightforward theory: AI capability is compute-constrained, frontier compute requires advanced chips, and if the US controls the supply of advanced chips through TSMC and ASML export restrictions, it can maintain a decisive and durable lead in AI. DeepSeek's performance appeared to invalidate the first premise. If algorithmic innovation can substitute for raw compute - at least partially - then chip denial is a necessary but not sufficient strategy for maintaining AI supremacy. Critics of export controls argue that DeepSeek proved their futility: restricting chips simply incentivised Chinese researchers to become more algorithmically clever, ultimately strengthening rather than weakening China's long-term AI capabilities. Defenders of the strategy counter that DeepSeek's efficiency gains, while impressive, still hit hard limits without access to cutting-edge hardware; that the performance gap at the absolute frontier remains significant; and that the controls are buying time for the US to extend its lead in more advanced chip generations.
The Open Source Dimension
DeepSeek released its models as open weights - meaning anyone in the world could download and run them. This decision, which has since been followed by other Chinese AI labs, fundamentally complicates the geopolitical framing of the AI race. American AI policymakers who worry about Chinese AI being used for surveillance, disinformation, or military applications now confront an awkward reality: Chinese-origin AI models are freely available to researchers, companies, and governments worldwide. The models can be fine-tuned, modified, and deployed without any oversight by either the Chinese government or American export authorities. Meanwhile, American AI companies like OpenAI have moved in the opposite direction - increasingly restricting API access and making their most capable models proprietary, citing safety concerns but also commercial interests. The open-vs-closed AI debate within the US - with Meta's LLaMA on one side and OpenAI on the other - has now collided with the China AI debate in ways that make both conversations significantly more complex.
China's AI Ecosystem: Stronger Than the Narrative
Western media coverage has often portrayed Chinese AI as a capable but derivative follower of American innovation. The DeepSeek moment challenged that framing, but it was not the first evidence that it was wrong. Baidu's ERNIE Bot, Alibaba's Qwen series, Tencent's Hunyuan, and a dozen smaller labs have been producing competitive models across multiple domains for years. In specific areas - code generation, mathematical reasoning, and multimodal understanding of Chinese-language content - Chinese models have been competitive with or superior to American equivalents. China's AI research publication output surpassed the US years ago and continues to grow. The country produces more AI PhD graduates annually than any other nation and has the world's largest dataset of digitised human behaviour through its domestic internet ecosystem. The talent pipeline, the research output, and now the demonstrated engineering capability are all substantially stronger than the 'China is copying, America is inventing' narrative suggested.
The Military Dimension: Where the Stakes Are Highest
Beyond commercial competition, the AI race has a military dimension that makes it qualitatively different from previous technology competitions. The Pentagon's use of AI for drone swarm coordination, autonomous targeting, intelligence analysis, cyber operations, and logistics optimisation is already substantial and growing rapidly. China's military-civil fusion doctrine - which mandates knowledge transfer between commercial AI companies and the People's Liberation Army - means that civilian AI advances directly strengthen military capability. The most concerning scenarios involve AI-enabled hypersonic missile guidance, autonomous submarine operations, and - most alarmingly - AI systems that compress decision-making timelines in crisis scenarios to the point where human oversight becomes practically impossible. Arms control experts have been warning for years that unlike nuclear weapons, which require specific materials and infrastructure that can be monitored, advanced AI can be developed covertly and deployed with minimal warning. The DeepSeek shock has made US defence planners acutely aware that their previous assumptions about the pace and character of Chinese AI military capability were likely wrong.
The Case for AI Competition as Beneficial
A contrarian but increasingly mainstream argument holds that US-China AI competition, while geopolitically fraught, has been economically and technologically beneficial. Competition drives innovation: the pace of AI advancement over the past three years has been extraordinary partly because both US and Chinese labs have been racing each other. DeepSeek's efficiency breakthrough - which has already been absorbed into subsequent OpenAI and Anthropic model designs - ultimately benefited the entire global AI ecosystem. Proponents of engagement argue that attempts to wall off Chinese AI development are both futile and counterproductive: futile because algorithmic knowledge cannot be export-controlled like hardware, and counterproductive because decoupling removes the channels through which AI safety norms, standards, and risk assessments could be shared between the world's two most capable AI powers. In a world where both the US and China are racing toward artificial general intelligence, some argue the most dangerous scenario is not one where both reach the milestone together but one where they do so in complete strategic isolation with no shared understanding of the risks.
The Case for Treating AI as a Strategic Technology
The opposing view - held by most US national security officials across party lines - is that advanced AI is more analogous to nuclear weapons than to, say, the smartphone. The potential of AI to provide decisive military advantage, to enable mass-scale economic disruption, and to undermine democratic institutions through disinformation means that maintaining a significant capability lead over strategic adversaries is a genuine national security imperative. Export controls on chips and design software, restrictions on AI-related investment into China, and visa constraints on Chinese AI researchers are all imperfect tools - but they are not futile. The argument is not that China can be permanently locked out of frontier AI; it is that every year of delay in Chinese capabilities reaching military-relevant thresholds is a year in which the democratic world can develop better AI governance frameworks, defensive capabilities, and alliance structures. Time, in this view, is strategically valuable even if the race cannot ultimately be won.
Where India Fits In
India occupies a uniquely complicated position in the global AI race. It has the world's largest English-speaking AI talent pool, a booming AI startup ecosystem, and government ambitions - expressed through the IndiaAI Mission with a Rs 10,000 crore budget - to become a global AI power. Indian engineers are disproportionately represented in the leadership of every major US AI lab. Yet India lacks the semiconductor manufacturing capability, the sovereign computing infrastructure, and the large proprietary datasets that give the US and China their structural advantages. India's strategic response has been pragmatic: use both US and Chinese AI tools where they serve Indian interests, invest in sovereign AI capability for sensitive government applications, and position itself as a neutral AI governance broker in international forums. The DeepSeek moment was watched carefully in New Delhi - not as a threat, but as evidence that algorithmic ingenuity can compensate for hardware disadvantage, which is precisely India's situation.
The Governance Gap That Neither Side Is Filling
Regardless of who wins the AI race in commercial or military terms, the most consequential failure of 2025-26 has been the almost complete absence of binding international AI governance. The Bletchley AI Safety Summit of 2023 produced a declaration but no enforcement mechanism. The EU's AI Act is primarily a product regulation framework, not a global governance instrument. The UN's advisory body on AI produced recommendations that no major power has committed to implementing. Meanwhile, AI systems are being deployed at speed in hiring, lending, criminal justice, healthcare, and military targeting in ways that have no international oversight framework whatsoever. The US-China AI rivalry has made bilateral governance agreements politically impossible at precisely the moment they are most needed. The world's most powerful technology is being developed at maximum speed by two strategic competitors with diametrically opposed political systems and values, with no shared safety standards and no mutual transparency. Whether you believe the US should win the AI race or that competition is healthy, this governance gap is the most dangerous dimension of the entire situation.
The DeepSeek shock of 2025 was a genuine inflection point - not because China has overtaken the US, but because it revealed how fragile the assumptions underlying US AI strategy were. The race is closer, the stakes are higher, and the governance framework is almost entirely absent. Has American AI supremacy been a force for good - or is competition the better path? Argue it on Debatrix.