Deep Tech Brief #01

Welcome to the very first issue of Deep Tech Brief.

Every Saturday, we break down deep tech companies building genuinely hard things, in plain language. No jargon, no hype.

This week's three happen to line up into one story.

Modern AI is only as good as its grip on the real, physical world. One of these companies manufactures that reality, one reads it, and one keeps it honest.

Let's get into it.

Here is a problem the AI industry rarely advertises. The large models now used to predict how a protein folds or how a cell responds to a drug are running low on fuel, and the fuel is data.

Text-based AI was trained on much of the public internet. Biology has no equivalent. Its data sits in databases that were built to store the results of individual academic experiments, not to train machine learning models, so they are patchy and skewed. By one analysis, 68 percent of all the data in the largest public sequence archive comes from just five species. Foundation models for biology are now showing signs of plateauing, not because the models are weak, but because they have run out of good, diverse, real-world data to learn from.

SURFACtoBioTech, led by co-founder and CEO Jacqueline De Lora, is attacking that shortage at the source.

Its tool is droplet-based microfluidics. In plain terms, the company takes a tiny volume of biological sample and breaks it into a fine emulsion, so each droplet becomes its own sealed micro-reactor running an experiment in parallel. De Lora's headline number: 10 microliters of sample can become 100,000 droplets, and every single droplet is the source of a data point.

That matters because the data AI needs most is the hardest to get. Models are good at spotting patterns in what they have already seen, but weak at predicting what happens when you intervene: change a gene, add a compound, and read the result. Those "perturbation" experiments are exactly what droplet systems can run by the hundred thousand. SURFACto is effectively building a factory for the kind of cause-and-effect biological data the field is starved of.

The obstacle is the one most hard-science startups hit: capital that moves as fast as the science. De Lora says quick, flexible funding has been her biggest challenge as the company moves from pre-seed into a seed round and pushes its early product toward manufacturing scale.

Worth watching, because every flashy "AI for biology" model quietly depends on someone solving this unglamorous data problem first.

Modern biology has an odd shape. Generating data has become cheap and fast, while making sense of it has not. The specialists who turn raw sequencing output into an answer, bioinformaticians, are scarce, expensive, and usually booked out for weeks.

This is not a new complaint. The shortage of bioinformaticians has been flagged since the 1990s, and it has only widened as sequencing costs collapsed. Reviews have repeatedly called for large increases in bioinformatics staffing across both academia and pharma, and the field adds tens of thousands of roles a year that go unfilled. Meanwhile most bench scientists were never trained to code in R or Python, so the data they produce often sits waiting on someone else's calendar.

Galagos, a Hamburg company founded in 2025 and led by CEO Alexander Zahn, is building what it calls a virtual bioinformatician to close that gap. You upload your omics data, ask a question in plain English, and its AI agent runs the analysis, from RNA-seq to single-cell to multi-omics, then hands back documented, reproducible results. Work that once took weeks, delivered in hours, with no code required.

Galagos is not alone in the idea. A wave of "bioinformatics copilots" has appeared now that language models are good enough to write and run analysis pipelines. Its real edge is not the AI itself, but where the data lives. Most of the serious competitors are US-based, while Galagos is EU-hosted and built for data sovereignty. For European pharma and biotech, bound by strict rules on where sensitive data can be processed, that is less a preference than a compliance requirement, and it is genuinely hard for a US rival to copy.

Is the traction real? By Zahn's account: a free open beta pulling in academics and startups, paid pilots with smaller pharma and biotech firms, and signed NDAs with four of the top ten pharma companies, with a first live pharma project expected within a month or two.

Still early, and success will hinge on getting cautious scientists to trust an agent with work they once would have checked line by line. But the sovereignty moat is real, and the timing is good.

"AI is a black box. That's a problem."

That is the starting point for Prof. Thomas Kopinski, co-founder of GAIA Lab. In high-stakes fields like energy, an AI that just says "trust me" is not good enough. If a system makes a decision that affects a power plant, you need to know why.

So GAIA builds what Kopinski calls "White Box" AI: transparent, auditable systems built on a neurosymbolic approach that the company says needs far less data and energy than mainstream deep learning.

Its most striking project is a world model that predicts and controls the plasma inside a fusion reactor. To see why that is hard, look at what it takes. When DeepMind and Switzerland's EPFL first trained an AI to hold a fusion plasma in 2022, the controller had to read 90 sensors and adjust 19 magnets roughly 10,000 times a second, because the plasma is violently unstable and a fraction of a second too slow lets it touch the wall and die. Doing that reliably, at that speed, is the whole game.

That is where GAIA's second bet comes in. The company runs its model on photonic hardware from Q.ANT, the Stuttgart firm that computes with light instead of electrons, aiming for responses in microseconds rather than milliseconds and at a fraction of the energy. The proof of concept already works on standard GPUs; getting it onto the custom photonic chips is this year's job.

On the AGI panic filling headlines, Kopinski is refreshingly grounded. He reads most recent progress as clever engineering rather than a leap toward superintelligence, and echoes Yann LeCun's call for calm, informed debate. The risks he actually worries about are already here: AI quietly deciding who gets what, and steering autonomous weapons.

One to watch, carefully. Not for a fusion breakthrough next quarter, but because pairing explainable AI with light-based computing is a genuinely different bet on how safe AI for the physical world gets built.

Three companies, one thread: AI is only as useful as its hold on the physical world. One side makes the data, one reads it, and one keeps it honest, and all three are being built in Europe.

Know a deep tech company we should break down next? Hit reply and tell us. The best editions come from real tips.

See you next Saturday.