Europe’s plan to catch up in AI largely rests on building better machines; bids to host up to seven EU-backed AI gigafactories, giant data centers that would each run between 25,000 and 100,000 top-end chips, close on November 12, and up to €10 billion in public money is on the table.
Brussels expects at least €20 billion more from private investors.
Zuzanna Stamirowska thinks the more consequential race is happening somewhere else. Speaking to 150sec at the One Way Summit in San Francisco on October 29, the Polish co-founder and CEO of AI lab Pathway argued that new model designs needing far less computing power could change who gets to build, own and benefit from AI—including businesses that want a model of their own and people priced out of today’s systems.
Pathway has offices in Palo Alto, Paris and Wrocław, and reached a $500 million USD valuation in August after bringing its total seed funding to $30 million. Its bet is Dragon Hatchling (BDH), a post-transformer architecture built to keep learning from smaller datasets without being retrained from scratch.
Think of most AI models today as a printed encyclopedia: everything they know is fixed the moment they were printed, and any new information has to wait for the next edition—essentially retraining or fine-tuning the model at great expense.
A chatbot can scribble notes in the margins during a conversation, but they are gone once the chat ends. BDH is designed to work more like an encyclopedia that rewrites its own entries as new information comes in, without waiting for the next print.
In practice, this looks like a colleague who keeps learning on the job, or a security model which could pick up a new cyberattack the first time it appears on a company’s network without waiting the months it usually takes to retrain. A model watching financial markets could notice the early signs that conditions are changing, the moment Stamirowska said traders most want to catch, instead of rebuilding every single new signal through what worked last year.
Or, in healthcare, a model could adjust as doctors log cases that don’t fit the textbook pattern, such as heart attack symptoms that show up differently in women, instead of repeating the gaps it was trained on.
The idea behind BDH, Stamirowska explained, started with ships. Her doctoral research at the Complex Systems Institute of Paris used 30 years of daily movements of the world’s entire merchant fleet, and she found that trade between two ports could be predicted with a single measure: the number of connections they share.
“It’s like Facebook. If you have many friends in common with someone, it’s very likely that you’re also connected,” she said. When her team studied how attention works in BDH, “it literally works like a routing mechanism in transportation.”
The case for smaller models
Stamirowska’s pitch rests on cost: transformers, the architecture behind most of today’s leading AI systems, are what the gigafactories are being built to run.
Introduced by Google researchers in 2017, the transformer is the design underneath ChatGPT, Gemini and Claude. It is powerful yet hungry,, as the bigger the model and the more text it handles, the more processing power it needs.
A “post-transformer” model like BDH tries to keep the intelligence while dropping much of that cost. “At the same level of models, a couple of billion parameters, you will already get a huge cost advantage with post-transformers,” she said, “and way more reasoning power.”
Parameters are the internal settings a model adjusts as it learns, and their number is a rough measure of its size. A model with a couple billion is small by today’s standards, with the largest systems thought to be running on trillions or more, though their makers don’t disclose the figures.
Pathway says it publishes what its result actually costs to run. “We report the actual compute cost. There is no hidden cost. We’re not subsidizing this,” the founder explained. “The other folks are evaluated on their reported prices. You can do price cuts. [It] depends on your unit economics.”
On some benchmarks, she said, BDH’s costs are “ridiculously small, to the point that we need to shift the y-axis a little but to still show how little that is,” adding that DeepSeek “is not even close to that, and DeepSeek is the cheapest, most efficient that you can get… and this is a V1!”
In a published test, BDH solved 97.4% of Sudoku Extreme puzzles without writing out its reasoning step-by-step or calling on outside tools, as many chatbots do to solve hard problems. Problems like these, where every answer has to obey several rules at once, are where Stamirowska says the model performs best. “Capability is mostly on the reasoning side, where it shines.”
A model that keeps leaning raises an obvious question: what happens when it learns the wrong thing? Because BDH needs less data, Stamirowska explained, it can pick up when the rules of a system change.
“You’re not so much fixed, like, oh, I’m in this regime forever, and I just draw everything to that,” she said, adding that its neurons also tune out repetition, much as a person stops hearing the hum of the fridge.
“The neuron will literally tune out because it’s hearing the same thing. It’s more looking for those signals that would contradict its worldview.”
Who gets access
In June, 28.8% of working-age people in the Global North used generative AI tools, compared with 16.2% in the Global South, according to Microsoft’s AI diffusion report, and the gap is still widening. Bottom line, the benefits of AI are spreading unevenly.
The money and decision-making behind the technology are even more concentrated. Just 100 companies, mostly in the U.S. and China, account for 40% of global corporate research spending, per the UN Conference on Trade and Development, while 118 countries, most of them in the Global South, are missing from the main international talks on how AI should be governed.
Stamirowska was clear that this isn’t her main mission—yet. “I haven’t spent my time on solving inequalities yet, because I have my fight of intelligence to fight at this very moment in time,” she said.
Still, the economics trouble her. “The promise of AI is to cure cancer, and it’s being sold to us this way. But I have a big issue with the fact that looking at how the economy is shaping up, there will be maybe 1% of people that will have access to those cures.”
With AI, she added, there’s a concentration of capital and compute. “Since skill is just being distilled, it may potentially be even harder to advance socially based on just skill.”
She has also seen technology work the other way; one of her co-founders previously built the first and largest community for competitive programmers, along with one of the first online tools for writing code.
“f you can code on a mobile phone, then you can learn programming. And if you’re in India, that actually allows you to advance socially,” she said as an example. “All of a sudden people were able to code on their mobile phones in multiple languages, because they couldn’t afford computers. We didn’t know. We were just making technology.”
Making AI smaller and more efficient changes the economics, she argued. Adoption would be faster and easier, and a model could be yours to train. But Stamirowska doesn’t expect raw computing power to stop mattering, either.
“The scale of compute will matter, but it will be more accessible.”
This is where her idea of sovereignty parts ways with Brussels. When the EU talks about sovereign AI, it means chips, data and models kept under European control and European law; when Stamirowska uses the word, she means something more personal.
“When I say sovereign, I mean a model of my own, not sovereign as the state. The customer, whoever that customer is, owns the weights and owns the model, and hopefully we don’t even see the weights.”
The gigafactories will decide where Europe’s computing power sits, sure. If Stamirowska is right, the most pressing question is who owns what runs on it.
Featured image: Courtesy of One Way Summit / Daniel Vegera
Disclosure: This article mentions a client of an Espacio portfolio company.
