The Race to Build the Substrate of Intelligence
Every month I feature three founders building at the edge of deep tech and healthcare AI — and ask the one question nobody else is asking. This is the inaugural edition.
This column is part of an April series. The intellectual foundation is laid in We’ve Been Asking the Wrong Question About AI which argues that intelligence has always been a property of systems, not individual nodes. From Newton to Schrödinger, from Tomasello's cultural ratchet to Sculley's governance debt, the pattern is the same: the collective solves what the individual cannot. The three founders in this edition are building the next layer of that stack.
Enterprise AI's decisive contest is not bigger LLMs or slicker agents. It is the urgent push to engineer the physical substrate (neuromorphic chips, neocloud racks, catalyst lattices) that underpins every abstraction. Botch the foundation, and nothing stands. Recent analysis exposes AI's binding constraints: neuromorphic's energy-sovereignty leap (Loihi 3 transforming neoclouds), quantum's regulatory rush, and materials chokeholds felling Physical AI outside simulations.
Substrate rules—not algorithms, not architectures.
Lab notebook sketch, log axis 10⁻¹⁰ m (atomic) to 10⁻⁸ m (genomic). Markers: Loihi 3 (neuron lattice), Copernic Catalysts (atomic clusters), Lila Sciences (molecular vial), Sophia Genetics (DNA coil). Caption: Substrate reality: where AI prediction hits physical limits, scale by scale.
Each attacks matter's independence from computation at unique resolutions:
- 10⁻¹⁰ m (Atomic): Loihi 3 crams 8M neurons/64B synapses on 4nm, slashing GPU power for sparse inference—neoclouds' efficiency cornerstone.
- 10⁻¹⁰ m (Catalysts): Copernic reprograms surfaces atom-by-atom against quantum chaos.
- 10⁻⁹ m (Molecular): Lila cycles AI from synthesis to validation, forging reliable outputs.
- 10⁻⁸ m (Genomic): Sophia masters bio-variation at population scale, defying dataset ideals.
One constraint. Multiple scales. Clear stakes: The founder who conquers substrate across levels owns AI's next era.
This debut Frontier Founders column spotlights those builders making it real.
Lila Sciences
The Autonomous Lab That Is Teaching AI to Make Things
For as long as AI has been applied to materials discovery, the workflow has been roughly the same. AI suggests candidates, humans make them, nature reveals which predictions were wrong, humans feed the results back. The loop works. It is also slow, expensive, and bottlenecked at the synthesis step, the part AI cannot do.
Lila Sciences is betting that the synthesis step is now automatable. Not fully autonomous yet, but agentic enough to dramatically compress the loop. Their AI agents plan experiments, direct robotic laboratory equipment to execute them, and interpret the results, including the unexpected ones. A human scientist keeps close watch, approves next steps, and intervenes when the system encounters something outside its training distribution. The generative load is shared in a way that has never been possible before.
The problem they are solving: Materials discovery has always been constrained by the cost and time of physical synthesis. Computational prediction can be extraordinarily fast. Making the predicted material and testing it in the real world is where decades get lost. Lila Sciences is attacking the synthesis bottleneck directly, not by making better predictions, but by making the feedback loop between prediction and physical reality faster and cheaper.
This matters enormously beyond materials science. The same loop (simulate, predict, make, test, revise) governs drug discovery, catalyst development, battery materials, and agricultural chemistry. Every industry that runs on physical matter is bottlenecked by the cost of moving from digital prediction to physical reality. Lila Sciences is building the infrastructure to collapse that gap.
How they will win: Speed of the feedback loop. Every experiment that runs autonomously overnight is one that used to take a week of researcher time. Every failed candidate that the system learns from without human intervention is governance debt that does not accumulate. The labs that adopt autonomous experimentation first will compound their materials knowledge at a rate that manually-operated labs simply cannot match.
There is a substrate ceiling here worth watching. The AI agents running Lila's experiments are only as good as the physical instruments they control, the sensor data those instruments produce, and the materials they are working with. As I wrote this month in The Physics of Intelligence, the bottleneck in physical AI is almost never the model. It is the sensors, the actuators, the surface interactions, the materials science of the lab equipment itself. Lila Sciences will win by building AI that understands the physical constraints of its own laboratory environment, alongside building better AI.
There is a deeper frame here worth naming. Andrej Karpathy's concept of Software 2.0, where neural networks replace hand-written code, is now reaching the scientific method itself. In Software 1.0, a scientist writes the experimental protocol. In Software 2.0, the AI learns what experiments to run from the accumulated outcomes of previous ones. Lila Sciences is building the infrastructure for scientific Software 2.0. The human scientist in the loop is not a limitation. For now, that human is the entity that defines what a good experimental outcome looks like. The day Lila's agents can specify their own reward signal is the day autonomous science becomes genuinely recursive. The infrastructure being built now will determine who gets there first.
One governance note worth watching: I have seen autonomous experimentation fail not because the AI was wrong, but because the sensor degraded overnight and nobody in the governance layer was watching the substrate rather than the model. Lila Sciences will need to build governance infrastructure for its physical environment with the same rigor it applies to its models.
Copernic Catalysts
Redesigning Chemistry from the Atom Up
Ammonia production is responsible for over one percent of global carbon emissions, a number that seems modest until you realize it is higher than any other single industrial chemical process on earth. The Haber-Bosch process that makes almost all of it has been essentially unchanged since 1913. It runs at high temperature and pressure, consumes enormous quantities of natural gas, and produces nitrogen fertilizer that feeds roughly half the human population.
Copernic Catalysts believes AI can redesign the catalyst that makes the reaction happen, from the atomic level up, and in doing so dramatically lower the energy requirements of one of civilization's most critical industrial processes.
Their approach combines density functional theory, the quantum mechanical computational method that I spent years using in my own PhD research on computational materials science, with machine learning to understand and redesign catalytic behavior at the atomic level. DFT has been a research tool for decades. What Copernic has done is make it commercially useful by pairing it with machine learning that can navigate the vast space of possible catalyst compositions faster than any human-guided search.
The problem they are solving: Catalyst design has historically been a combination of deep theoretical knowledge and expensive trial and error. The theoretical knowledge, understanding why a particular arrangement of atoms at a catalyst surface accelerates a reaction, requires quantum mechanical simulation that is computationally intensive and expertise-intensive. Machine learning can learn the patterns from existing DFT data and dramatically reduce the number of full quantum simulations required to find promising candidates. Copernic is using this combination to attack a problem (green ammonia and e-fuel catalysts) where the stakes for getting it right are civilizational.
I want to be specific about why this resonates with me personally. At Caterpillar, as a chemical engineer, I worked on a problem that sits at the heart of what Copernic is building: how to clean and regenerate industrial catalysts so they can be reused rather than replaced.
The challenge was not finding a better catalyst. The catalyst was known. The problem was that over time, contaminants (sulfur compounds, carbon deposits, metal poisons) accumulate on the active sites and choke the reaction. The catalyst is not broken. It is dirty, in a very specific, chemistry-dependent way. Regeneration means understanding what is coating the surface, at what depth, and what conditions will restore the active sites without destroying the underlying structure. Get it wrong and you damage the catalyst permanently. Get it right and you extend operational life by months, sometimes years.
What we were trying to do by hand (characterize the fouling chemistry, model the regeneration conditions, optimize the process for reuse) is exactly what Copernic's AI platform can now do computationally, from the atomic level up, at a speed and precision that no human-guided process can match. The DFT plus machine learning combination they are using is not just useful for designing new catalysts. It is the right tool for understanding why existing catalysts degrade and what regeneration chemistry will actually work.
This matters enormously for industrial operators. A catalyst regeneration solution built on atomic-level understanding of degradation mechanisms is not a one-time design service. It is a continuous intelligence system for the most expensive consumable in industrial chemistry. If Copernic builds toward that vision, they are not competing in materials discovery. They are building the operating system for industrial catalysis.
This is the substrate ceiling applied to decarbonization. You cannot transition to clean energy without better catalysts for green hydrogen production, ammonia synthesis, and carbon capture. You cannot build better catalysts without understanding chemistry at the atomic level. You cannot do that at scale without AI that bridges quantum mechanics and machine learning in the way Copernic is building.
How they will win: First-mover advantage in a domain where the knowledge compounds. Every catalyst the system designs and validates becomes training data that makes the next design faster and more accurate. The team that builds the largest proprietary database of DFT calculations for industrially relevant catalysts, and the ML models trained on that data, will have a moat that is genuinely difficult to replicate. This is the materials equivalent of foundation model training data: the more you have, the better the model, and the harder it is for a late entrant to catch up.
There is also a governance angle here that I find fascinating. Copernic's AI is making atomic-level predictions about chemical reactions that will eventually run at industrial scale. The validation chain runs from DFT simulation to bench test to pilot plant to full industrial deployment. It is long and expensive. At each stage, the question is: how much do you trust the AI's prediction, and what evidence do you need before the next step? That is an AI governance question in the most literal sense. The answer will determine how fast green chemistry can actually scale
Sophia Genetics
Making Genomic Medicine Work Everywhere
The science of using AI to analyze genomic data for disease diagnosis, treatment selection, and patient monitoring is advancing faster than any other domain in clinical AI. The infrastructure to deliver that science to patients outside a handful of elite academic medical centers is not. A child with a rare genetic disorder in Nairobi or Manila or São Paulo has access to roughly the same genomic interpretation today as a child in Boston had ten years ago. The frontier moves. The floor barely shifts.
Sophia Genetics is betting that the floor is now movable. Their platform pools de-identified multimodal patient data (genomic, radiomic, clinical) from more than 800 hospitals across 70 countries, and uses that federated dataset to train AI models that work across the genuine biological diversity of human populations, not just the narrow slice represented in most published genomics research.
The problem they are solving: Most genomic AI models are trained on data from patients of European ancestry treated in well-resourced health systems. The models perform well on that population and degrade, sometimes dramatically, on everyone else. This is not a fairness problem in the abstract. It is a clinical accuracy problem. A variant classifier trained primarily on one ancestry group will miss pathogenic variants and flag benign ones in patients from underrepresented populations. The diagnostic odyssey for those patients stretches from months into years.
Sophia Genetics is attacking this directly. By federating data from hospitals across the geographies where the biological diversity actually lives, they are building training sets that reflect the full range of human variation. The AI models built on top of that substrate work for the population a hospital actually serves, not the population the research literature happens to over-represent.
This matters enormously beyond rare disease. The same distribution problem governs oncology biomarker testing, pharmacogenomics, and cardiovascular risk prediction. Every clinical AI application that depends on genomic input inherits the ancestry bias of its training data. Sophia Genetics is building the infrastructure to break that inheritance.
How they will win: Network effects on training data that no single institution can replicate. Every hospital that joins the platform contributes data that improves model performance for every other hospital. Every variant interpretation that gets validated in one geography becomes a signal that strengthens interpretation in another. The team that builds the largest federated, ancestry-diverse genomic dataset will have a moat that is genuinely difficult to replicate, because the moat is not the model. The moat is the trust relationships with 800 hospitals and the data governance infrastructure that makes federation actually work.
This is the substrate ceiling applied to biology. You cannot predict clinical outcomes without genomic data that reflects the patient in front of you. You cannot get that data at population scale from any single health system. You cannot federate across health systems without governance infrastructure that solves data sovereignty, patient consent, and regulatory variation in 70 different jurisdictions. Sophia Genetics has been building that infrastructure for more than a decade. The AI advantage is downstream of the governance work.
There is a frame here that connects all three companies in this column. Lila Sciences is collapsing the loop between AI prediction and physical synthesis. Copernic Catalysts is collapsing the loop between quantum simulation and industrial chemistry. Sophia Genetics is collapsing the loop between genomic prediction and clinical reality across the full diversity of human biology. Each is attacking a substrate constraint that no algorithm alone can solve. Each is winning by building the physical, chemical, or biological infrastructure that the AI layer depends on.
One governance note worth watching: Sophia Genetics operates at the intersection of clinical AI, federated learning, and cross-border health data. Three regulatory regimes that are each evolving rapidly, and that interact in ways no regulator has fully worked out. The companies that build genuine compliance infrastructure now, rather than retrofitting it later, will be the ones that scale when the regulatory dust settles. Sophia Genetics has been building that compliance posture since before federated genomics had a name. That head start compounds.
The Pattern Across All Three
Substrate is not a metaphor. It is the literal physical, chemical, and biological layer that AI predictions have to land on to be useful. Lila Sciences owns the molecular substrate. Copernic Catalysts owns the atomic substrate. Sophia Genetics owns the genomic substrate. Each is building infrastructure that the rest of the AI stack will eventually depend on, whether the rest of the AI stack realizes it yet or not.
The founders who understand this are not competing in the same market as the LLM and agent companies. They are building the floor those companies will eventually stand on.
The question nobody else is asking is the one I will keep asking in this column: who is building the substrate, and who is building on sand?
Next month: three founders working on the neuromorphic, quantum, and sensor layers.
About Frontier Founders
Every month I feature three founders building at the edge of deep tech and healthcare AI — and ask the one question nobody else is asking. I am particularly interested in companies navigating the gap between laboratory demonstration and production deployment, and in founders building for the Global Majority rather than optimizing for the narrow slice of the world that dominates most tech investment.
If you know a founder who should be featured, or if you are building something that is hitting the substrate ceiling in your own domain, I would like to hear from you.