Artificial Intelligence Must Not Become a New Cold War

The trap of the US-China AI race

The American debate on artificial intelligence is increasingly shaped by one simple but dangerous sentence: if we slow down, China will overtake us. This sentence has become a political shortcut. It allows major technology companies and their allies to describe regulation, safety testing and public accountability as threats to national survival.

This is more than geopolitical competition. It is a cold-war mentality that treats AI as a field of dominance rather than a social infrastructure that must be governed in the public interest. Every restriction is framed as a gift to Beijing. Every delay is described as a loss of strategic advantage. Every demand for transparency is treated as an obstacle to innovation. Under this pressure, even reasonable safeguards for powerful models can be dismissed before they are properly evaluated.

The result is a race in which speed becomes ideology. Advanced AI systems can support software development, education, research and public services, but they can also assist cyber operations, fraud, disinformation, manipulation and dangerous technical workflows. If the only policy command is to win, safety becomes a burden and accountability becomes optional.

The economic weakness of the race

There is also a second problem: the current AI race may be economically unstable. The dominant paradigm is built around massive data centers, expensive chips, huge energy demand and models that are costly to train and costly to serve. Yet the market is already moving toward price sensitivity and price wars. If many firms are building similar systems with similar data and similar techniques, there is no durable moat. A technical lead may last weeks, not decades.

Chinese open-weight models intensify this pressure. Z.ai’s GLM-5.2 and the broader DeepSeek experience show that the gap with leading closed models can narrow quickly and at much lower cost. Even more important is the shift from “larger model” to “more efficient system.” DeepSpec and DSpark are not just about making models smarter. They are about making model serving faster and cheaper, using speculative decoding techniques to reduce wasted computation.

This matters because inference is becoming the real deployment bottleneck. Standard LLM generation is sequential and memory-bound. Speculative decoding changes the economics: a smaller draft process proposes tokens, while the larger target model verifies them more efficiently. DSpark adds confidence scheduling and semi-autoregressive drafting, aiming to improve speed without changing the model’s correct output distribution. In plain terms, the same AI service can become faster and cheaper without simply adding more hardware.

If AI becomes cheaper, faster and easier to run locally, the argument for a race of ever-larger closed infrastructure becomes weaker. The future of AI may not belong only to the largest data centers. It may also belong to efficient, open and locally deployable systems.

Efficient AI as public-interest innovation

This is the real strategic lesson. The future of AI is not only about intelligence. It is about efficiency, accessibility, energy use, privacy and control. A model that can run on local infrastructure can protect sensitive data better than a remote black-box service. A faster serving stack reduces energy costs. Open tools allow universities, municipalities, small firms and public agencies to adapt AI to their own needs.

This does not mean that openness alone solves safety. The more capable and cheaper AI becomes, the more important governance becomes. Open does not mean uncontrolled. It means inspectable, reproducible and auditable. It allows independent researchers, public institutions and civil society to test systems rather than trust corporate assurances.

Why Europe needs a different path

The European Union should not respond to the US-China race by entering a deregulation contest. Europe’s choice is not between stagnation and reckless acceleration. Its task is to build a different model of AI sovereignty: open, secure, efficient and democratically accountable.

The EU AI Act should not be treated as a burden. It should become the basis for a mature industrial and social strategy. Europe needs open standards, public compute, open models where appropriate, local deployment, independent evaluation, model cards, dataset documentation, audit logs and retrieval systems grounded in verified sources.

For Greece and other smaller European countries, this is practical, not abstract. Public administration, education, health, culture and language technologies should not depend entirely on opaque foreign platforms. AI in the public sector should be built as a public capability: open-source software, high-quality national datasets, local low-cost models, human final responsibility and democratic oversight.

The best answer to the new AI cold war is not to run without rules. It is to build AI that is safe, efficient, accessible and useful. Europe should treat responsible innovation not as weakness, but as strategic maturity.

Sources:

1. Robert Wright, “What if Trump is right to pump the brakes on the most advanced AI?”: Wright’s article examines how nationalist pressure to beat China can push US AI policy away from precaution and toward reckless acceleration: https://www.washingtonpost.com/opinions/2026/06/26/trump-fable-ai-ban-was-not-gift-china/,
2. Robert Wright & Ryan Fedasiuk, “The US-China AI Race”: The conversation explores US-China dynamics, AI safety, global pause proposals and the strategic difficulty of cooperation between rival powers: https://www.nonzero.org/p/the-us-china-ai-race-robert-wright
3. YouTube, “The US-China AI Race | Robert Wright & Ryan Fedasiuk”: The video provides accessible context on the geopolitical framing of frontier AI and the risks of treating the field as a zero-sum race: https://www.youtube.com/watch?v=5FTm0hS-vrE,
4. Gary Marcus, “Why things will eventually fall apart”: Marcus formulates the “no moat, more competitors, price wars, scarce profits” argument and connects it to the economic fragility of the current AI race: https://garymarcus.substack.com/p/why-things-will-eventually-fall-apart,
5. Gary Marcus, “The illusion of Generative AI, the insanity of massive bets on hyperscaling”: This post links criticism of hyperscaling to the case for world models, neurosymbolic AI and more reliable architectures: https://garymarcus.substack.com/p/the-illusion-of-generative-ai-the,
6. DeepSeek, “DeepSpec”: The open repository presents a full-stack codebase for training and evaluating draft models for speculative decoding, highlighting the importance of efficient model serving: https://github.com/deepseek-ai/DeepSpec,
7. DeepSeek, “DSpark paper”: The DSpark technical report describes confidence-scheduled speculative decoding with semi-autoregressive generation for faster large-model serving: https://github.com/deepseek-ai/DeepSpec/blob/main/Dspark_paper.pdf,
8. Hugging Face, “DeepSeek-V4-Pro-DSpark”: The model page shows that DSpark checkpoints can be used with tools such as Transformers, vLLM and SGLang, supporting broader deployment beyond a single closed platform: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-Dspark,
9. Reuters, “After Anthropic shutdown, China’s Z.ai closes frontier gap”: Reuters documents GLM-5.2’s strong benchmark performance, lower operating cost and strategic relevance for China’s AI ecosystem: https://www.reuters.com/world/asia-pacific/after-anthropic-shutdown-chinas-zai-closes-frontier-gap-it-plans-dual-listing-2026-06-25/,
10. Z.ai, “GLM-5.2: Built for Long-Horizon Tasks”: Z.ai’s technical post presents GLM-5.2’s results on coding and long-horizon benchmarks and its positioning against closed frontier models: https://z.ai/blog/glm-5.2,
11. European Commission, “AI Act”: The official Commission page explains the EU’s risk-based AI framework, its timeline and the governance architecture for trustworthy AI: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai,
12. NIST, “AI Risk Management Framework”: The NIST AI RMF provides a technical framework for mapping, measuring and managing AI risks, showing that risk governance is compatible with innovation: https://www.nist.gov/itl/ai-risk-management-framework.