Steven Latré

Steven Latré's thinking centers on a single mathematical alarm bell: the compute demands of AI have decoupled from what chips can deliver. Until 2010, he explains, AI compute needs roughly tracked Moore's law, doubling every two years; since then, models have grown by a factor of about 100 per year, while chip improvements have stayed on their old two-year cadence ¹². The result is a structurally widening gap between what AI wants and what hardware can efficiently provide, and he distills the consequence bluntly: "100 keer meer rekenkracht betekent 100 keer meer energie, 100 keer meer water" ¹². This is not a marginal efficiency problem for him but the central engineering and policy challenge of the field. He compounds the point by noting that agentic AI, where models chain steps and talk to each other (his example is Manus computing for over an hour), multiplies inference compute by another factor of roughly 100 compared to a single question-answer exchange ¹². Layered on top of that is "sleep-time compute," where models now pre-compute or effectively "dream" while idle, meaning systems burn energy around the clock even without anyone using them ¹².

Despite this alarm, Latré is not a pessimist about what can be done technically. He repeatedly points to DeepSeek as proof that the trajectory is not fixed: a team of a few hundred people, by carefully bridging hardware and software optimization, produced a model roughly 36 times more energy-efficient than a standard GPT-style system ¹². For him this is the strongest evidence that Europe, despite lacking the chip fabrication and capital of the US and China, can still leapfrog rather than simply follow. His own work at OpenChip reflects this conviction in hardware terms: chips designed so that consumption can flex with the actual supply of renewable energy, deliberately trading peak performance for efficiency and resilience ¹. He also points further out on the horizon to superconducting digital chips that could run about 100 times faster at the same energy cost, though they only function below -35°C, effectively meaning a datacenter's compute could be shrunk into "a shoebox, wrapped in a freezer" ¹².

On infrastructure and policy, Latré argues that datacenters have quietly become critical national infrastructure comparable to railways or water mains, and that their siting should not be left to pure market economics ¹². His go-to illustration is Amsterdam, which hosts a disproportionate number of datacenters simply because transatlantic cables land nearby, not because of any advantage in energy supply, a siting logic he considers exactly backwards for an AI era defined by energy scarcity ¹². He extends this into a vision of "AI at multiple speeds," where fast, high-compute answers cost more, both financially and environmentally, and where energy impact becomes something priced into everyday model usage rather than externalized ¹². Underneath this sits a broader unease: he states plainly that "de economische waarde die AI-modellen vandaag leveren, komt niet overeen met de infrastructuurinvestering erachter," and that the current trajectory of ever-larger compute spending is not sustainable ¹. It is telling to him that every major LLM provider, Google, OpenAI, and Microsoft among them, is now signing deals with nuclear power providers just to secure the energy these systems require ¹.

Latré is equally direct in dismissing near-term AGI as a serious prospect. "Eigenlijk moeten we eerlijk zijn, we hebben geen enkel pad richting AGI," he says, and when pressed on 2027 timelines the answer is "absolutely no" ¹². His reasoning rests on Moravec's paradox: language models pattern-match brilliantly on exams and benchmarks but fail at simple physical or common-sense reasoning, which he takes as evidence there is no known technical route from today's LLMs to general intelligence ¹². But he treats the AGI debate itself as something of a distraction, a "false discussion" in his framing, arguing that AI is already disruptive enough through its energy footprint and its effect on jobs without ever needing to become AGI, and that Terminator-style what-if scenarios pull attention away from problems that are already real and present ¹². On jobs specifically, he notes the disruption has landed in an unexpected place: five years ago the expectation was that drivers would be automated first, yet it is the creative sector, his example being Albert Heijn using AI-generated voice-overs for its ads, that is being hit early ¹.

Running through all of this is a moral register that Latré returns to more than once. He describes AI as having outgrown its status as a mere technological revolution to become "een maatschappelijke revolutie," and argues that the people building it, himself included, carry a societal duty to bring it forward "op de juiste ethische manier" ¹². He frames his own move into European AI development in explicitly personal, almost generational terms, describing "een soort van bijna morele verantwoordelijkheid" he felt, even toward his children later, to be able to say he at least tried to advance European AI ¹². That same instinct pushes him toward a pragmatic call to action on Europe's chip and AI ambitions: rather than dwelling on the continent's disadvantages, he insists, "laten het ons gewoon echt proberen" ¹².

Latré is also careful to define the role he wants AI to play, resisting framings that put humans and AI in competition or in some shared category of agency. "Voor mij is niet de rol van wij als mens in AI," he says. "Ik zou het willen zeggen: de rol van AI is die van een assistent in de wereld van mensen" ¹². In the same spirit, he undercuts any impulse to anthropomorphize the technology, joking that users need not feel obliged to thank a chatbot, "want voorlopig hebben ze nog geen gevoelens hè" ². This combination, technical alarm about energy, confidence that efficient European alternatives are possible, skepticism about AGI hype, and insistence on a clearly bounded, assistant-like role for AI, forms the throughline of his public thinking.

Finally, Latré situates the current AI moment within a longer historical pattern he is wary of repeating. Thirty years after the internet's invention, he observes, society is still issuing "course corrections" for the damage caused by unchecked techno-optimism, citing efforts like Solid personal data vaults as examples of correcting mistakes that were foreseeable at the time ¹². His clear takeaway is that AI development should not wait three decades to learn the same lesson: the efficiency gains, infrastructure choices, and governance questions around AI need to be confronted now, deliberately, rather than corrected after the fact. He even raises efficiency gains as a matter of social choice rather than pure output,