Less Wood, It's War! What if the "third way" of AI spoke with a European accent?

In the legendary scene from Go West, Groucho desperately shouts: “More wood!” while they dismantle the train cars to feed the boiler. Today, observing the AI race between American hyperscalers and the Chinese state deployment, the déjà vu is inevitable. They are burning the train to be the first to reach a goal that no one has finished defining.

But, what if winning didn't require more fuel, but a better engine?

The Taalas factor: when numbers change the conversation

A few weeks ago, Taalas, a Canadian startup founded in Toronto in 2023, came out of stealth mode. Its proposal is radical: instead of running AI models on general-purpose GPUs, it fuses the model directly onto the silicon. The resulting chip is the model.

Their numbers are a quantum leap. Their first chip, the HC1, runs Llama 3.1 8B at 17,000 tokens per second per user. To put things into perspective, an Nvidia H200 (the most currently available on the market) reaches 230. But that's not all: the construction cost of the HC1 is 20 times lower, and its energy consumption is 10 times lower.

Perhaps the most astonishing thing is that they have reached this point with 24 people and $30 million in effective investment.

It is not optimization. It is a paradigm shift in the economy of inference.

The geostrategy that Taalas doesn't mention (but which exists)

Taalas is Canadian. And that matters more than it seems.

Europe and Canada share a framework agreement for technological collaboration (CETA), but above all, they share an uncomfortable strategic position: neither the American model of “surveillance capitalism” nor the Chinese model of “social control” is acceptable within their own institutional and business fabric.

If inference stops being a question of who pays Microsoft's or Google's electricity bill and becomes a matter of architectural efficiency, the map of European assets takes on a new prominence.

The European pieces of the puzzle

ASML. It manufactures the only machines in the world capable of producing chips on nodes below 7 nanometers. Without them, the Taiwanese TSMC would not exist as we know it, nor could it supply Nvidia with its current GPUs. It is the piece of the value chain that no actor can replicate in the short term.

ESMC (Dresden). The joint venture between TSMC, Bosch, Infineon, and NXP, with more than 10 billion euros invested and direct support from the European Chips Act. It is the first European foundry, although only capable of operating between 28 and 16 nanometers, which does not yet cover the needs for latest-generation AI chips. Its importance is given because it establishes a crucial operational and political precedent: Europe knows how to build chip foundries.

Mistral. From France, they have demonstrated that algorithmic efficiency competes with sheer size. Their language models obtain top-level results with a fraction of the parameters of their American rivals. This is not an academic curiosity: it is exactly the profile of a model that a Taalas-type architecture can profitably fuse onto silicon.

The strategic question that remains unanswered is whether the Chips Act 2.0, currently being debated in Brussels, will direct investment toward advanced nodes equipped with ASML technology. If that happens, the complete value chain—efficient model + specialized chip + sovereign foundry—would cease to be speculation and would provide Europe with unexpected competitiveness in that future that seemed so unattainable for us.

A window of opportunity, not a certainty

But it would be dishonest to present this as the anticipation of a certain victory. We must acknowledge that Taalas currently manufactures in TSMC Taiwan, and not in Dresden. That its benchmarks are internal, without independent validation. And that the business model—a new chip for every new frontier model—raises scalability questions.

But the strategic question is not whether Taalas will win. It is whether in Europe we have enough courage left to dare to be present when the inference market fragments from the training market. Because that fracture has already begun.

The question for politicians (and for citizens)

For institutional leaders, the question is more urgent and more concrete than any corporate budgetary decision. What does it mean, in practice, that the infrastructure that processes the health data of our hospitals, the communications of our defense systems, or the financial records of our companies resides on servers that comply with foreign legislation? And I fear that this is not a rhetorical question, but the current state of affairs. But the Chips Act 2.0 can become a political lever with sufficient substance to change it.

For citizens, the question is even simpler: do we prefer that the technology that organizes our lives—healthcare, justice, education, information—be subject to the strategic decisions of companies that answer to their shareholders in California or Beijing, or to institutions that, with all their flaws, are accountable to us?

Digital sovereignty is not a technical concept. It is a political decision that, if not taken consciously, someone will take for us.

The question for executives

For years, the fundamental technological decision in many organizations has been reduced to choosing between Azure, AWS, and GCP. Three American companies, and three models of equivalent dependence.

The question worth asking today is not how much cloud budget to allocate next year. It is whether we are paying attention to the window that is opening: an AI that does not require more fuel, but a different engine. More efficient, closer, and, for the first time in a long time, potentially more ours.

Perhaps, after all, it is not necessary to burn the train to arrive on time. We just need to stop copying the neighbor's engine and start manufacturing our own.