Europe's Power Play

Last week, Denmark's newly appointed energy minister called a crisis meeting. It was long overdue. Electricity has become a scarce resource, and it will stay that way for a long time. We are entering a period where AI creates such an acute shortage of power that it will force an unprecedented reckoning with how we, as a society, distribute the little we have. 

The reason is that the world has stepped into the next phase of the AI race, the inference era, where our consumption of and demand for AI is a steeply rising curve. That demand feeds a structural problem that reaches well beyond Denmark. It is a geopolitical race, and the investments being made right now are of historic proportions.

The problem, which I also address in my video AI and the Stack of Decisions, is that however you slice the numbers, the EU's investments are dwarfed compared to the astronomical sums the American hyperscalers are pouring into AI infrastructure. For every million euros we invest, they invest 300 times as much.
Europe cannot win a race measured in cables and kilowatts. The more useful question is whether that is even the right race to run? The argument of this essay is that Europe's constraints, paradoxically, may be its most important strategic asset.

Explore the interactive insights on Power Play microsite

Where does the pressure come from?

To understand what is happening right now, we need to consider the bottlenecks and the commercial pressure that create the current situation, which is full of complex dynamics.

We know that token prices halve roughly every other month, and from that a reasonable assumption would be that AI is getting cheaper. Unfortunately it is more complicated than that. While prices drop the business built on AI consumption at Anthropic and OpenAI is climbing by a factor of 10. A large part of this is driven by agentic workflows that can run around the clock. Where a single user action once triggered one API call, it now triggers a chain of calls. Volume grows faster than price falls. The total AI bill rises, even as the price per token sits at a record low.

But that is not the full picture. This pricing regime, which European companies budget against today, is subsidised by a volatile mix of venture capital, hyperscaler partnerships, and deliberate strategies to capture market share. Leaked financials on OpenAI recently disclosed a projected loss of 20 billion dollars in 2026. That’s the equivalent of the GDP of a small country like Malta. The losses are budgeted precisely to capture market share, so the investment pays back over the long run. Investor patience has limits. When it runs out, the subsidy does not end. It finds a new sponsor.

Raw power or energy intelligence 

In other words, Europe shares this interest in eliminating bottlenecks with everyone else. It sits high on the agenda everywhere. The token economy rests on a foundation of suppressed costs, bottlenecks, and borrowed time but the players have radically different approaches. 

Brute force remains the strategy of the US and China. Build more data centres, generate more power, deploy more chips. Build, baby, build. It is the only strategy that truly works in the short term, which is why the capital flowing in can be measured in historic investments of billions of dollars annually. For Europe it is the least likely path to success. The infrastructure gaps, regulatory timelines, and capital disadvantages across 27 countries are substantial, and they resist removal by tomorrow's new directive, however determined the attempt. 

Energy intelligence is the European counter-move. Rather than chasing the competition on scale, the goal is to extract more intelligence per watt through software efficiency, flexible grid integration, mixed-precision computation, and optimised hardware. In April 2026 the International Energy Agency (IEA) made clear that policymakers must give data centres incentives to operate "flexibly". The IT and grid architects of the future must master what the IEA calls smart integration. Data centres should move beyond being passive power guzzlers and gain the intelligence to throttle heavy AI training tasks during the hours when the public grid is at maximum load, and to accelerate when local wind or solar energy runs in surplus. It works as a design philosophy more than a single technology: compute with energy awareness rather than compute until the power runs out. The hope is that a Europe forced to innovate rather than scale will have its "DeepSeek" moment, where China showed a new path for model architecture because it was cut off from access to the most advanced chips.

Three power hogs

If Europe's path forward runs through efficiency rather than scale, it matters enormously to understand exactly what is inefficient and why. Three structural weaknesses drive the current power consumption, and all three are, in principle, attackable. They are consequences of choices made when energy was cheap and scaling was the only strategy anyone needed.

1. The memory wall

One of the things discussed a great deal in Silicon Valley is what they call the memory wall. Every current chip architecture builds on the same principle, the Von Neumann architecture from the 1940s, which separates compute from memory. It corresponds to the processor and RAM in a modern computer.  

A large AI model carries its entire knowledge base in memory at all times, roughly the equivalent of 500 GB of densely packed information. Every time the model produces a single word in a response, it must read through that entire knowledge base to decide what comes next. Amazingly the bottleneck does not sit with the billions of simultaneous calculations as you might expect. It is in speed at which memory can deliver the data. And since modern AI chips can calculate far faster than memory can supply them, the chips spend most of their time waiting. Scale that across millions of simultaneous conversations, and you have a structural energy and capacity problem built into the hardware itself

That waiting time has a direct commercial consequence. Today, memory is the single most important constraint on how many tokens the world can produce per second. That is why High Bandwidth Memory (HBM), the on-chip RAM that makes fast AI possible, is sold out at every major manufacturer through the end of 2026. Memory prices have roughly doubled since the start of 2025, and data centre capacity is absorbed faster than new capacity can be built. Samsung and several other manufacturers have collectively redirected 93 percent of their production to AI data centres. 

A chip spends up to 70% of its energy moving data back and forth rather than computing. That is an enormous overhead built into the entire current generation of AI hardware. And it serves you badly when energy is scarce. All of it because of a chip architecture that has remained structurally unchanged for 80 years.

2. The tokenizer overhead

Language models do not read words. They read tokens: small text fragments produced by a data compression algorithm called Byte Pair Encoding, invented in 1994. In the newer AI models, built predominantly in the US and China, the result is that English and Chinese respectively compress efficiently, while other languages require considerably more tokens to represent the same content. In Nowable’s Great Token test from June 2026, Danish required roughly 50 percent more than equivalent English content. Other languages such as Asian or Arabic ones generate 4 to 10 times as many tokens. That is 4 to 10 times as many calculations, which call for corresponding chips and power. 

This is a direct consequence of the way AI models are designed and trained today, and from the US/China concentration of power we are seeing in the field right now. As of June 2026 Google Search is now switching to become AI-generated across 200 countries. That is an estimated 2.5 billion of people AI-searching everyday in their native languages. An upsurge in inference, which will surely claim its share of global energy resource.

 3. The transformer architecture

The third bottleneck is the architecture that drives virtually every AI model in production today. The transformer, introduced in 2017, has proven extraordinarily good. It has also proven extraordinarily hungry. The more you scale on data and compute, the better the models get.

Some of that hunger comes from a structural limitation: transformers have no persistent memory. Every conversational turn processes the whole conversation again, which means the transformer depends deeply on on-chip memory to deliver a coherent experience. The cost grows quadratically, and when you finish a conversation that nets out at 6,000 tokens, 63,000 tokens have passed through the data centres.

Algorithmic improvements help offset this in part. Epoch AI's analysis from February 2026 estimates software efficiency gains of a factor of 10 each year and other measured gains appear to come from better training data. But those are still just tweaks.

The transformer was one of the rare leaping innovations, and a comparable leap has eluded us for 9 years. That means the very architecture that has given us such strong AI models is an inseparable part of the cost problem. 

Towards analogue chips and a tokenfree future

Taken together, these three bottlenecks describe a paradigm optimised for abundance: of energy, capital, and English-language data. Europe operates under none of those conditions. That is a disadvantage on the current terms. On different terms, it is a design brief: three concrete problems to solve, a set of real constraints, and precisely the kind of pressure that has historically produced genuine innovation.

Neuromorphic chips represent the most ambitious near-term hardware breakthrough. The ultimate goal of energy intelligence is to imitate biology. In spring 2026, researchers from the University of Cambridge published a landmark study in Science Advances presenting a new, brain-inspired chip material (hafnium-based memristors). The chip is analogue by design and is effectively breaking down the memory wall by integrating memory and compute.

This is the reason why it is expected to cut AI hardware's energy consumption by up to 70% by processing and storing data in the same place. Our human brain can solve complex AI tasks on under 20 watts, while data centres require megawatts. Europe's great deep-tech strength could become designing software and industrial robots (Physical AI) built for this neuromorphic architecture. Commercial rollout sits some years out, but the materials, hafnium oxide, already feature in chip production, so the road from lab result to production integration is shorter than it would otherwise be.

Token-free architectures take a different approach to the tokenisation architecture by eliminating it entirely. Meta's Byte Latent Transformer (BLT) was extended with significant improvements in May 2026 by Meta and Stanford researchers. The method operates directly on raw bytes and drops the fixed tokeniser as a compression intermediary. The model groups bytes dynamically, assigning more compute to regions that are harder to predict and less to predictable ones.

In controlled scaling experiments up to 8 billion parameters, BLT matched the performance of token-based models and reduced memory bandwidth by over 50 percent. Crucially, because BLT has no fixed vocabulary, it also has no structural reason to treat English differently from Danish or Finnish. The language skew is created by tokenisers. Remove them and you remove the mechanism that creates the tax. 

Tokeniser parity. While we wait for token-free architectures, several European researchers has succeeded in isolating the tokeniser from an AI model, training it on European languages to reach parity, meaning 1-to-1, with English, and then reintroducing it into the AI model so the language tax disappears. This kind of modular AI surgery is remarkable for several reasons.

Partly because it reduces token consumption while keeping the model intact, partly because it creates a genuine alternative to the enormously power- and resource-hungry process of finetuning a model to produce an equivalent adapted local language model. The prospects are exciting if the method proves able to scale, since on open models you could in future simply swap and retrain a tokeniser each time a new model arrives, and start fresh from there rather than from scratch.

World models and radically different architectures operate on the longest time horizon while pointing toward the most fundamental rethink. Yann LeCun's AMI Labs is the most striking recent signal that the architectural rethink is moving from theory toward capital. Born in France and long the chief AI scientist at Meta, LeCun departed in 2025 and returned to Europe to found AMI Labs in Paris, raising one billion dollars in 2026 to pursue a fundamentally different vision of intelligence. He argues that current vision-language models waste large amounts of resources by failing to prioritise them intelligently. For example, a self-driving car spends as much energy decoding leaves on a tree as reading the traffic.

His JEPA architecture builds on world models rather than language models, learning abstract representations of how the world works and predicting meaning rather than pixels. In experiments so far it has proven to require markedly less compute. In the US, Stanford researcher Fei-Fei Li, also called the Godmother of AI, has created World Labs, pursuing the same intuition from a spatial angle with her project Marble. Both leave traditional tokens and transformers behind entirely. If architectures like these mature, the entire cost structure beneath today's AI consumption is potentially turned on its head.

What it means across three time horizons

 The near future, the next two to three years, belongs to the race for infrastructure. Whoever secures power and compute fastest will run the largest models. For European nations and companies, this period is navigated best by understanding the existing cost structure precisely: which models carry the lowest overhead for the languages we use, how to design the harnesses and architectures that optimise AI consumption, and how to treat your AI stack as a strategic procurement decision rather than a default. Parity-aware tokenisation, the modularly trained tokenisers, may become a standard option for multilingual model training.

The medium horizon, three to seven years, is where efficiency becomes competitive. Neuromorphic hardware will begin to appear in specialised industrial contexts. Token-free architectures will move from research labs toward production systems. European deep-tech investment in energy-efficient AI hardware and software will bear fruit and compete on the terrain where the gap to US and Chinese infrastructure spending is smallest. 

The long horizon, beyond seven years, is the most uncertain and the most decisive. If world models achieve what their architects believe they can, the entire cost structure of AI changes. Inference that today requires megawatts because it leans on memoryless transformers across vast redundant conversational contexts could be replaced by a new and more efficient architecture that leans on learned internal world models, which carry their understanding forward instead of rebuilding it from scratch each time. 

Europe did not win the first round of the AI infrastructure race. But the first round was fought on terrain chosen by others: abundant capital, abundant energy, and an architecture built for English. The next round will be fought on terrain that is still being shaped. The constraints in the current paradigm are real, and they are many. So is the potential that follows from solving them. Whether Europe captures that potential depends on capital, coordination, and competence at a scale the continent has yet to demonstrate. The next paradigm is not yet owned by anyone. Europe should treat that as the emergency it is.