The AI Arms Race: From GPT to DeepSeek

The AI Arms Race: From GPT to DeepSeek

The competition for artificial intelligence supremacy between the United States and China has entered a new phase. The old assumption that hardware dominance guarantees AI leadership no longer holds.

For most of the past decade, the global AI race had a clear frontrunner. American companies — OpenAI, Google DeepMind, Anthropic, Meta — led the development of large language models, powered by vast computing budgets and unrestricted access to Nvidia's most advanced processors. China was seen as a capable but distant competitor, constrained by US export controls designed to limit its access to the very hardware that frontier AI requires.

Then came DeepSeek.

The DeepSeek Moment

In January 2025, the Chinese startup DeepSeek released R1, an open-weight reasoning model that matched the performance of OpenAI's o1 on key benchmarks — at a fraction of the training cost. The company claimed to have spent roughly $6 million in computing power to train a model that rivaled systems developed with budgets orders of magnitude larger. Nvidia's stock fell 17 percent in a single day. Markets wiped over half a trillion dollars in value from the company in what became the largest single-day decline for any stock in history.

The shockwave was not just financial. DeepSeek demonstrated that algorithmic efficiency could compensate, at least partially, for hardware constraints. Trained on tuned-down Nvidia chips rather than the latest restricted processors, R1 proved that high-performance AI was achievable without access to the most advanced silicon. The implications for US export control strategy were immediate and uncomfortable.

DeepSeek was not an isolated event. In the same period, several other Chinese labs — including Alibaba, Moonshot AI, and Zhipu AI — released models claiming comparable performance to leading American systems. Alibaba's Qwen series became the most downloaded open-source model family in the world, surpassing Meta's LLaMA, with over 700 million downloads and more than 180,000 derivative models built on its architecture.

Two Ecosystems Diverge

The AI race in 2026 is no longer a simple question of who builds the most powerful model. It has become a contest between two increasingly distinct ecosystems.

The American ecosystem remains centred on proprietary models backed by enormous capital investment. Goldman Sachs projects that AI companies may invest more than $500 billion in infrastructure in 2026. Microsoft and Google are more than doubling capital expenditures to meet demand for computing power. Nvidia's valuation surpassed $5 trillion in 2025, making it the first company to reach that milestone.

The Chinese ecosystem operates under different constraints and with a different philosophy. Facing restricted access to the most advanced chips, Chinese developers have prioritised algorithmic efficiency, open-source distribution, and rapid deployment into commercial applications. ByteDance's Doubao chatbot, Alibaba's Qwen assistant, and Tencent's Yuanbao have become part of daily life in China, typically offered free to individual users. AI adoption has accelerated across the private sector, government, and state-owned enterprises.

The open-source strategy carries particular geopolitical weight. By releasing powerful models freely, Chinese companies are building influence across developing markets. DeepSeek has seen two-to-four times higher adoption interest in Africa than in other regions. Several major US technology companies are already using Chinese open-source models in their own applications — Airbnb, for example, relies on Alibaba's Qwen models for its AI-driven customer service.

The Hardware Gap Persists — But Narrows

Despite DeepSeek's efficiency breakthroughs, the hardware gap between the two countries remains significant. Yet the trend line is clear. Chinese firms are accelerating efforts to deploy domestic alternatives to Nvidia, from Huawei as well as newer chipmakers including Moore Threads, Biren, and Enflame. DeepSeek has reportedly received conditional approval to purchase Nvidia H20 chips and is preparing to launch its next-generation V4 model in early 2026, which early reports suggest could be competitive with leading Western coding-optimised models.

The deeper question is whether hardware superiority still translates into AI superiority as cleanly as it once did. But DeepSeek's success has demonstrated that the relationship between compute and capability is more complex than a simple linear function.

The Geopolitical Stakes

AI is not just a commercial competition. It is increasingly treated as a national security asset by both Washington and Beijing.

The US AI Action Plan, released in 2025, frames the competition in explicit terms: the United States must export its full AI technology stack — hardware, models, software, and standards — to allies and partners, or risk ceding global influence. The strategy envisions AI partnerships with allies, similar to those forged with Saudi Arabia and the UAE in 2025.

China's approach combines state-directed investment with an aggressive open-source distribution strategy. By making powerful models freely available worldwide, Beijing is building a parallel AI infrastructure that does not depend on American platforms or hardware. This carries implications not only for commercial markets but for information ecosystems: models trained under different regulatory and political frameworks embed different assumptions about what content is acceptable.

Europe, meanwhile, has adopted a regulatory-first approach through the AI Act, emphasising safety and transparency requirements. While this has positioned the EU as a standard-setter, it has also raised concerns about competitiveness. European AI companies remain significantly smaller and less well-funded than their American and Chinese counterparts. The EU's challenge is to ensure that regulation enables rather than constrains innovation.

What Comes Next

The AI race in 2026 is entering what some analysts describe as the year of AI agents — systems that move beyond conversational chatbots to autonomously execute complex, multi-step tasks over extended periods. This shift will increase computing demands per user even further, sustaining the boom in infrastructure investment.

At the same time, the competition is becoming more plural. India, the Gulf states, Japan, and South Korea are all making significant investments in AI capabilities. The era of US-China duopoly in frontier AI may be giving way to a more distributed landscape, though the two superpowers will continue to exert the greatest influence.

What is certain is that the assumptions that governed the AI race even twelve months ago no longer hold. Hardware dominance alone does not guarantee leadership. Export controls can slow but not prevent progress. And the most consequential AI developments may come not from the largest budgets but from the most efficient algorithms.

The AI arms race is not slowing down. It is becoming more complex, more distributed, and more consequential for the global balance of power.

This is part of Tech Cold War's coverage of the artificial intelligence battleground. Subscribe to receive weekly analysis on the technologies reshaping global power.

Sources
CSIS, "DeepSeek's Latest Breakthrough Is Redefining AI Race," December 2025 (csis.org)
PIIE, "How the AI boom shrugged off the DeepSeek shock and keeps gaining steam," February 2026 (piie.com)
Foreign Policy, "How DeepSeek's AI Model Changes U.S.-China Competition," February 2025 (foreignpolicy.com)
EU Institute for Security Studies, "Challenging US dominance: China's DeepSeek model and the pluralisation of AI development," July 2025 (iss.europa.eu)
Atlantic Council, "Eight ways AI will shape geopolitics in 2026," January 2026 (atlanticcouncil.org)
Capacity Media, "DeepSeek one year on: How a Chinese AI model reshaped the global AI race," February 2026 (capacityglobal.com)
Rest of World, "DeepSeek's new model could push China ahead in the global AI race," December 2025 (restofworld.org)
Caixin Global, "Year in Review: DeepSeek's Breakout Rewrites U.S.-China AI Race," January 2026 (caixinglobal.com)