Europe’s
Power Play

Understand the three constraints that make AI a power hog - and four bets that could change it.

This is what is driving the energy crunch.

Europe’s entire five-year plan is less than what one hyperscaler spends in five weeks.

The EU has committed €20 billion to build AI gigafactories across five years. Google, Microsoft, Meta and Amazon have each allocated between $135 and $200 billion for 2026 alone.

Token prices fall. The total AI bill rises. Here is why.

Volume grows faster than price falls. Agentic AI runs around the clock. The cheaper AI gets per token, the more the world consumes. Consumption drives the demand for more compute up - and that means more power.

Three structural problems are built into every AI system running today.

The memory wall

A chip spends up to 70% of its energy moving data - not computing. The bottleneck is memory speed, not processing power. An architecture unchanged for 80 years.

Tokenisation

Models read fragments called tokens, and non-English pays a tax. The same legal clause costs 51% more tokens in Danish than in English - some languages need 4 to 10 times as many. More tokens means more chips, more watts.

The architecture that forgets everything

Every conversational turn reprocesses everything from scratch. A conversation producing 6,000 tokens pushes 63,000 through the data centre. The cost grows quadratically.

Four bets on a more efficient AI.

The new path Europe should be leading.

Modular AI surgery

European researchers has found a way to eliminate the language tax on open models. Detach the tokeniser. Train it for European languages. Reattach. Same model, no language tax.

Goodbye token economy - hello token-free future

Meta and Stanford researchers invented Byte Latent Transformer (BLT), which could replace the tokeniser. It operates directly on raw bytes, with no fixed vocabulary - and no structural reason to treat English differently from any other language. BLT cuts memory bandwidth by over 50%.

The chip that breaks the wall

Cambridge researchers have built the Memristor, an analogue chip that puts memory and compute in the same place - the way the brain does - on under 20 watts.

Models that understand the world

Yann LeCun left Meta, returned to Europe, and founded AMI Labs in Paris, raising one billion dollars in 2026. His bet: AI that models how the world works, not just the next token - world models could cut the energy of AI video by up to 75%.

The next paradigm is not yet owned by anyone.

Europe should treat that as the emergency it is.

AI and the Stack of Decisions

Get nowable in your inbox.

AI, energy, and the architecture of what’s next. A few sharp reads a month, no noise.