Among the hundreds of conferences devoted to artificial intelligence, some reveal not so much an innovation as a shift in doctrine. The discussion with Aidan Gomez, co-founder of Cohere and co-author of the seminal paper “Attention Is All You Need,” falls into this category.
For the past two years, the dominant narrative surrounding AI has centered on a race for power: more GPUs, more parameters, more capital. Yet behind the spectacular announcements from American tech giants, another battle is unfolding—one that is less visible but perhaps more strategic: the battle for operational sovereignty.
The Return of Politics to Artificial Intelligence
The day before his speech, Aidan Gomez attended a G7 meeting bringing together heads of state and leaders of major AI companies.
The key point is not so much the presence of industry leaders as the consensus that seems to be emerging: Western democracies now view artificial intelligence as a strategic infrastructure comparable to energy or telecommunications.
For several years, Europe has approached AI primarily from a regulatory perspective. Today, the conversation is shifting. The issue is no longer just about regulating AI, but about controlling part of the value chain.
Gomez puts it simply: it is essential for a democracy to occupy at least second place globally in the AI hierarchy.
Behind this statement lies a deep geopolitical concern. If these models become the cognitive infrastructure of government agencies, hospitals, power grids, or banks, dependence on a few non-European players becomes a matter of national security.
The End of the “Everything Is Open” Myth
One of the most interesting developments concerns the partnership between Cohere and the German company Aleph Alpha.
This partnership illustrates a transformation in the market. For a long time, the European ecosystem spread its resources across a multitude of promising startups. However, according to Gomez, modern AI requires a concentration of capital, talent, and infrastructure rarely seen in previous technology cycles.
The logic is therefore changing.
The goal is no longer to fund a hundred different initiatives in the hope that one will succeed. Instead, it is to create a few champions capable of reaching global critical mass.
This approach is more reminiscent of the European aerospace industry and the creation of Airbus than the traditional startup world.
The Real Battle: Critical Infrastructure
Cohere’s positioning is particularly revealing of the market’s evolution.
Unlike other players focused on consumer applications, the company targets critical infrastructure:
- energy;
- telecommunications;
- health care;
- finance;
- public administration.
In these environments, the main issue is not the models’ creativity but their reliability, security, and ability to operate in sovereign environments.
This distinction is fundamental.
Media debates often focus on conversational assistants. Yet the most important markets likely lie elsewhere: in systems that manage power grids, banking operations, or government infrastructure.
When these systems fail, the economy grinds to a halt.
The Silent Gap Between Demonstration and Deployment
The article also highlights a rarely discussed phenomenon.
Most of the most advanced models are designed to demonstrate what is technically possible. But companies must deploy what is economically viable.
Gomez explains that many customers simply cannot run models that are constantly growing in size and cost.
Cohere’s true competitive advantage, therefore, isn’t raw power.
It’s efficiency.
Being able to run a model at one-tenth to one-thirtieth the cost becomes more valuable than gaining a few points on an academic benchmark.
This logic echoes a classic rule of the luxury industry: invisible sophistication often counts for more than ostentatious display.
From POC to ROI
For the past eighteen months, one question has been on everyone’s mind across all executive teams: where is the return on investment?
The prevailing narrative has long suggested that companies struggle to turn their experiments into economic value.
Gomez offers an interesting perspective.
According to him, the problem lay not in the technology but in the maturity of the organizations. The first waves of Proofs of Concept were often poorly defined, inadequately governed, or disconnected from business needs.
Today, the situation is changing.
Companies are better able to identify high-value use cases. Vendors have a better understanding of industry constraints. The models are more reliable.
As a result, pilot projects are beginning to move into production.
This transition from POC to industrial deployment is likely one of the most significant changes observed in 2026.
The Employment Paradox
On the issue of the impact on work, Gomez takes a measured stance.
At this stage, Cohere observes more instances of augmentation than replacement. Organizations are using AI to accelerate their growth rather than to drastically reduce their workforce.
This analysis is consistent with what we see in many sectors: productivity gains are real but are largely offset by an increase in demand.
The executive also refers to the Jevons paradox, according to which a technology that greatly improves efficiency may ultimately increase consumption rather than reduce it.
Economic history is replete with similar examples.
However, there is no guarantee that this trend will remain stable in the long term. Gomez acknowledges that governments will need to anticipate more complex scenarios involving training, retraining, and new forms of social protection.
Acceptability Is Becoming a Strategic Issue
One of the most insightful passages concerns the growing distrust of AI.
Unlike some tech executives who view this resistance as merely a communication problem, Gomez believes it is legitimate.
Concerns about employment, energy consumption, or economic concentration do not stem from a lack of public understanding.
They reflect a genuine uncertainty about the future trajectory of technology.
This acknowledgment is important.
Social acceptability could become just as crucial as technical performance itself.
A global cognitive infrastructure cannot be imposed on the citizens who use it.
The Details
Perhaps the most surprising observation from the interview concerns the Transformers themselves.
Nearly ten years after the publication of *Attention Is All You Need*, the architecture remains the foundation of virtually all major models.
For Gomez, this longevity is almost unusual.
In the history of computing, few paradigms have dominated for so long without being replaced. His hope, therefore, is not to see Transformers triumph indefinitely, but to witness the emergence of a new generation of more efficient architectures.
A remark that serves as a welcome reminder: despite the hundreds of billions invested in AI, we may still be at the very beginning of this story.
Cette publication est également disponible en :
