The Third Way That Doesn't Exist

In February 2024, Emmanuel Macron made a speech in Paris that contained a phrase that became briefly famous in AI policy circles: “We do not want to be the colony of either.” The sentiment was directed at Europe’s dependence on American AI infrastructure and the risks of Chinese technology alternatives, and it captured something real about how European (and, by extension, many non-aligned countries’) leaders understand their situation.

The speech did not come with a budget that would change that situation. The phrase was diagnostically accurate and strategically insufficient in a way that describes most of the “third way” thinking about AI sovereignty.

The honest arithmetic of frontier AI development is brutal for ambitions that are not backed by brutal amounts of money. Training a frontier model in 2027 requires a compute cluster of perhaps 10,000 to 100,000 high-end AI accelerators, running continuously for weeks to months. The hardware alone costs somewhere between $500 million and $5 billion, depending on the tier of the model being trained. The electricity to run the cluster for a frontier training run costs tens of millions of dollars. The engineers who design and execute the training run are among the most sought-after in the world, earning compensation that strains academic budgets beyond recognition.

The companies currently training frontier models — OpenAI, Google DeepMind, Anthropic, Meta, and a handful of others — are spending billions of dollars per year on AI infrastructure. Microsoft’s investment in OpenAI alone exceeded $13 billion before 2025. Google’s internal AI compute budget, which includes both training and inference, is a significant fraction of the company’s enormous capital expenditure. These are not R&D budgets. They are industrial spending at a scale that blurs the line between technology company and heavy manufacturer.

No European country or company is in this league. The EU’s combined AI computing initiatives — the EuroHPC Joint Undertaking, the French sovereign AI program, the German AI network — represent real investment but an order of magnitude below what the American frontier labs spend annually. Mistral AI, the most credible European frontier lab, has raised approximately €1 billion in venture funding across all rounds. OpenAI raised more than that in a single round in 2023.

Japan’s GENIAC program, announced with considerable fanfare, is investing ¥100 billion (roughly $700 million) in domestic AI compute. India’s IndiaAI mission has committed ₹10,372 crore (approximately $1.2 billion) over multiple years. The UK’s AI Safety Institute has secured compute commitments that are measured in hundreds of millions of pounds. These are meaningful investments in a domestic computing infrastructure that previously didn’t exist. They are not investments that produce the ability to train GPT-5 class models.

The honest technical implication is that most countries’ AI sovereignty programs will produce capabilities in the range of the third or fourth tier of current frontier models — useful for many applications, genuinely valuable for domestic language and domain-specific use cases, significantly short of the most capable systems. That is not nothing. A domestically trained model that understands Hindi idiom, Indian legal frameworks, and Indian government processes well is more useful for Indian applications than a frontier American model that knows English perfectly and Indian languages adequately. The domain-specific advantage is real.

What it does not produce is independent capability in the most strategically sensitive applications: autonomous military systems, intelligence analysis, scientific discovery at the frontier, the kind of reasoning capability that might produce qualitative advantage in complex strategic planning. Those capabilities track the frontier, and the frontier is being set by the two countries whose resource commitments are categorically different from everyone else’s.

The European situation has one distinctive characteristic that makes the “third way” framing more than empty aspiration: regulatory leverage. The EU has approximately 450 million consumers, a sophisticated regulatory apparatus, and a demonstrated willingness to use market access as a bargaining chip in technology negotiations. The GDPR generated compliance investments from every global technology company. The AI Act will do the same. The Digital Markets Act is reshaping how American platform companies operate in Europe. These instruments do not produce European AI capabilities, but they do give European regulators a seat at the table in decisions about how AI is deployed, on what terms, with what consumer protections and data rights.

Regulatory power is a different kind of AI sovereignty from compute sovereignty. It is more achievable for Europe’s current resource level, and it addresses a real set of concerns — about AI accountability, algorithmic discrimination, and data rights — that matter for European citizens regardless of where the models were trained. The strategic risk is conflating regulatory influence with technological independence: being able to set the rules for how American AI is deployed in Europe is not the same as being able to train competitive AI in Europe, and the gap between those two things is where European AI strategy is currently weakest.

India’s situation is structurally different from Europe’s but arrives at a similar constraint from a different direction. India has AI talent (its diaspora is prominent in the leadership of every major American AI lab), growing compute investment, and a domestic market of enormous scale. What it lacks is the accumulated research infrastructure — the dense ecosystem of research labs, university programs, and corporate R&D that produces the regular stream of architectural advances that the field depends on.

The talent diaspora issue is particularly sharp for India. The researchers who trained at IITs and then completed PhDs at MIT, Stanford, and Carnegie Mellon, and then built their careers at Google Brain, OpenAI, and Meta — these are some of the most consequential people in the current AI development ecosystem. They are not, by and large, building India’s AI infrastructure. They are building American companies’ AI infrastructure. Whether that changes, and how quickly, depends on whether the Indian AI ecosystem develops enough to offer the research environment, compensation, and opportunity that keeps researchers who currently prefer the US.

The observation that “India’s AI talent is in California” is as accurate as the observation that “California’s AI talent is at a few companies on the Peninsula.” Talent concentrations are real phenomena with compounding effects. Breaking them requires sustained investment, compelling opportunity, and time.

The third way — AI independence for countries that are neither the US nor China — is not a fantasy. But it is a decades-long project, not a policy announcement or an investment round. The countries that will achieve meaningful AI sovereignty in the 2030s are those that are making the foundational investments now: in compute that gives domestic researchers and companies the ability to experiment and train; in research institutions that can attract and retain talent; in data infrastructure that makes domestic-language training feasible; in regulatory frameworks that protect the interests of domestic users without creating barriers to accessing the best available capabilities.

Macron’s “not the colony of either” aspiration requires more than aspiration. It requires the building of infrastructure that does not currently exist, at costs that European political will has not yet fully endorsed, over timelines that exceed electoral cycles. That is the arithmetic of sovereignty in the AI age.

It is not an unsolvable problem. It is a harder problem than the speeches make it sound.