The G7 summit in Italy was not just attended by President Trump. Neither were Macron, Scholz, and Meloni alone. Alongside them, key figures such as Sam Altman from OpenAI, Dario Amodei from Anthropic, Sundar Pichai from Google, and Demis Hassabis from DeepMind were present during the meetings at the historic Palazzo Ducale in Genoa. Importantly, they were not there as observers or sidelined advisors. Their presence at the main table signals a paradigm shift: when did heads of state cease to be the only ones capable of shaping geopolitics?
Photo: Igor Omilaev on Unsplash
This G7 summit sets an unprecedented precedent in the recent history of international diplomacy. At past negotiations, like the Kyoto Protocol, CEOs from oil companies did not attend alongside their leaders, nor did central bankers participate directly in Bretton Woods. However, in 2026, leaders from the companies that dominate critical technology are sitting next to those who, in theory, are meant to regulate them. What's at stake is not just who writes the rules for AI, but who truly has the capacity to implement them.
The Changing of the Guard: When Technical Power Surpasses Political Power
The presence of AI CEOs at the G7 is not merely a diplomatic courtesy. It is a clear acknowledgment of a reality that governments have been slow to accept: the power to develop and control cutting-edge AI lies in the hands of a few private companies, not in state institutions.
For instance, OpenAI controls GPT-5, considered the most advanced model for general reasoning. Anthropic has Claude 4, which excels in interpretability and safety. Google owns Gemini Ultra 2.0, along with the most advanced TPU infrastructure. DeepMind has progressed so much that it is three years ahead in predictive proteomics. These are not just competitive advantages; they are technical capabilities that no government can replicate in the short term, even with unlimited budgets.
France made an attempt with its sovereign AI initiative in 2024, investing €5 billion in Mistral and public research centers. However, two years later, Mistral has just announced a partnership with Microsoft to scale its models. Germany launched Aleph Alpha with €500 million in state investment, but today it relies on Anthropic models for critical services. Meanwhile, the UK promised to create its own AI "national champion," but ended up regulating instead of building.
The lesson is crucial: building cutting-edge models requires not only capital but complete ecosystems of talent, computing infrastructure, and organizational culture that cannot be improvised. Governments can fund basic research, but it is private companies that are uniquely capable of turning that research into real-world systems on a global scale.
The Hidden Agenda: Cooperative Regulation vs. Regulatory Capture
Photo: Luke Jones on Unsplash
The official narrative of the summit mentions "collaboration between governments and industry to ensure the safe development of AI." The curious thing is that the reality is more complex and, honestly, more interesting.
On one hand, these CEOs are actively calling for regulation. Altman clearly stated he seeks mandatory licenses for training models that exceed certain computational thresholds. Amodei has proposed "constitutional safety" frameworks that require external audits. Hassabis, for his part, advocates for binding international treaties on military AI.
Is it altruism? Perhaps in part. But it can also be viewed as a business strategy. Regulation that requires licenses to train models creates unbeatable barriers to entry for other competitors. If you need government approval to use 10^26 FLOPs, only companies with established relationships with regulators will be able to compete. Meta tried to train Llama 4 without the required certifications, but after pressure from the U.S. Department of Commerce, it abandoned the project.
On the other hand, governments are aware of their dependence on these companies. Not only to maintain technological leadership against China but also to implement AI in areas such as defense and public health. The Pentagon relies on Google and Anthropic for intelligence analysis, while the UK’s NHS uses OpenAI models for medical triage. This interdependence creates what political scientists call "reverse regulatory capture": the industry and governments become mutually dependent.
China, the Absent (Yet Omnipresent) Elephant
In the G7's official statement on AI, China is mentioned 17 times. However, companies like ByteDance, DeepSeek, and Baidu are not named. This omission is deliberate and reveals a central tension throughout the summit: how can one maintain leadership in AI when your main geopolitical competitor does not have to follow the same rules?
China does not have "AI companies" in the Western sense. It has "companies with AI capabilities operating under the strategic direction of the State." This distinction is key, as Beijing can coordinate resources, data, and talent in ways that would be illegal or unfeasible in liberal democracies.
DeepSeek, the Chinese government-backed AI lab, has just published a study on distributed training, which reduces computing costs by 40%. ByteDance has integrated generative AI into TikTok in ways that violate European regulations, but that doesn’t matter because TikTok does not seek European approval; it seeks market penetration.
The CEOs present at the G7 understand this asymmetry perfectly. Altman noted that "competing with actors who have no restrictions on privacy, transparency, or security is like running a race with weights on your ankles." However, they know that using this argument to soften regulations in the West is politically impossible after the Cambridge Analytica and Clearview AI scandals.
The solution seems to be a two-speed system: strict regulation for consumer applications and almost total freedom for "national security" applications. Google can use your search history to train models for the Pentagon but not to sell you ads. Anthropic can process medical data without consent if it's for biodefense research, but not for commercial diagnostics.
This double standard is not new in dual-use technology, but AI amplifies it because the same technology can serve both purposes. The model that analyzes proteins to develop vaccines can also design pathogens, and the system that detects misinformation on social media can generate it on a large scale.
The Precedent Set by This Summit
What is most significant about the G7 in Genoa is not just the agreements reached (which are vague and non-binding). It is the institutional precedent that establishes that cutting-edge tech companies are now key geopolitical actors, with a voice in forums traditionally reserved for sovereign states.
The World Bank does not invite Goldman Sachs to its annual meetings, despite Goldman possibly having more impact on capital flows than many member countries. OPEC does not include ExxonMobil in decisions about production quotas. The WTO does not give Amazon a vote in trade negotiations. However, the G7 does include OpenAI, Anthropic, and Google in discussions about the future of AI.
The difference is critical: these companies are not just important economic players. They are the only ones with the real technical capacity to implement or block any regulatory framework that governments design. If the G7 decides that all AI models must include cryptographic watermarks to detect synthetic content, only these companies will be able to carry it out. If the international treaty on military AI requires security audits before deployment, only these companies have the necessary technical knowledge.
This technical dependency translates into political power. And the G7 in Genoa formally recognizes this. It is the 2026 equivalent of when oil company CEOs were included in discussions on energy security in the 1970s.
What This Means for the Rest of the Ecosystem
For founders and developers who were not at the G7 table, this shift has concrete implications.
First: The companies present will define the de facto technical standards that the rest of us will have to follow. If OpenAI implements a watermarking system in its APIs, every startup using GPT-5 will have to incorporate it. If Anthropic requires specific safety protocols to access Claude 4, there will be no viable alternative for enterprise applications.
Second: The window for building independent alternatives is closing. Mistral, Cohere, and Inflection emerged promising to be counterweights to the giants. However, two years later, Mistral is now partnered with Microsoft, Cohere relies on government contracts, and Inflection was acquired by Microsoft. The pattern is clear: either you rapidly develop into a strategic player, or you end up being acquired or marginalized.
Third: Open-source AI faces an existential challenge. Meta continues to release Llama models but with increasingly severe restrictions. Stability AI is in financial trouble, and Hugging Face is profitable but cannot compete in training cutting-edge models. If the regulation emerging from the G7 favors auditable models, open source will be limited to non-critical applications.
Fourth: For European and Latin American startups, dependence on U.S. infrastructure becomes a geopolitical issue. If your product relies on APIs from OpenAI or Anthropic, you subject yourself not only to their terms of service but also to U.S. foreign policy. When Washington imposed restrictions on AI chip exports to China, those restrictions automatically applied to services based on those chips. Your startup in Barcelona or Buenos Aires became subject to the geopolitics of Silicon Valley, without having chosen it.
The Question No One Is Asking
All analyses of the G7 summit focus on the agreements and proposed regulations. But the more important question is: what conversations took place in the closed-door meetings that we didn't have access to?
What was publicly shown was a carefully choreographed theater. Press statements about "responsible AI development" and "international cooperation" camouflage the real decisions about who can train which models. The red lines, the negotiations between security and competitiveness... those talks happened in private.
In those meetings, the balance of power is different from what is observed at the official table. Heads of state have democratic legitimacy, but AI CEOs possess technical leverage. A prime minister can threaten regulation, but if that regulation makes competing with China impossible, the CEO can simply respond, "then we’ll lose," diluting the threat.
It is the technological equivalent of what economists call "too big to fail," but it's even deeper. These companies are "too critical to regulate meaningfully." You can impose fines or superficial restrictions, but you cannot fragment them without ruining the technical capabilities you need for national security, economic competitiveness, and digital infrastructure.
The G7 summit in Genoa will be remembered not for the specific agreements it produced but for the institutional precedent it established. It marked the moment when governments formally recognized that the power to shape the future of AI resides in the hands of a few tech companies.