Edge AI Is Leaving the Demo Stage, and Funding Is Following

For most of the past decade, the AI story ran through the cloud: bigger models, bigger data centers, bigger bills. That's still true for training. It's less true for where AI actually runs once it's deployed. Across manufacturing floors, drone fleets, vehicles and security cameras, more inference is happening on the device itself, not in a data center three hops away. Vendors have started calling the broader trend "physical AI" — the point where generative reasoning meets systems that have to act in the real world — and it shows up in both the hardware roadmaps and the funding numbers.

Estimates of the edge AI market's size vary meaningfully depending on who's counting and what they include. Grand View Research and Precedence Research put the 2026 market somewhere in the $25–30 billion range, while other analysts put the figure closer to $37 billion. Most agree the market is growing at a compound annual rate somewhere between 21% and 29% through the early 2030s, with several projecting it to clear $100 billion by 2030–2033. Treat the exact number as directional rather than precise; the underlying trend is the more reliable part of the story. Hardware — chips, NPUs and accelerators — reportedly accounts for roughly half of edge computing revenue today, which says this is still as much an infrastructure buildout as a software one.

Funding data points the same direction. Between August 2025 and July 2026, edge AI startups raised roughly $2.55 billion across about 21 disclosed deals, according to an analysis by New Market Pitch. Chip and inference-hardware companies — names like Hailo, SiMa.ai and Axelera — accounted for close to $1.9 billion of that on their own. Security and surveillance applications running inference on-device raised over $2.8 billion combined. Autonomous vehicle and logistics players raised considerably more than either category, north of $8 billion, though a good share of that is buying autonomy stacks broadly rather than edge compute specifically.

The technical case for edge AI has existed for years. The constraints pushing workloads there have simply gotten harder to ignore. Real-time decisions — a drone avoiding an obstacle, a robotic arm adjusting mid-motion — can't tolerate the round-trip latency of a cloud call. Privacy rules, GDPR chief among them in Europe, make it costly to ship sensor or camera data off-site when it doesn't need to leave the building. And connectivity can't be assumed: a factory floor, a construction site or a moving vehicle will lose signal, and any system built around a permanent link to the cloud stalls right when it's needed most.

The industry has responded by standardising around purpose-built edge silicon — NVIDIA's Jetson and DRIVE Thor platforms are the most commonly cited examples — and by rethinking how models are trained in the first place. Running inference at the edge is the easier half of the problem. Updating a model once it's already deployed, without shipping it back to a data center every time conditions change, is the harder one, and it's where a lot of the current engineering effort is going.

Rhizome, a Paris-based company Critical Ventures backed in 2025, is built around exactly that problem. Its platform lets AI models keep learning directly on the hardware they run on — drones, industrial sensors, robotics — instead of requiring a cloud retraining cycle every time the environment shifts.

Almetra, a Berlin-based company also from our 2025 vintage, works on a narrower but related problem: giving manufacturers real-time visibility into what's happening on the shop floor using cameras alone, with no additional sensors and no line disruption. Its system is designed to meet union-approved data protection standards — a detail that matters for adoption inside large European manufacturers as much as the technology itself. It's a smaller-scale example of the same shift: intelligence that runs where the work happens, not in a dashboard that updates an hour later.

We think the founders worth backing in this space are building for constraints — power budgets, intermittent connectivity, hardware that can't be swapped out easily — rather than assuming compute is cheap and connections are stable. That's a harder engineering problem than shipping another cloud-hosted model, and it's a large part of why we were drawn to teams like Rhizome and Almetra. It also fits squarely within our thesis on where durable value gets built in applied AI.

If you're building infrastructure, tooling or applications for AI that has to run outside the data center, we'd like to talk. Reach out through our deal submission page.