Arthur Mensch at the French National Assembly: Why AI is Electricity Turned into Tokens by FPDieulois :: 2026-05-17
On May 12, 2026, Arthur Mensch, co-founder and CEO of Mistral AI, delivered one of the clearest
Regulation, Sovereignty, and the European Choice Excellente intervention de Mensch cette semaine à la commission d’enquête sur la souveraineté numérique.
& most urgent testimonies yet on the strategic importance of artificial intelligence.
Speaking before a French National Assembly inquiry commission on digital dependencies and systemic vulnerabilities,
Mensch reframed AI not as abstract software, but as a heavy industrial process: the transformation of electrons (electricity) into tokens (units of intelligence).
His 1-hour-22-minute hearing, alongside Public Affairs Director Audrey Herblin-Stoop, was both technically sharp and strategically alarming.
Here is a comprehensive summary of the key messages, with special focus on the physical realities, energy scale, tokens, and the broader implications for Europe.
AI Is Not Magic — It’s an Energy-to-Intelligence Factory
Mensch’s central metaphor is powerful and physical: AI converts electricity into tokens.
A token is the fundamental unit that large language models process.
It is not exactly a word, nor a single character. Modern tokenizers (using techniques like Byte Pair Encoding) break text into statistically optimal chunks.
For example, Victor Hugo’s famous poem Demain, dès l’aube… (written after the death of his daughter)
is only a few hundred words long, but represents roughly 110–130 tokens for a model like Mistral.
The entire poem fits comfortably in the early part of a model’s context window.
Behind the elegant French you read sits a stream of numerical vectors — the “green code rain” of The Matrix, or the underlying reality a replicant like Roy Batty might perceive.
Every multiplication and attention operation on these token embeddings consumes real electricity.
Thanks to massive progress in cooling technologies (direct liquid cooling, immersion, better facility design),
a much larger share of each electron now goes to actual computation rather than waste heat.
Power Usage Effectiveness (PUE) in modern AI data centers has dropped toward 1.1–1.3, meaning the vast majority of power delivered to the building powers GPUs, not air conditioners.
This makes AI a true energy-intensive industry, comparable to aluminum smelting or data-heavy manufacturing.
The Gigawatt Scale: France’s Window of Opportunity
Mensch highlighted concrete numbers that should wake up European policymakers:
France has an average surplus of about 9 GW of low-carbon (mostly nuclear) electricity.
Future AI clusters are measured in gigawatts.
A single 1 GW site represents an enormous concentration of compute
— hundreds of thousands of high-end GPUs (H100/H200/B200-class chips consume 700 W to over 1 kW each, before cooling and losses).
At European scale, he foresees the possibility of roughly 1 kW of AI power per person within a few years
— translating to tens of GW for France and hundreds of GW continent-wide.
These figures are not science fiction.
They represent the new industrial race. Building a 1 GW AI cluster costs tens of billions of euros in GPUs, power infrastructure, networking, and specialized facilities.
The interconnection density and low-latency requirements favor large, well-planned sites.
The Physical Limits Are Real
Mensch repeatedly grounded the discussion in material reality
— crucial for anyone who has worked in electrical machines, power electronics, or hardware manufacturing:
Semiconductors: Advanced nodes (3 nm, 2 nm, soon smaller) are extremely concentrated, primarily in Taiwan (TSMC).
Supply is tight and geopolitically vulnerable.
Helium: Essential for semiconductor manufacturing processes and certain cooling/test environments.
Global supply is constrained.
Rare earths and critical materials: Neodymium, dysprosium, and others are needed for magnets, capacitors, and high-performance components.
China dominates processing.
Power electronics: High-efficiency converters, transformers, and distribution systems become critical at this scale.
Every percentage point of efficiency gained or lost has massive financial and energetic consequences.
Progress in liquid cooling has dramatically improved the percentage of electrons that reach computation, but the supply-chain vulnerabilities remain hard physical constraints.
Europe cannot afford to be purely dependent on foreign supply for any of these.
Hyperscalers: The Silent Land Grab
One of Mensch’s strongest warnings concerns American hyperscalers (Microsoft, Google, Amazon, Meta, etc.).
They are signing long-term contracts with European utilities — including EDF — to reserve massive blocks of power.
Their global CapEx plans run into trillions of dollars.If Europe does nothing, French and European electrons will be transformed into American tokens.
The energy stays local, but the economic value, IP, data advantage, and strategic control flow abroad.
Europe risks becoming a low-value “energy battery” for foreign AI empires — a vassal continent in the intelligence age.
The Transformation of Work: From Coders to Managers of AgentsInside Mistral today, software engineers no longer write most code line by line.
They have become managers and orchestrators of AI agents. They set specifications, direct specialized agents, review, debug, ensure security, and validate outputs.
This shift already delivers productivity gains of 5× to 10× on many tasks.
At Mistral, AI consumption already represents roughly 10% of total payroll — about €10,000 per employee per year (equivalent to roughly 1 kW of continuous GPU power per person).
However, Mensch is nuanced on scalability. In small, agile teams the gains are spectacular.
In large organizations, organizational bottlenecks (meetings, alignment, validation, culture) quickly absorb individual productivity improvements.
AI can help reduce these frictions over time, but it requires deliberate redesign of companies and processes.
Toxic middle management may be exposed and reduced, but orchestration, accountability, and high-stakes decision-making remain fundamentally human.
Mensch was sharply critical of Europe’s regulatory stack (AI Act, GDPR, copyright rules, etc.).
While well-intentioned, the cumulative burden favors large American players who can afford massive compliance teams.
European startups pay a heavy overhead.His recommendation is pragmatic rather than protectionist:
Prioritize European champions in public procurement, especially for defense and critical sectors.
Build local data centers that capture R&D, jobs, and value.
Use France’s nuclear advantage aggressively while the window remains open.
Treat AI infrastructure as strategic national capability, similar to energy or semiconductors.
Conclusion: A Historic Window Closing
Arthur Mensch’s testimony was not alarmist for effect — it was a sober engineer’s warning.
AI is not coming; it is already an energy-hungry, material-intensive industry that will reshape economies and geopolitics in the coming years.
Europe has unique cards: decarbonized electricity, strong engineering talent, and companies like Mistral.
But the window is narrow — measured in months to a couple of years.
If Europe fails to build its own GW-scale AI infrastructure, it will import intelligence at massive cost, lose strategic autonomy, and watch value flow westward.
The choice is clear: become producers of tokens, or remain passive consumers of foreign intelligence.
As Mensch made clear, the electrons are already here.
The question is who will turn them into intelligence — and who will capture the value.
This hearing deserves to be studied carefully by policymakers, engineers, and citizens across Europe.
The age of intelligence is also the age of energy and materials.
Arthur Mensch has drawn the map.
Now comes the hard part: building.....
Il aborde à peu près tous les thèmes avec brio, et expose parfaitement les enjeux.
C'est à se taper la tête contre les murs qu’il n’y ait eu qu’une poignée de personnes…
< B > My Personal Blog < B > List of all my Blog posts < B >
Article generated & inserted in SQL tables by AIs
FREDDY FRANCE Blog personnel
(c)FPe COPYRIGHT @FPDIEULOIS 2026