The global race in artificial intelligence has moved beyond the competition for the best chatbot. The real battle is now over infrastructure: data centers, chips, power, memory, storage, networking, model-training pipelines, inference platforms and the cloud services that will deliver AI to billions of users. AI is no longer treated as a feature inside software. It is becoming a general-purpose layer for the economy.

This explains why a small number of very large companies in the United States and China are investing at a scale normally associated with energy systems, railways, semiconductors or national infrastructure. Microsoft, Alphabet, Amazon, Meta, OpenAI with Oracle and SoftBank, Alibaba, ByteDance, Tencent and Baidu are not behaving like ordinary software firms. They are trying to secure control over the next computing platform.

The paradox is that spending is rising much faster than proven AI profits. Some of these companies are highly profitable overall because of advertising, cloud, commerce, enterprise software or consumer platforms. But the new AI infrastructure itself has not yet shown returns proportional to the capital being committed. The pure AI labs face an even sharper challenge: training, inference, talent and data costs are enormous, while revenues are still immature. The question is therefore not whether AI matters. It clearly does. The question is whether the current investment cycle will produce broad social value, durable profits and open innovation, or whether it will deepen digital dependency on a handful of closed platforms.

The first major trend is pure scaling. Its premise is that larger models, larger datasets and more compute produce better performance in predictable ways. The scaling-laws tradition made compute the central resource of frontier AI. This view naturally favors the United States, where cloud hyperscalers, venture capital and public markets can finance massive infrastructure buildouts. It also explains why AI has become a data-center and energy question, not just a software question.

The second trend is reasoning models. Here progress does not come only from bigger pre-training runs. It comes from using more computation at inference time. The model is encouraged to work longer on a problem, explore alternative solutions, check itself and produce more reliable answers in mathematics, coding, science and complex decision tasks. OpenAI, Google DeepMind and DeepSeek represent different versions of this approach. DeepSeek was especially important because it showed that reasoning capabilities can be improved through more efficient methods and open model releases, challenging the assumption that only the richest closed labs can participate at the frontier.

The third trend is world models. This school starts from the idea that language is not enough. An intelligent system must understand video, space, motion, physical causality, bodies, tools and environments. Meta’s JEPA line of work, robotics labs and video-generation systems point in this direction. If language models are systems for manipulating symbols, world models aim to become systems for understanding situations.

The fourth trend is neurosymbolic AI. It recognizes that neural networks are excellent at pattern recognition and weak at formal verification, strict logic and mathematical reliability. That is why leading systems increasingly combine neural models with symbolic solvers, theorem provers, knowledge graphs and rule-based verification. AlphaGeometry and AlphaProof are strong examples. The neural model proposes. The symbolic system checks.

The most important point is that these schools are no longer isolated. The frontier is hybrid. The future AI stack is likely to combine scaled models, test-time reasoning, multimodal world understanding and formal verification. That makes the infrastructure race even more capital-intensive, because every layer adds new compute needs.

The companies investing hundreds of billions are not doing so because the profit model is already fully settled. They are doing so because they believe the cost of being absent from the next platform would be existential. If AI becomes the main interface for search, productivity, coding, education, research, public services, commerce and media, then whoever controls the infrastructure will control access to knowledge, users, developers and markets.

Microsoft wants to turn cloud and enterprise software into an AI-native productivity layer. Alphabet wants to defend search, advertising, YouTube, Android and Google Cloud. Amazon wants AWS to be the default compute platform for training and inference. Meta wants to embed AI into social media, advertising, creator tools and future consumer interfaces. OpenAI, Oracle and SoftBank are trying to secure dedicated compute capacity at a scale that reduces strategic dependence. Alibaba, ByteDance, Tencent and Baidu are moving inside a Chinese context where AI is at once a commercial opportunity, a state priority and a matter of technological self-reliance.

There is also a structural reason. AI infrastructure has winner-takes-most characteristics. More compute can produce stronger models. Stronger models attract more users and developers. More users create more feedback, data and revenue opportunities. More revenue supports more compute. The loop rewards scale and punishes late entry.

This creates a powerful defensive logic. Even if AI products are not yet fully profitable, the incumbent platforms cannot afford to let a rival become the default AI layer. A search company cannot ignore AI assistants. A cloud company cannot ignore model hosting. A social platform cannot ignore synthetic content and personal agents. An e-commerce and logistics giant cannot ignore AI planning, recommendation and automation. The spending is partly offensive, but also deeply defensive.

The danger is that this race could recreate the worst features of the platform economy at a larger scale. AI could become a new layer of dependency where states, universities, hospitals, municipalities and small firms rent intelligence from a few proprietary clouds. That would create technological lock-in, opaque decision-making, weak auditability and a permanent transfer of public and private value to foreign infrastructure owners.

The environmental dimension is also serious. AI data centers require electricity, water, land, chips and supply chains. If the benefits remain private while the energy, environmental and democratic costs are socialized, the legitimacy of the AI transition will weaken. The question is not only whether AI can increase productivity. It is whether the infrastructure of AI will be accountable to society.

This is where open source, open standards and local models become strategically important. The alternative to the hyperscaler race is not technological isolation. It is a plural AI ecosystem built around open models, interoperable APIs, public datasets, verifiable systems, local deployment and shared infrastructure. Europe does not need to imitate the US and China by trying to outspend them in every layer. It can compete through trust, efficiency, openness, democratic governance and sector-specific public value.

For Greece, the lesson is clear. The country should not become a passive buyer of token-based AI services from closed platforms. Public administration, education, health, language technology, culture and research need controlled infrastructure, open models where possible, transparent datasets, model cards, datasheets, RAG systems with verified sources and human responsibility for final decisions. Public money should create public code, public knowledge and public AI capacity.

The giant platforms are spending hundreds of billions because they are trying to control the next era. Democratic societies should invest differently. They should build AI as a public capability, not merely consume it as a private service. The goal should not be the largest model at any cost. It should be useful, transparent, accountable and reusable AI infrastructure that strengthens citizens, researchers, small firms and public institutions.