Twin Prime has raised $10 million in pre-seed funding to build frontier AI models for defence and security. The round was led by Expeditions, with backing from American and European venture firms, THEON, family offices and angels connected to Palantir, Anduril, Quorum and other security technology circles.

The company is young, founded in 2025 by George Lentzas, Stephane Sezer, Drew Calcagno and Michael Leite-Garcia. Its pitch is not another general AI assistant wearing camouflage. Twin Prime says it is developing multimodal models that can reason across sensor data in the physical world and compress the distance between perception and decision. In defence, that distance can be the whole product.

Here’s the awkward truth. Most AI coverage still treats the interface as the story. Defence buyers do not. They care about latency, robustness, edge deployment, messy inputs and whether a system works when conditions are ugly. Mud, fog, noise, jamming, panic. The opposite of a demo room.

The Tech.eu report describes Twin Prime as a lab focused on models that fuse data from many sensor modalities. Expeditions co-founder and GP Mikolaj Firlej argued that current AI models are not specialized enough for modern war and wider security challenges, especially at the edge. That is the market opening Twin Prime is trying to occupy.

The Nordic relevance is not a neat headquarters story. It is a capital and security story. Expeditions has become one of the more visible European investors in defence and resilience startups, while Nordic governments and industrial groups are rethinking procurement after Russia’s invasion of Ukraine. The region has discovered that resilience is not a policy word. It is a product category.

Specialization also changes the venture math. A defence AI company can’t only chase users and usage. It needs security clearances, pilots, procurement trust and hardware partners. The sales cycle can be slow, then suddenly very large. Strange rhythm.

A separate THEON release said the company is making a $3 million strategic investment in Twin Prime and plans a joint venture owned 60% by THEON and 40% by Twin Prime. The goal is to integrate AI models into THEON products and accelerate internal AI research and development.

That matters because defence AI cannot live as a slide deck. It has to attach to sensors, optics, platforms and operational workflows. THEON’s electro-optical products give Twin Prime a potential path into physical systems rather than pure software pilots. It also gives THEON a way to avoid treating AI as a feature bolted on after the hardware is already designed.

The risk is obvious. Joint ventures can become bureaucratic furniture if incentives drift. But the signal is still useful: European defence primes and suppliers know they need software-native partners, and startups know they need routes into procurement and deployment. Neither side gets to pretend the other is optional.

Twin Prime’s founding team reads like a deliberate rejection of single-domain thinking. Frontier AI, quantitative finance, national security, government and military backgrounds all show up in the company’s story. That mix is not decorative. Defence AI sits at a place where model performance, probabilistic thinking, operational constraints and institutional trust collide.

Quantitative finance veterans know what it means to make decisions under uncertainty and latency. National security operators know that context can be incomplete and consequences asymmetric. AI researchers know how fragile models can become outside a training distribution. The job is to make those instincts live in one product culture.

That is also why generic model wrappers feel insufficient. A defence customer does not only ask whether a model can label an image. It asks what happens when three sensors disagree, when bandwidth drops, when the adversary adapts, and when a human operator needs confidence fast. Those questions are deeply unglamorous. They are the market.

The presence of THEON is a reminder that defence startups do not scale the way consumer startups do. A brilliant model without a route into hardware, field testing and procurement may remain a research artifact. A hardware supplier without advanced software risks becoming a component vendor in a market that increasingly values integrated decision support.

The planned joint venture is therefore a strategic shortcut for both sides. Twin Prime gets a path toward deployed systems and product feedback. THEON gets access to specialized AI without pretending it can build every layer internally. The structure may look old-fashioned, but defence is full of old-fashioned mechanisms because accountability matters.

For Nordic and European investors, this is one of the sharpest lessons of the new defence wave. The best companies may need to partner early with primes, governments and dual-use customers. Independence is valuable, but isolation is expensive.

There is a civic side to the story. Defence AI funding is rising because threats are real, but the category will draw scrutiny. Voters and policymakers will ask where autonomy begins, how humans remain accountable, what data is used and whether private companies should shape battlefield decision systems. Those questions cannot be postponed until after deployment.

A company like Twin Prime can help itself by being precise about what it does. Sensor fusion and decision support are not the same as autonomous weapons. Edge intelligence is not a magic wand. Clear language matters because vague language invites fear, and in defence, fear can slow adoption as quickly as technical failure.

The most successful European security startups may be the ones that combine technical ambition with institutional maturity. They will need to speak engineer, soldier, regulator and citizen. Not many teams can. That is why the category is hard, and why the winners could matter.

A lot of enterprise AI can fail gracefully. A summary is mediocre. A chatbot misunderstands a prompt. A workflow needs a human check. Edge defence systems have less room for graceful failure. They may operate with limited connectivity, imperfect sensors and pressure that makes long deliberation impossible.

That is why Twin Prime’s focus on perception-to-decision compression is important. Speed is not useful without trust, and trust is not useful if it arrives too late. The technical work sits between those poles. Models need to process messy inputs, expose uncertainty and fit into human command structures that already have doctrine and procedure.

The companies that solve this will probably not look like consumer AI winners. They will be quieter, more integrated and more willing to do painful field work. Less viral. More consequential.

Ten million dollars is not much money in frontier AI terms. It is also not much money in defence procurement terms. Yet the round is meaningful because it points to where budgets may move. Defence ministries and suppliers are looking for software that improves existing systems rather than waiting for entirely new platforms. That favors companies that can plug intelligence into sensors and workflows already in the field.

This is where European startups may have an opening. The continent has deep industrial defence assets, but it has often lagged in software speed. Startups can bring that speed if they respect the constraints of the sector. Move fast and break things is not a defence doctrine. Move carefully and integrate well is less catchy, but it sells.

Twin Prime will now have to prove that its technical ambition can survive customer specificity. Each defence customer will have different sensors, data policies, threat models and integration demands. A model lab can celebrate generality. A defence supplier gets paid for useful specificity.

There is a temptation to describe every defence AI company in science-fiction terms. Twin Prime’s work sounds more prosaic and more consequential: make sense of sensor feeds quickly enough that operators can act. That includes data quality, compute constraints, integration and trust. None of those are tidy.

The unexpected angle is that the best defence AI companies may look less like model labs and more like translators between physics and command. They will need to understand data from cameras, radar, infrared, acoustic systems and whatever else modern battlefields throw at them. Then they need to make the output boringly reliable.

The immediate next milestone is not a glossy benchmark. It is likely a fielded pilot where a customer can compare Twin Prime’s output with existing workflows. Defence buyers are skeptical by design, and they should be. A model that looks impressive in isolation has to prove it can sit inside a chain of command without adding confusion.

That makes the round a beginning rather than a validation stamp. The validation will come when operators trust the system enough to change how they work. In defence AI, adoption is measured in procedures, not downloads.

Twin Prime’s round lands in a Europe that is spending more on defence while still trying to build its own technology base. For Nordic readers, the takeaway is clear: security tech is no longer a side channel in the startup market. It is becoming one of the main routes through which deep tech gets funded, tested and bought. Original report