NobodyWho, a Copenhagen-based startup, has raised €2 million to help developers run small AI models directly on smartphones—shifting more everyday AI features away from the cloud and onto the device in a bid to improve speed, privacy, and reliability.
The funding underscores a growing push toward on-device AI, where tasks like text generation, summarisation, and smart assistants can operate locally even when connectivity is patchy or when users prefer not to send sensitive data to remote servers. While large cloud-hosted models still dominate the AI landscape, a new generation of compact models—optimised for mobile chips—has made local inference a practical target for startups and platform players alike.
Why on-device AI is gaining momentum
The appeal of running AI on a phone is straightforward: reduce the round-trip to a data centre, lower ongoing compute costs, and keep data closer to the user. For consumers, that can translate into snappier features and fewer privacy trade-offs. For businesses, it can mean less spending on cloud inference and fewer compliance headaches when handling personal or regulated information.
In the last year, smartphone makers and chip companies have also leaned hard into edge computing and neural processing units (NPUs), marketing devices as “AI phones” capable of handling more tasks locally. That hardware trend is opening a lane for software companies like NobodyWho to provide tooling that makes it easier to deploy and manage compact models on mobile devices at scale.
What NobodyWho says it’s building
Based on the deal headline, NobodyWho is positioning itself around enabling local execution of small AI models on phones—an area that typically involves model optimisation, quantisation, runtime performance, and device-specific compatibility. The practical challenge is that mobile environments vary widely: different chipsets, memory limits, thermal constraints, and operating system rules can all affect whether a model runs smoothly.
To be useful to product teams, an on-device AI layer must do more than “run a model.” It needs to help developers ship features that behave consistently across devices and update safely over time. That can include:
- Model compression techniques to reduce size and memory usage
- Quantisation to speed up inference with minimal accuracy loss
- Device-aware performance tuning to avoid overheating and battery drain
- Packaging and deployment workflows that fit app store policies
- Monitoring and fallback logic when a device can’t meet performance targets
While the fundraising note does not provide full product specifics, the direction aligns with a broader industry shift: moving from “one model in the cloud for everyone” to a hybrid approach where a phone handles what it can, and the cloud handles what it must.
The economics: cutting cloud inference costs
One of the less visible drivers of the on-device push is cost. Running AI features in the cloud can create an ongoing bill that scales with usage. For consumer apps, viral growth can turn inference into a major expense line. For enterprise products, predictable unit economics matter, especially when customers expect fixed pricing.
By moving some inference onto the user’s device, companies can reduce dependence on paid compute—though it introduces new costs in engineering, optimisation, and QA across devices. The bet behind startups like NobodyWho is that tooling and infrastructure can standardise that complexity, making local AI cheaper and easier to adopt.
Privacy and regulation: keeping sensitive data local
On-device AI also speaks to privacy concerns that have intensified as AI features expand into messaging, email, photos, and voice. When inference happens locally, fewer raw inputs need to leave the device. That can reduce exposure in the event of a breach and help companies minimise the amount of personal data processed in the cloud.
In Europe, where compliance expectations are high, local processing can be attractive for teams navigating data protection obligations and user expectations. However, on-device AI is not a silver bullet: apps still need to be transparent about what data is collected, what telemetry is sent, and how models are updated—especially if updates change behaviour in ways users can notice.
Technical hurdles: performance, battery, and fragmentation
The hard part of phone-based AI is not the demo; it’s the day-to-day experience. Even small models can strain older devices, and sustained inference can heat up a phone or drain the battery. Developers also face fragmentation between iOS and Android, plus wide variance across Android hardware tiers.
That’s why the “small model” category matters. Compact models can be tuned for specific tasks—like extracting entities from text, generating short responses, or performing lightweight summarisation—without attempting to replicate the full breadth of large general-purpose systems. In many products, that narrow focus is enough to deliver user value while staying within mobile constraints.
What the €2M round signals for the market
A €2M raise is modest by late-stage standards, but it is meaningful in a market where many early AI startups are expected to show clear differentiation beyond “wrapping a model.” Funding into on-device infrastructure suggests investors see room for enabling layers—tools that help other companies ship AI features reliably and cost-effectively.
It also reflects a more mature phase of the AI cycle: after the initial wave of cloud-first experimentation, teams are now wrestling with latency, cost, privacy, and product stability. If NobodyWho can reduce the friction of deploying on-device AI, it could benefit app developers looking to keep AI features responsive while limiting cloud dependence.
For users, the near-term impact will be measured in small moments—features that work faster, continue offline, and keep more personal data on the phone. The next test for NobodyWho will be turning that promise into tooling that developers can ship at scale across the messy reality of modern mobile hardware.