If you're following the humanoid robot space, you're probably aware that a few Chinese companies are becoming extremely dominant. There are two companies in particular, Unitree and Engine AI (and to a lesser extent, AGI Bot), which will start shipping very strong sub-$6k consumer-oriented robots to the US in Q1 2026, and intend to dominate the humanoid ecosystem in the same way that DJI dominated the early drone ecosystem. I think they will probably succeed - they have very high-velocity, well-capitalized engineering teams, with the ability to do both hardware and software very effectively.
On one hand, I think this is pretty cool, because I think a good, low-cost humanoid robot will be the most transformative hardware product of my lifetime, and it will be here much more quickly than many people expect. On the other hand, what the world looks like by this time next year, I expect, is thousands of hackers and startups beta testing Unitree and Engine AI robots in diverse scenarios - robot boxing, elder care, apple picking, logistics, home security, and more - in the same way that early adopters tried out drones for many applications. They will provide the cash flow, supply chain effects and product roadmap to cement one or both of these companies as the robot of choice for anyone who wants to experiment with general-purpose robots.1
Besides the obvious downsides - like Unitree working with the Chinese military, supply chain aggregation, and siloing of humanoid robotics expertise - this really concerns me as someone who wants to build great robotics AI models. The basic difference between robotics and other AI domains is that while the Internet provided the train tracks for other domains to ride on, those same train tracks do not exist in robotics. With LLMs, both data acquisition and value creation can be scaled very quickly thanks to the internet and data centers, but with robots, you need to have a big fleet before you can really start to build interesting machine learning models, or start to monetize those models. In that way, my expectation is that the development of AI models for robotics will be more similar to self-driving cars than to ChatGPT, and there will be strong network effects for whoever can get to fleet-scale first, stronger than any other consumer electronics products built before. Put differently, robotics companies with positive cash flow, fleet-scale data, and strong engineering cultures will be able to translate methodological approaches from any competitor or research lab into their own flywheel, begetting strong network effects for first movers. I think that right now these network effects will probably aggregate to Unitree, and it will be almost impossible for anyone else to build a viable general-purpose robotics business.
So, that brings me to my company, K-Scale. The basic premise for K-Scale is that I believe a Silicon Valley-based AI startup which can partner effectively with Chinese hardware startups would have the potential to take on not only Tesla, but also Unitree and Engine AI, by leveraging the strengths of both ecosystems. To that end, our goal is to commoditize the humanoid robot as quickly as possible, by building open-source robot reference designs, finding great hardware companies that want to build their own humanoid, and working with them to get a competitive product to market. We're able to move pretty quickly and cheaply this way. Getting a hardware product to market on a $4 million seed round speaks to how capital-efficient this approach is, while also keeping our balance sheets clean of the sort of risky capital expenditures that typically kill early-stage hardware companies with variable demand.
But there is a more important goal here, besides shipping a lot of robots. In the same way that a free and open internet laid the foundation for most of AI today, a free and open "robonet" will lay the foundation for the future of embodied AI, if executed well. This robonet will provide both the vehicle for capturing lots of data, as well as the vehicle for monetizing embodied intelligence. The data for training LLMs was created by people using the internet in interesting ways, and the most likely path towards large-scale robotics data creation will be along a similar route - something like "Reddit for humanoid robot data".2
Based on experience, I strongly expect that it will be very difficult to build an extremely capable embodied intelligence without first building this sort of infrastructure layer. There are some interesting research ideas which might succeed in leveraging large-scale video data for robotics pre-training which are exciting, but even if we can solve the research problem, getting these ideas to work at scale on real robots will still present its own challenges. Fortunately, this is an exciting and motivating thing to build, it is something that no one else is building, and it's something that K-Scale is uniquely positioned to contribute to the world, which is great for getting smart people to work really hard.
I believe it is the only strategy that is likely to succeed in the current ecosystem long-term, where even Tesla is struggling. I think that the closer you are watching the hardware ecosystem, the more likely you are to agree with this. Anyone working on humanoid robot hardware today should probably assume that IP and R&D ownership will be nebulous at best, even for Chinese firms. In scaling to mass-market production, it will be almost impossible for manufacturers to avoid this, except for bespoke, high-margin applications like defense or other things that overlap with government operations. On top of that, there are captive audiences for a robotics platform which is not built around a company like Unitree, which K-Scale can use as a foothold.
It is clear that we will need to accelerate our engineering efforts if we do not want to be eaten alive, which requires engineers and capital. We can leverage an exciting mission to find great, motivated people, but at the same time, there is a conundrum in raising significant capital for an idea which amounts to giving away a lot of intellectual property, with the goal of shaping the ecosystem.
One avenue I've been thinking about is partnering with some well-capitalized companies that are interested in building a sort of "open humanoid consortium", in the way that Android leveraged Google's backing to build an open smartphone ecosystem. There are natural partners here among OEMs who benefit from a heterogeneous ecosystem with common standards and shared intellectual property. I suspect that the real strength of such a consortium would come once there is an acknowledgement from a few key players that the American robotics ecosystem is trailing China, and that building large, in-house, proprietary R&D teams will probably be a waste of capital, and will likely not lead to a competitive advantage. This might become an easier sell as firms like Unitree become more and more dominant and competing products fail to reach meaningful market share. Such a consortium, or even simply the backing of one or two well-capitalized companies, would provide a lot of leverage with different suppliers (as OpenAI is doing) and make executing on this strategy much easier.
If you find this idea interesting, please reach out. I am very actively exploring what the right way forward looks like - I think we're on the precipice of potentially the most exciting technology race in human history, and I think it is the right time to make a large impact.
Footnotes
It's worth mentioning that there are a few companies who are trying to build foundation models for robots - something like one brain for all robots. I think this is fairly naive, personally (I'm not even sure if these companies still believe that this approach is correct) - I doubt it will be possible to make rapid progress on building ML systems for robots without building your own hardware. As evidence, you can see how 1X was able to build full-body visio-locomotion before Skild, despire having raised less money and with a smaller (and ostensibly weaker) machine learning team. Put differently, I don't think that there are many methodological moats in machine learning these days, and the value of having a team of people who are very good at churning out research papers is pretty low compared to the value of having a bunch of highly-motivated product hackers, like the team that Unitree has built. ↩
We can contrast this approach with a more pilot-first approach focused on using a bespoke solution to build a data flywheel with specific verticals, before expanding. The principle differentiator here is that the internet-style approach unlocks significantly more experimentation in what real product-market fit will look like. Given that we are in the very early days for humanoids, it's probably a better idea to explore the space very broadly and listen carefully to early adopters than to over-index on a few bespoke customers. ↩