Europe's AI Window Is Closing Fast
From autonomous software engineering to frontier AI sovereignty: five startups to watch, five EU takeaways, and why the next two years matter.
Swedish AI coding company Lovable just raised another $400M at a $13.3B valuation. Days earlier, AI code-review startup CodeRabbit raised $143M at $1.5B, while processing more than 2M code reviews every week. AI software engineering is no longer a side market — it is becoming one of the clearest commercial battlegrounds in AI.
And the opportunity is already moving beyond code generation.
This week, we look at five early-stage European startups attacking what comes next: autonomous maintenance, security architecture, legacy modernisation, QA and safety-critical software.
At the same time, the European AI Office is asking the bigger question: can Europe build enough compute, capital and strategic leverage to remain competitive at the frontier? Its latest expert report argues that the next one to two years could be decisive.
Also in this issue: in The Leap, DTM's founder Martin Schilling argues that winners do not enter crowded categories — they define the narrow category they can own, and the metric that proves they are winning. McKinsey is benchmarking what makes corporate–Deep Tech partnerships scale, and founders and corporates can contribute via an 8-minute survey.
Dear all,
This is part of the book NORTH STAR: The 12 Formulas Behind the World's Most Valuable Deep Tech and AI Companies. How Clarity Compounds into Capital, Customers, and Enterprise Value. Sign up HERE for the waiting list to be amongst the first to receive the book.
By the morning of 17 June 1928, the crew had been stranded almost two weeks in a Newfoundland harbour, pinned down by fog, gales and engines that refused to start. Three attempts had failed. When a narrow break in the weather finally opened, the commander wanted to wait. One of the crew disagreed, and by the terms of the expedition the call was hers. They would go now.
For twenty hours and forty minutes they flew out over the open Atlantic, through cloud, no radio, fuel running down. That year, several people had died attempting the same crossing. When the seaplane set down off the coast of Wales, a crowd was already forming.
She was not the first to cross the Atlantic by air, nor the second, by 1928 the ocean had been flown many times, and those firsts were taken. What was still unclaimed was a category she could own outright: the first woman across, and four years later, the first woman to fly it solo.
Her name was Amelia Earhart.
She had the nerve to go after something she could win. If you cannot be first in the race everyone is running, find the race you can win, and win it so completely that your name becomes the category. "The global leader in visual robot programming for manufacturing" is a race you can win. "Robotics software company" is one you have already lost.
The Formula
The [region] leader in [narrow category], measured by [metric].
Everyone judging you, an investor across the table, an enterprise board weighing your proposal, is quietly sorting you into a market map they already carry in their head. If you don't name your box, they will pick one for you, and it will hold a well-funded incumbent already. You find categories by narrowing, not inventing from nothing: start from the maps investors already use and work down: sector, sub-sector, to a specific product category you can credibly lead within seven to ten years. Then pick your North Star metric, the single number that shows whether you are actually winning.
The three levels, and why Level 3 is where you win
L1 · Sector | L2 · Sub-sector | L3 · The product category you lead |
|---|---|---|
Defence tech | Counter-UAS | Affordable, sovereign kinetic interceptors |
Space | Space access / launch | Sovereign small-lift orbital launch |
Advanced materials | Critical raw materials | Circular rare-earth permanent magnets |
Manufacturing and robotics | Physical AI / humanoids | General-purpose humanoids, built in the West |
Enterprise AI | Vertical AI agents | Autonomous procurement agents |
TechBio | Protein engineering | Generative protein design |
HPC | AI datacentre interconnect | Graphene-based optical chip interconnect |
Clean energy | Fusion | Stellarator-based fusion |
Nobody becomes "the leader in Defence"; those levels belong to incumbents. You win at Level 3: narrow enough to be first, but still on a map they recognise.
Categories across the verticals
Defence — not another missile maker fighting the primes, but the one that redefines the category as low-cost, mass-producible interception, the corner TYTAN Technologies is building into. The European leader in affordable, mass-producible air-defence interceptors, measured by cost per intercept.
Space — not "a rocket," but the continent's independent route to orbit, the category Isar Aerospace is racing for. Europe's leader in sovereign small-satellite launch, measured by launches per year.
Advanced materials — not "a materials company," but the self-driving lab that finds the winning materials (like catalysts) before anyone else does, the loop Dunia AI is closing for the energy transition. Europe's leader in autonomous materials discovery, measured by validated material candidates delivered per quarter.
Robotics — not fixed-code machines, but the Western-built general-purpose humanoid that NEURA Robotics is chasing. The Western leader in general-purpose humanoid robots, measured by humanoids in series production.
Enterprise AI — not "AI software," but agents that recover real money, the category Magentic is carving out. The category leader in autonomous procurement agents, measured by euros of spend recovered.
BioAI — not "an AI-for-bio tool," but the design layer itself, where Cradle is planting its flag. The leader in generative protein design, measured by lab-validated designs shipped.
HPC — not "AI hardware," and not another optical module bolted beside the chip, but light woven onto the die itself, the corner Black Semiconductor is opening in graphene. Europe's leader in graphene-based optical chip interconnect, measured by terabits per second per watt.
Energy — not "fusion" in general, but one defensible corner of it, the one Proxima Fusion is betting on. Europe's leader in stellarator-based fusion, measured by the path to first commercial plant.
Two mistakes kill it. Too broad: "an AI platform," "a robotics company" and you are a feature inside a giant's market, where the incumbent wins by default. Too niche and you have a project, not a category. When in doubt, err narrow: Amazon started as a bookshop. And never bury it, your category belongs in the first three sentences you say.
In your first minutes with an investor or customer, you are not really being asked what you do. You are being asked which race you are running, and whether you can win it. Name the category yourself, big enough to matter, narrow enough to own, and hand them the number that proves you are ahead. Do that, and like Earhart, you become the one the world files under a category of your own.
With wishes that they name the category after you,
Martin
Software production moves towards machine speed. But code still needs to be maintained. Architectures need to remain secure. Decades-old enterprise systems need to be understood. Products need to be tested. We looked across Europe for the early-stage companies attacking these harder layers of the software engineering stack.
LogicStar — Zurich, Switzerland
Technology: Autonomous software maintenance using AI agents, program analysis and software verification.
Customers: Software teams managing complex production codebases; named enterprise customers have not yet been publicly disclosed.
Use cases: Reproducing software issues, generating tests, diagnosing bugs, proposing fixes and autonomously maintaining existing applications.
Funding: $3M Pre-Seed led by Northzone, with participation from angels from DeepMind, Snyk, Spotify, Fleet and Sequoia Scouts.
Why it matters: LogicStar is targeting what could become the next major bottleneck after AI-generated code: maintaining the rapidly expanding volume of software already in production. Its combination of LLMs with program analysis and verification could make software maintenance increasingly autonomous rather than simply faster.
→ The founding team includes former DeepCode founders and ETH Zurich professor Martin Vechev; LogicStar has also built its own software-agent benchmarks, including SWT-Bench.
Oplane — Malmö, Sweden
Technology: Agentic security architecture and automated threat modelling for AI-native software development.
Customers: Miro, Tandem Health, and additional enterprise pilots.
Use cases: Mapping software architectures, generating threat models, identifying security requirements and pushing contextual remediation directly into engineering workflows.
Funding: €4.5M Seed led by Seed Capital, with participation from Icebreaker.vc and angels including Neo4j co-founder Emil Eifrem, Robert Lagerström and Joakim Nydrén.
Why it matters: As AI agents generate and modify software at machine speed, traditional security reviews risk becoming a structural bottleneck. Oplane is building an automated security-architecture layer designed to understand system-level risk rather than simply scan individual lines of code.
→ Oplane says deployments have already expanded across several hundred repositories, with thousands of threat models executed within months.
Kodesage — London, United Kingdom & Budapest, Hungary
Technology: AI-native software intelligence and modernisation for legacy enterprise systems.
Customers: Enterprise and regulated-industry customers, including HOLD Asset Management and NETI.
Use cases: Understanding and documenting legacy codebases, generating tests, analysing issues, supporting migrations and converting ageing software stacks including COBOL, Oracle Forms, PL/SQL, PowerBuilder and RPG.
Funding: €2.3M Pre-Seed followed by a $6.6M Seed led by VentureFriends, with Portfolion returning and angels including Christian Szegedy and Mario Götze.
Why it matters: Much of Europe's critical enterprise infrastructure still runs on software written decades ago. Kodesage is applying AI to the brownfield systems that generic coding assistants struggle to understand — opening a route into some of the largest and most persistent IT modernisation budgets in banking, insurance, energy, telecoms and government.
→ Kodesage can run in private-cloud, on-premise and air-gapped environments, making it particularly relevant for regulated and security-sensitive organisations.
Holmes — Ghent, Belgium
Technology: Autonomous AI-native software quality assurance.
Customers: 30 design partners, with product input from engineering and product leaders at Collibra, Luzmo, Upsellplus and Lighthouse.
Use cases: Automatically generating and maintaining end-to-end tests across authentication, checkout, navigation, search, forms, responsive layouts, accessibility and error recovery.
Funding: €1.1M Pre-Seed led by Syndicate One, with participation from NewSchool.vc, RDY Capital, 100IN and founders from Aikido and Showpad.
Why it matters: AI is accelerating software creation faster than manual QA can scale. Holmes is building the verification layer required for an environment in which code is produced continuously by machines and needs to be tested at the same speed.
→ Despite launching only in 2026, Holmes is already working with 30 design partners.
Hypercritical — London, United Kingdom
Technology: AI-generated and verified control software for safety- and mission-critical systems.
Customers: Engineering teams in automotive, aerospace, defence, robotics and other industrial sectors; named customers have not yet been publicly disclosed.
Use cases: Generating software architectures and requirements, producing executable test suites and discovering control algorithms for systems including electric powertrains, robotics and other safety-critical machines.
Funding: £2M Pre-Seed led by Join Capital, with participation from Octopus Ventures, tiny.vc and Plug and Play.
Why it matters: Hypercritical is taking AI software engineering beyond web and enterprise applications into systems where incorrect code can cause physical failure. Its approach combines AI generation with test-driven algorithm discovery and verification — a prerequisite for bringing AI deeper into Europe's automotive, aerospace, defence and industrial value chains.
→ Co-founders Elie Talj and Alexey Barinov previously worked on advanced automotive systems at McLaren and Arrival, giving the company unusually strong domain expertise for AI-native industrial software engineering.
Five key takeaways from the European AI Office's new report, "Enhancing Competitiveness, Sovereignty and Security of the European Union in Frontier AI".
Frontier AI is becoming a strategic asset — and Europe may have only one to two years to avoid falling into a structurally weaker position.
That is the central warning from the European AI Office's new report, "Enhancing Competitiveness, Sovereignty and Security of the European Union in Frontier AI", which synthesises the first meeting of the European Expert Forum on Frontier AI.
The forum brought together more than 100 experts from European AI developers, industry, investors, technical research, academia and think tanks in April 2026. The report is not an official Commission position, but it offers a revealing view into how Europe's frontier-AI problem is increasingly being framed: not just as an innovation challenge, but as a question of competitiveness, sovereignty and security.
Here's 5 key takeaways:
1. Europe's AI problem is increasingly an infrastructure problem. Experts estimate that the EU currently hosts only around 5% of global AI compute, despite accounting for roughly 15% of global GDP. The report argues Europe should aim to close that gap. Compute and the energy required to power it are described as an urgent two-year priority, alongside growth-stage capital, legal clarity around training data and talent retention. Translation: Europe may not have a research deficit. It has a scale-up deficit.
2. The next 12–24 months may matter more than the 2030 targets. The reason is not simply that frontier models are improving quickly. Leading developers are increasingly using their own models internally to accelerate AI research itself — potentially creating a compounding advantage for whoever already sits at the frontier. If AI begins materially helping to build its own next generation, capability gaps may stop growing linearly. Europe can still set ambitious targets for 2030. But the decisions determining whether those targets remain realistic may have to happen now.
3. Sovereignty does not mean Europe has to build everything itself. One of the more interesting shifts in the report is its definition of AI sovereignty: experts disagree on whether Europe should build a fully domestic frontier stack, but broadly converge on four capabilities Europe needs regardless of where the best models are developed: the ability to access, choose, control and benefit from frontier AI. That means securing reliable access to leading systems, maintaining provider choice, building independent European audit and verification capacity, and ensuring that the economic upside does not accrue entirely elsewhere. The report also proposes deeper coordination with trusted partners including the UK, Canada, Japan, South Korea, Australia, New Zealand and India. Sovereignty, in other words, may be less about autarky and more about leverage.
4. Europe should place fewer, bigger technology bets. The report does not suggest Europe should simply build another version of every US frontier lab. Instead, experts point towards selective high-risk, high-return bets where Europe could develop differentiated leadership: open-weight frontier models, more reliable and secure AI, world models, Physical AI and robotics, compute verification and independent AI evaluation infrastructure. The institutional ideas are equally ambitious: a "CERN for AI," DARPA-style funding vehicles or large public-private consortia. But the report carries an important warning: betting on the next paradigm cannot become an excuse for being weak in the current one.
5. Europe already controls important chokepoints — but needs to use them strategically. The report highlights its position in advanced semiconductor manufacturing equipment as one of the clearest examples of existing strategic leverage. Additional leverage could come from hosting more global compute capacity and building leadership in secure, auditable and verifiable AI infrastructure. The implication is significant: Europe does not need to dominate every layer of the AI stack to have strategic power. It needs to understand which assets others cannot easily replace — and protect, expand and deploy those positions deliberately.
The core message of the report is therefore less about building a European ChatGPT than about building a European position of strength across the AI value chain. And according to the experts advising the AI Office, the window to do that is narrowing fast.
Europe produces strong Deep Tech research but often struggles to scale it. Corporate partnerships are a big part of why some ventures scale and others stall.
McKinsey is running a benchmark study that asks the same questions of both start-ups and corporates. What does actually help: integration intensity, commitment (C-level sponsorship, dedicated resources, binding milestones), and alignment on what each side gives and gets?
The survey takes about 8 minutes. Every participant receives the aggregated benchmark before the public release later this year, ahead of a McKinsey report around The Spark 2026. The survey is open until 14 September.
Deep Tech start-up? Take the founder survey! Corporate? Take the corporate survey!
Copenhagen is Calling
More than 10,000 founders, investors, operators and ecosystem players are expected to descend on Copenhagen next week for TechBBQ — one of the Nordics' most important startup gatherings. What started as a small founder BBQ in 2013 has grown into a major Scandinavian meeting point for capital and technology, with dedicated investor matchmaking, an LP Forum, Investor Day and focused tracks spanning deep tech and life sciences, defence & dual-use, fintech and AI.
For founders, the value is less about another two days of panels and more about concentration: international investors, corporates and Nordic startups in one place, with structured matchmaking built around it. If the Nordics are on your fundraising, partnership or scouting map, Copenhagen is probably where you want to be next week.
→ 26–27 August 2026 | Bella Center, Copenhagen
