The Dry Lab Revolution
New materials have always taken 10 to 20 years to reach the market. AI, simulation and robotic labs are removing most of the slowest part. This note explains what changed, how much time it really saves, and why that matters for early-stage returns.
Why materials have been slow, expensive and hard to fund
Almost every energy and climate technology is limited by a material: the electrolyte in a battery, the catalyst in an electrolyser, the sorbent in a carbon-capture unit. Many also rely on raw materials Europe imports from a handful of countries, such as iridium for electrolysers and graphite for batteries, so a better material is often a cheaper and more secure one too. Finding a better one has meant trial and error at the lab bench. The US Materials Genome Initiative put the typical time from discovery to commercial product at 10 to 20 years or more.1
For venture investors that timeline is the problem. A VC fund lives about 10–12 years. When a single material journey takes fifteen, the exit often falls outside the fund, and the capital burned in the early lab years sets the valuation bar for everything after.
Discovery moved from the wet lab to the computer
A “dry lab” does the searching in software. AI models trained on physics simulations predict which of millions of candidate materials are stable and useful. Self-driving labs then make and test the shortlist with robots that choose their own next experiment. Six data points show the scale of the shift, including two from European players focused on climate.
New crystal structures predicted by one AI model
Google DeepMind's GNoME predicted 2.2 million new crystals, of which about 380,000 are expected to be stable. DeepMind compares the output to roughly 800 years of prior discovery.
From 32 million candidates to a working battery material
Microsoft and US national lab PNNL screened about 32 million inorganic materials down to 18 candidates in 80 hours. PNNL then built a working solid-state battery with the new electrolyte.
Median speed-up from self-driving labs
A 2025 review of published self-driving lab studies found a median sixfold reduction in the experiments needed to reach a target, with larger gains in more complex problems.
Autonomous synthesis, and a lesson in rigour
Berkeley's A-Lab ran 17 days without human hands and made 36 of 57 target compounds. Critics showed many were not truly new, and Nature corrected the paper in 2026. Throughput is real; expert judgement still decides what counts.
Simulations to find better carbon-capture materials
Meta, Georgia Tech and Europe's CuspAI built Open DAC 2025, the largest open dataset for direct air capture: nearly 60 million quantum-chemistry calculations across about 15,000 porous materials. It trains AI to spot CO₂-capturing sorbents before anything is made. CuspAI, co-founded by Amsterdam's Max Welling, raised $100 million in 2025.
Target time to a new power-to-X catalyst
Berlin-based Dunia runs a self-driving lab for power-to-X electrocatalysts. It aims to cut the time to identify and manufacture a new catalyst to about three years. In a proof of concept, its platform beat human researchers on both speed and performance. These are company claims. Dunia is backed by the EIC and European VCs.
How much time does each lever really remove?
Each technology speeds up specific stages, not the whole journey. Switch the levers on and off to see where the years go. AI and self-driving labs shrink discovery and lab work from eight years to just over one in the base case. Pilot and scale-up gets somewhat faster, but qualification and commercial ramp do not move at all. Those are the physical steps, and they still need wet lab and pilot assets and people who have done it before, so the whole journey compresses far less than the headline numbers suggest.
The fourth lever is our own view of what experienced operators and a committed industrial customer can take off those steps. Switch it on to see where we make the difference.
| Stage | Traditional | Dry lab | Speed-up |
|---|---|---|---|
| Discovery & screening | 4.0 yrs | 0.8 yrs | 5.2× |
| Lab synthesis & optimisation | 4.0 yrs | 0.5 yrs | 7.8× |
| Pilot & scale-up | 4.0 yrs | 2.4 yrs | 1.7× |
| Qualification | 2.0 yrs | 2.0 yrs | – |
| Commercial ramp | 1.0 yrs | 1.0 yrs | – |
| Total | 15.0 yrs | 6.7 yrs | 2.2× |
Illustrative model by AIXTerra, not a forecast. Stage lengths reflect a typical energy-materials journey within the 10–20 year range cited by the Materials Genome Initiative. Self-driving lab factors are anchored on the median 6× acceleration in the literature [4]; AI factors on published screening results [2][3]. Quantum factors are speculative and assume fault-tolerant hardware, which most estimates place in the 2030s [6]. Where two levers act on the same stage, their effects are multiplied. The operators & industrial pull factors are AIXTerra's own view, not market data. Commercial ramp excludes building a full-scale first-of-a-kind plant.
The levers do not multiply across the whole journey. They collapse the front end. The winners are the companies that also shorten what comes after.
Time is the biggest driver of early-stage returns
Take the same outcome, an exit at 5× the money invested, and change only how long it takes. Shaving five years off the journey nearly doubles the annual return, and moves the exit comfortably inside a ten-year fund.
Annual return (IRR). Exit falls after a typical fund's life.
Annual return (IRR). Exit within the fund, with room for follow-ons.
Simple arithmetic illustration of time-to-exit on IRR, before fees. Not a projection of any fund or investment.
Two further effects compound this. Fewer failed experiments mean less capital is spent before the first proof point, so founders give away less of the company early. And a company that can show results in two years instead of five raises its next round on evidence rather than on promise.
The market has noticed. AI-for-materials startups raised more than $1.3 billion in two years, according to PitchBook, including a $300 million seed round for Periodic Labs in 2025.7 Most of that capital sits in large US platforms, which leaves the European, application-specific layer comparatively under-served.
Discovery is becoming a commodity. The moat moves downstream.
When a big tech lab publishes 380,000 stable materials for free, predicting a material is no longer a business on its own. Value accrues to whoever owns the scarce steps that follow.
Generic prediction models and open databases
Improving fast, widely available, and funded by the largest tech companies. Hard to defend as a standalone product.
Proprietary experimental data from closed-loop labs
Every experiment a self-driving lab runs creates data no one else has. The model improves with each loop.
Scale-up know-how and industrial pull
Turning a powder into tonnes at spec and cost, with a customer ready to qualify it. This is where most of the remaining years sit, and where experienced operators make the difference.
The risks a careful investor should price in
Predicted is not manufacturable
Many AI-predicted materials cannot be made cheaply or at scale. The A-Lab correction shows how easily results are overstated.
The slow stages stay slow
Pilot plants, certification and customer trials follow physical and commercial clocks that software does not change.
Valuation froth
Large rounds for AI-science platforms can inflate prices across the sector before revenue is proven.
Quantum timing
Useful quantum chemistry is widely expected in the 2030s, not today. We treat it as upside, never as part of the base case.
The Dry Lab Revolution, applied to energy and climate
AIXTerra backs top-tier European teams with a sharp niche focus and defensible IP in energy and climate materials. They pull every lever of the dry lab revolution (AI-driven discovery, simulation and automated experimentation) and pair it with first-class wet lab assets and the hands-on skill to make, test and scale what the models propose. That combination lets them own the steps after discovery: proprietary data, scale-up and industrial customers. We invest early, with initial tickets of €0.5–3 million from pre-seed to Series A.
General-purpose discovery models are becoming shared infrastructure. Their real impact is downstream: dozens of focused ventures, each pairing AI with deep domain IP and wet lab skill to own one hard material problem. These are illustrative niches where we expect that tail to grow.
Niches where quantum and quantum-AI hybrid simulation could add the most, for example strongly correlated catalysts and superconductors. We treat this as a 2030s option, not a base case.
Why AIXTerra is positioned to win in the Dry Lab Revolution
Spotting the shift is not enough. Winning means seeing the best teams first, judging both the models and the lab work, and helping founders through the slow stages that software does not shorten. Four strengths set us apart, and two facts show they are already working.
Operator-Led
Our VC team includes materials scientists, energy operators and AI engineers who have built and scaled ventures in each domain. We can test a model's claims and a pilot plant's numbers with the same rigour.
Full-cycle track record
Partners have backed deep tech materials companies such as Thorizon, Thermagy, E-magy, Chaincraft and MX3D since before the current AI wave, from first ticket through to exit.
Industrial pull
Partnerships with leading industrial corporates give portfolio companies a route to pilots, qualification and first customers.
Focus on the hard part
We help founders compress scale-up and qualification, where most of the remaining time and value sit, and where wet lab excellence still decides who wins. This is the fourth lever in the model above.
A pipeline ready to deploy
We have already mapped and engaged a full pipeline of European ventures applying the Dry Lab levers to energy and climate materials. From day one we can deploy into teams we know, rather than starting a search.
Diligence that opens doors
Our approach to due diligence is valued by the companies themselves. We write a vendor-style report that founders can share with co-investors, and that report doubles as our own investment homework.
- Materials Genome Initiative, About MGI; 2014 Strategic Plan, US National Science and Technology Council.
- Merchant et al., Scaling deep learning for materials discovery, Nature 2023; Google DeepMind summary.
- Microsoft Azure Quantum blog, Jan 2024; Microsoft Source.
- Adesiji et al., Benchmarking self-driving labs, Digital Discovery 2025.
- Szymanski et al., An autonomous laboratory for the accelerated synthesis of inorganic materials, Nature 2023 (corrected Jan 2026); C&EN on the correction.
- PostQuantum, Quantum chemistry timeline review, May 2026.
- PitchBook, Apr 2026; Fortune, Sep 2025.
- Sriram et al. (incl. Welling, CuspAI co-founder), The Open DAC 2025 Dataset for Sorbent Discovery in Direct Air Capture, arXiv Aug 2025; Carbon Herald.
- SiliconANGLE, Dunia raises $11.5M, Oct 2024.
Important information. This note is issued by AIXTerra Ventures for information and discussion with qualified investors only. It is thought leadership, not an offer or solicitation to buy or sell any security or fund interest, and not investment, legal or tax advice. Illustrative figures and models are simplified assumptions, not forecasts or indications of future returns. Early-stage investing carries a high risk of losing all capital invested and is illiquid. Past performance of the team is no guarantee of future results. Any investment is made solely on the basis of the relevant offering documents.