For most of the last decade, the hardware side of robotics was treated by institutional investors as a garnish on the software plate. Autonomy stacks, perception libraries, simulation platforms — those were the things that attracted tourist capital. Anything with a bill of materials above $15,000 was quietly filed under 'too capital intensive' and handed off to the one partner at the firm who had an engineering degree.
That pattern broke in 2025. According to PitchBook, global venture funding into hardware-inclusive robotics crossed $18.4B, up 71% year-over-year, while software-only robotics rounds grew just 9%. The shift is not a blip. It is being driven by three forces that we believe are structural.
The supply side finally caught up
Five years ago, a mid-complexity humanoid or mobile manipulator took 18–24 months to go from CAD to a fieldable prototype. Today, teams coming out of Shenzhen, Guadalajara and Gdansk routinely ship v1 hardware in under nine months. Contract manufacturers that used to demand 10,000-unit minimums are quoting 200-unit pilot batches at margins that weren't available in 2020.
That compression of the prototype loop has done for robotics what AWS did for software in 2008: it has made iteration cheap enough for venture math to actually work.
You used to underwrite a robotics deal assuming three prototype generations would eat $12M. Now it's closer to $4M. That single change rewrites every line of the model.
LLMs solved the wrong half of the problem — and that is good news
Large models turned out to be extraordinary at planning, language grounding and high-level task decomposition. They are still mediocre at the low-level control problem: grasping, contact-rich manipulation, sub-millimeter insertion. The market has internalized this. Foundation-model companies are not going to vertically integrate into warehouses and hospitals. Hardware-native teams will.
This is why we saw Physical Intelligence, Skild, Figure, 1X and Apptronik close rounds at valuations that would have been unthinkable in 2023 — and why the next cohort of Series A rounds is being led by generalist funds, not corporate strategics.
What 'structural' actually means for allocators
For an LP deciding between a sleeve allocation to AI-native funds and one to hardware-robotics specialists, the critical question is no longer 'will hardware robotics produce outlier returns' but 'which managers can actually diligence bill-of-materials risk.' That is a narrower shortlist than most LPs realize.
- Median Seed round size in hardware robotics rose from $3.1M in 2023 to $6.8M in 2025
- Time between Seed and Series A compressed from 22 months to 14 months for the top quartile
- Japanese, Korean and Gulf sovereigns accounted for 28% of Series B capital in the sector
- Eight of the ten largest 2025 rounds included a strategic manufacturing partner on the cap table
Where we think the next 18 months go
We expect a bifurcation. The top tier of generalist humanoid and dexterous-manipulation companies will continue to raise at frontier-AI valuations and will be priced primarily on talent density and data flywheel. Below that, a much larger set of application-specific hardware teams — logistics, construction, agriculture, eldercare — will attract investors who care about payback periods and service contracts.
The decade-long disconnect between where the value was created in robotics and where the capital was deployed is closing. That is uncomfortable for investors who spent the last cycle developing pattern matching around pure-software SaaS — and it is exactly the kind of dislocation that produces the best vintages.