"AI for Life": a framework for investing in intelligence under real-world constraints
From narratives to infrastructure: why AI must be invested within the constraints of the living world.
Vous souhaitez lire en français ? → Cliquez ici
AI Wars — Episode 2
In the first episode of AI Wars, we mapped the narratives shaping AI’s future.
This second essay moves from narratives to allocation, examining how assumptions about intelligence translate into infrastructure and capital deployment (and Asterion’s own investment approach.)
This article draws on conversations across the Asterion ecosystem and was co-written by Antonin Léonard (partner at Asterion) and Arthur De Grave.
In less than a decade, AI has become one of the fastest-growing sources of pressure on the physical systems that sustain life.
Data centers now compete with cities and industries for electricity, water, land, and critical minerals. In some regions, AI workloads already absorb a significant share of national power supply. Indirect emissions are rising faster than official decarbonisation claims can conceal. Local pollution around compute infrastructure produces health effects comparable to heavy industry.
And yet, AI continues to be discussed as if it were weightless - a realm of code, models, and abstraction, detached from the conditions of the living world.
This mismatch is not accidental. It reflects a deeper assumption: that intelligence can scale independently from life.
This article starts from a different premise.
Intelligence did not emerge outside the living world. It emerged from it - and it remains constrained by it.
Treating AI as an autonomous force, rather than as an extension of life-bound systems, leads to a structural error in how we build, deploy, and finance “intelligence”.
AI for Life is a framework for investing under this constraint. Not as a moral posture, but as a condition of viability.
Intelligence vs. Life
A Bergsonian counterpoint
French philosopher Henri Bergson offers a more demanding point of entry. In Creative Evolution, intelligence is not presented as a natural ally of life. On the contrary, Bergson describes intelligence as being shaped by action on inert matter, not by an intuitive understanding of living processes.
“Intelligence is characterised by a natural incomprehension of life.”
Intelligence excels at isolating, measuring, and manipulating. It breaks continuity into discrete elements and turns movement into structure. These operations are powerful, but they come at a cost: when applied to living systems, they tend to flatten what is fluid, reduce what is evolving, and mistake optimisation for understanding.
From this perspective, there is no reason to believe that AI would depart from this logic. Quite the contrary: AI radicalises it. By extending abstraction, formalisation, and optimisation to unprecedented levels, AI turns a structural limitation of intelligence into a systemic force, thus intensifying the gap between intelligence and life.
This is why intelligence cannot be left to operate without constraints. Once detached from the conditions of the living, its capacity to act expands faster than its ability to perceive what it erodes.
This tension cannot be wished away. If intelligence tends to misread life, then any attempt to align AI with the living world cannot rely on spontaneity or good intentions alone. It requires limits and constraints (and explicit choices! We will get to that).
AI for Life
A general principle of continuity
This is where the idea of AI for Life takes shape.
AI can extend the capacities of the living, refine its modes of action, and improve the conditions under which life develops. This only holds as long as intelligence remains situated within the world it acts upon, rather than positioned above it.
From this starting point, several limits become unavoidable.
When this perspective is taken seriously, it leads to a small number of constraints that gradually structure both technological choices and investment decisions:
Pillar I — Ontological boundaries
What we refuse to ask AI to be
Given the current state of knowledge, super-intelligence – whether you call it AGI or ASI – appears too speculative to serve as a credible investment horizon. It rests on assumptions that are quite wild and encourages abstraction from the worlds AI already reshapes.
The same drift appears in techno-messianic claims presenting AI as a solution to climate change, healthcare, or education. AI can certainly accelerate research, improve coordination, and reduce certain frictions, but it cannot replace political decisions, industrial transformations, or collective responsibility all by itself.
A further boundary concerns creation: AI systems generate, recombine, and explore vast spaces of possibilities. Creation remains tied to intention and meaning. These dimensions belong to the living (Bergson, again!). Displacing them would not simply change production processes; it would shift the locus of meaning itself.
Pillar II — Material constraints
Frugality as a condition of viability
AI operates within strict physical limits. Energy, hardware, networks, and land define its footprint. And it is well known that gains in efficiency frequently trigger rebound effects that increase overall consumption rather than reduce it.
In such a context, frugality becomes a condition of durability. Model size, training regimes, inference efficiency, hardware lifecycles, and infrastructure choices all matter, and not every use case deserves to scale. We must collectively agree on this inconvenient truth: performance metrics detached from physical costs offer a distorted view of progress.
Proposals to relocate compute infrastructure off-planet – for instance through orbital data centers projects – illustrate this tension clearly: by externalising material limits rather than integrating them, they treat planetary boundaries as obstacles to be bypassed and deepen the abstraction from the living world that this thesis rejects.
Pillar III — Situated impact
Where AI actually makes a difference
Not all AI applications deserve the same level of attention or capital.
The most meaningful contributions emerge in domains where complexity is real and gains compound over time: industry, energy systems, mobility, health, and public infrastructure. In these contexts, AI acts as decision support and coordination infrastructure, while human judgment remains structurally central.
By contrast, spectacle-driven applications and content saturation consume resources without strengthening the systems they touch. They draw attention while leaving underlying capabilities largely unchanged.
So how do we move from these principles to investment choices?
Taken together, these constraints form an investment doctrine.
Over the past years, we have reviewed several hundred early-stage companies claiming to build “AI-powered” solutions. Roughly two hundred placed AI at the core of their value proposition. Only a small fraction combined technical ambition with a clear understanding of material constraints, sector-specific realities, and long-term viability. This filtering process gradually shaped our investment practice.
At Asterion, AI is approached as an infrastructure of the living world. Investments focus on AI where it strengthens life-supporting systems, improves conditions of production and coordination, and remains compatible with the material realities it depends on. To date, this has resulted in a limited number of deliberately concentrated investments, across domains where AI acts as a lever rather than a substitute.
Living Models (formerly .omics) applies AI to experimental biology, where the complexity of living systems resists brute-force abstraction. Rather than replacing biological reasoning, its technology supports researchers in navigating experimental uncertainty, accelerating discovery while remaining grounded in the material realities of life sciences.
Argile.ai operates at the intersection of AI and the built environment. Its systems assist in building renovation by structuring technical choices, regulatory constraints, and energy performance scenarios. Here, AI functions as coordination infrastructure, improving the conditions under which physical transformation takes place.
Futurail focuses on rail infrastructure, a critical and highly constrained system. By optimising train paths and capacity allocation, its technology increases the efficiency of existing networks, supporting modal shift without demanding new extractive infrastructure. Intelligence is deployed in service of a system whose limits are explicit and non-negotiable.
Together, these investments reflect a consistent logic: AI is deployed where complexity is real, constraints are binding, and gains translate into durable improvements of life-supporting systems. Over time, this approach has imposed its own discipline, shaping not only what is supported, but also what is deliberately left aside.
Choosing continuity over hubris
AI stands at a point of bifurcation. It can deepen existing crises or help navigate them. This is why we borrowed the concept of pharmakon from the French philosopher Bernard Stiegler. The difference lies less in the sophistication of models than in the constraints imposed on their development and use.
AI for Life frames innovation as something that must remain anchored in the conditions of the living world.
Intelligence did not emerge against the living world. It emerged from it. It has also learned, time and again, to abstract itself from the conditions that sustain life. Holding these two movements together may be the most demanding form of progress we can choose.
This is why investing in AI today means taking responsibility for how capital constrains intelligence within the limits of the living world.
Three questions for… Cyril Véran, CEO & Co-founder of Living Models
Cyril Véran is the co-founder and CEO of Living Models (currently operating as .omics), a biotech company developing AI systems to model plant biology for breeding and agronomy. With BOTANIC, its foundation model, Living Models aims to reduce biological experimentation by structuring uncertainty upstream, before costly lab and field trials. Because in biology, AI only matters when it operates within physical constraints and delivers real-world impact.
AI’s environmental impact is increasingly debated. How do you approach this at Living Models?
The problem is that we talk about “AI” as if it were a single thing. Training massive, generic models to generate content is very different from training specialised models designed to replace years of biological experimentation.
In biology, mistakes are costly. A wrong prediction doesn’t just waste compute — it wastes months of lab work, field trials, land, water, and chemical inputs. That forces discipline. At Living Models, our models are targeted, constrained, and built to reduce real-world experimentation, not multiply it.
The question isn’t whether AI consumes resources. It’s whether it replaces something far more expensive. In our case, it does.
Living Models is building BOTANIC, a foundation model for plant biology. What sets this approach apart from generic AI applied to biology?
Biology doesn’t behave like language. Genomes aren’t sentences, and plants aren’t static systems. The same genetic sequence can behave very differently depending on environment, stress, and interactions.
Generic AI approaches tend to flatten that complexity. BOTANIC is built around it. It integrates multiple layers of biological information — genomics, transcriptomics, metabolomics, phenotypes, and environmental data — to model how traits actually emerge in plants.
Even then, predictions remain hypotheses. BOTANIC doesn’t replace the lab or the field. It structures decisions upstream, reducing uncertainty before experiments begin.
What’s the most credible path from more powerful models to measurable impact in the physical world?
The path is slower and more fragmented than most narratives suggest. More powerful models only matter if they are embedded in existing workflows and institutions. In plant breeding, that means working with seed companies, researchers, and regulators, each with their own timelines and constraints.
The real bottlenecks are rarely computational. They are access to high-quality data, the cost and duration of experimental cycles, and the ability to translate predictions into decisions that practitioners trust. Trust is earned when models fail visibly and improve transparently, not when they promise too much.
Measurable change comes when AI shortens feedback loops. If breeders can test fewer hypotheses, with higher confidence, under real environmental conditions, then traits that reduce water use, chemical inputs, or climate vulnerability become economically viable faster.
That is where AI stops being an abstract capability and becomes an infrastructure: something that operates within biological limits and helps navigate complexity rather than pretending to resolve it.
A quick look at three companies in our portfolio currently raising — all building real-world impact across critical infrastructure, mobility, and biology.
→ Futurail — building the next generation of autonomous, high-frequency rail operations to boost Europe’s train capacity.
→ Bubble Robotics — designing autonomous underwater robots to inspect, monitor, and protect critical ocean infrastructure.
If you’re interested in any of these rounds, contact me: antonin@asterionventures.com
Humanoid robots or human connection? (The Conversation)
Using Tesla’s Optimus as a case study, The Conversation explores what humanoid robots reveal about our deeper AI ambitions. Beyond productivity gains, Musk’s vision taps into emotional and cultural expectations: companionship, empathy, and social presence. The article argues that the humanoid form is not just an engineering choice, but a symbolic one, blurring functionality with intimacy. The core risk is not technical failure, but social displacement — outsourcing care, conversation, and tolerance to machines. The real design challenge, it concludes, is whether AI systems will quietly replace human connection or actively steer us back toward one another.
AI’s memorization crisis (The Atlantic)
The Atlantic reports that several leading language models — including GPT, Claude, Gemini, and Grok — can reproduce long passages from copyrighted books they were trained on, contradicting years of public denials from AI companies. Researchers showed that, when prompted strategically, models could output near-complete texts from major literary works. The issue, known as “memorization,” challenges the claim that models merely learn patterns rather than store content, and raises serious legal, economic, and reputational risks for the industry. If models copy rather than abstract, the foundations of generative AI — from copyright defenses to training practices — may need to be rethought.