The platforms are buying. The question is who's building what they need.
Shield AI raised $2 billion at a $12.7 billion valuation this week. Revenue above $540 million, growing 80%+ year on year. Combat-validated AI autonomy across drones deployed in Ukraine. Selected for the U.S. Air Force's Collaborative Combat Aircraft programme. Impressive round, real traction.
But the move that caught our attention was buried in the announcement: Shield AI is acquiring Aechelon Technology, the simulation company behind the Pentagon's Joint Simulation Environment. High-fidelity flight simulation, physics-based sensor modelling, synthetic training environments. The simulation capability Shield AI needs but can't build fast enough internally.
This is a pattern, not an anecdote.
The platforms are acquiring their enabling layers
ABB and NVIDIA announced RobotStudio HyperReality earlier this month: 99% sim-to-real correlation for industrial robots. Google DeepMind partnered with Agile Robots to push foundation models onto factory floors. Shield AI just bought its simulation layer outright. Anduril ($60B valuation, ~$2B revenue) is building its own through Lattice OS.
Every major Physical AI platform is reaching the same conclusion: the hardware works, the autonomy stack works, but the training pipeline, the simulation fidelity, the edge compute, and the vertical skill libraries are bottlenecks. They're solving those bottlenecks by acquiring or partnering with smaller, specialised companies that got there first.
Where the early-stage opportunity sits
At Unlock Capital, we invest at the enabling layer. We're not writing cheques into $12.7 billion Series G rounds. We're looking at the companies building the components that Shield AI, Anduril, ABB, and the next wave of Physical AI platforms need to buy or build.
In our pipeline, we're seeing early-stage companies working on exactly these bottlenecks: simulation environments that generate synthetic training data for autonomous systems, and edge compute architectures that run AI models on hardware where cloud connectivity doesn't exist. These are small teams solving hard physics and engineering problems, often with deep domain expertise from defence, mining, or industrial automation.
The Aechelon acquisition validates the thesis. A $12.7 billion platform decided it couldn't build its simulation capability fast enough, so it bought a specialist. That buy-vs-build pressure only increases as more platforms reach production scale and need to train faster, simulate more scenarios, and deploy to more edge environments.
The Smil Test
Remove Hivemind's AI and the drone stops functioning. That's Essential AI at the substrate level. The same test applies down the stack: remove the simulation pipeline and the AI can't be trained. Remove the edge compute and it can't be deployed. The enabling layer is essential to the essential.
That's where we're focused.