NestAI launches defence models for European drone autonomy

NestAI, the AI lab founded by Silo AI cofounder Peter Sarlin, has released its first models for military use, with an initial focus on drone autonomy and battlefield mission planning. The Helsinki-linked company raised €100 million in November from Nokia and Finnish state-owned investor Tesi, but has not disclosed revenue, valuation or commercial contract values.

The launch puts NestAI into the crowded but politically favored European defence tech field, where startups are pitching domestic alternatives to US-controlled AI infrastructure. Sarlin told Sifted the company’s aim is to reduce Europe’s dependence on foreign providers for the model layer used in military systems.

NestAI was founded in 2025 after Sarlin sold Finnish AI lab Silo AI to AMD in 2024, in a deal Sifted described at the time as Europe’s highest-value AI acquisition. NestAI has grown to 200 employees over the past year and has hired people with backgrounds at Tesla, Nokia, Intel and Silo, according to the company.

Drone models and mission orchestration

The company is building two main products. One is a set of foundation models for autonomous drones that can run at the edge. The other is an orchestration model delivered through NestOS, its existing platform, for planning and executing missions involving unmanned systems.

Sarlin said the training data combines synthetic datasets with real-world data. The company has not detailed the size of those datasets, the model architecture, performance benchmarks or the extent to which the systems have been tested outside pilots.

NestAI is currently running pilot programs with the Estonian and Finnish armed forces. Sarlin said the platform can support end-to-end missions with unmanned drone fleets and handle tasks armed forces want drones to carry out. The company’s longer-term plan is to work with other allied militaries and use operating data from deployments and pilots to keep improving the models.

Sarlin framed the technical problem as closer to autonomy in changing physical environments than to general-purpose chatbot development. He told Sifted that battlefield systems must cope with conditions that can change quickly, including damaged terrain and altered maps after attacks. NestAI’s pitch is that domain-specific models, rather than general-purpose frontier systems from companies such as Anthropic or OpenAI, can be competitive for defence use cases.

Sovereignty pressure after Anthropic export controls

The timing follows a US government-ordered suspension on exports of Anthropic’s most advanced models, an episode that intensified European debate over reliance on foreign AI providers for critical systems. Sarlin said the incident showed why European militaries need to own and control the foundation models behind military capabilities.

The war in Ukraine is the operating reference point for NestAI’s defence work. Data compiled from Ukrainian Air Force reports cited by Sifted showed Russia launched about 13,300 air attacks involving drones, missiles and cruise missiles against civilian targets in Ukraine in 2024. In 2025, that figure rose to roughly 56,700, with the increase driven largely by drones.

Sarlin said European defence procurement and development still often run at peacetime speed, while Ukraine has connected research and development more directly to battlefield feedback. That is the model NestAI says it is building around, although it has not said how quickly its own models can be updated or redeployed in operational settings.

Compute partnerships

NestAI is also leaning on external infrastructure. The company has a partnership with AMD for compute capacity and another with Finland’s LUMI AI factory, where it plans to use the LUMI supercomputer to train models.

It will also work with Qutwo, a quantum startup cofounded by Sarlin. Sarlin said Qutwo simulates quantum computing on GPU clusters to compress large AI models without reducing performance, with the goal of running more capable models on lower-cost edge hardware. The company has not disclosed technical results for that compression work.

This story draws on original reporting from Sifted.