Chinese Automakers Bet on Humanoid Robots: What It Means
Xpeng, BYD and Chery are pivoting into humanoid robots because car margins are thin. The bodies are getting cheap — the control software is where the skill premium moves.
Chinese automakers are piling into humanoid robots, and the reason is not that robots suddenly work. It is that cars stopped paying. TechCrunch reported on 28 August 2026 that Xpeng, BYD, Chery and several others are all building humanoids, following the hype Elon Musk generated with Optimus.
I want to pull one thing out of that story, because it decides what a Sri Lankan engineer should actually do about it: the companies rushing in have the bodies solved and the brains unsolved. That gap is the whole opportunity.
🏭 Who is actually in the race
The list from the reporting is longer than most people realise, and it is almost entirely carmakers:
| Company | Robot / unit | What's publicly known |
|---|---|---|
| Xpeng | Iron | Robotics unit raised over $900M at a $6.3B post-money valuation |
| Chery | AiMOGA | Robotics unit reportedly began preparing for an IPO in August 2026 |
| BYD | Xiao Di | Humanoid robot unveiled |
| Changan, GAC, Li Auto, SAIC, Seres | — | All developing humanoid robots |
The Xpeng round was led by IDG Capital, with Gaorong Ventures, Tencent and Alibaba participating. Founder He Xiaopeng and co-president Brian Gu put in roughly $100 million of their own. Xpeng called it the largest single-round private financing recorded in China's "embodied AI" industry. That is the company's characterisation, not an independently audited claim, and I'd read it as such.
💰 The motive is margin, not magic
Michael Dunne, CEO of Dunne Insights, gave the clearest read in the piece. He said Xpeng's founder "sees razor-thin profit in cars on the near horizon. Robots look much more promising."
That is a very different sentence from "humanoid robots are ready." It's a sentence about where the next margin might be when your current business is being competed to zero.
Key takeaway: These are not robotics companies that happen to make cars. They are car companies escaping a price war, carrying an actuator-and-battery supply chain with them. The capability follows the desperation, not the other way round.
Dunne's second observation is the one worth pinning to your wall. Chinese automakers, he said, "have all the hardware to get the job done" — but the open question is whether "they can catch Tesla on the AI side."
🧠 Bodies get cheap, brains do not
An electric car and a humanoid robot share most of a bill of materials: motors, reduction gears, battery packs, power electronics, sensors, and a factory that can assemble all of it at volume. If seven or eight Chinese automakers are attacking that at once, the predictable result is that robot bodies get cheaper, faster, on the same curve EV batteries did.
The control stack does not fall the same way. Here's how I'd split it:
| Layer | Capital needed | Gets commoditised? | Can you practise it from Sri Lanka? |
|---|---|---|---|
| Actuators, gearboxes, chassis | Enormous — needs a factory | Yes, quickly | No |
| Battery + power electronics | Enormous | Yes | No |
| Sensors, compute modules | Large | Yes | Partly |
| Perception, control policy, data pipeline | Small — mostly compute time | Slowly, if at all | Yes |
| Task-level reasoning and safety | Small | No | Yes |
Look at the bottom two rows. That is the half of the industry where a person with a laptop, a free-tier GPU quota and time is not structurally locked out. You cannot compete with BYD on gearboxes. You can absolutely learn to write the policy that runs on one.
🛠️ What to actually learn, in order
If you're a student or a small-team builder here and you want a real position in this rather than a LinkedIn opinion about it, the sequence I'd follow:
- Simulation before hardware. Open-source physics simulators run on a normal machine. You can build, break and retrain a controller a thousand times without owning a single servo. Hardware teaches you patience; simulation teaches you iteration speed. Start with speed.
- Classical control before learned control. Kinematics, PID, state estimation. Teams that skip this end up unable to explain why their learned policy fell over.
- Vision-language models as the task layer. The interesting part of current humanoid work is not the walking — it's turning "put the cup on the shelf" into a plan. That's an LLM problem, and it's the part you can prototype today against an API.
- Data pipelines. Demonstration data, logging, replay, evaluation. Unglamorous, and the single most transferable skill on this list.
- Deployment constraints. On-device inference is memory-bound before it is compute-bound. Before you assume a model fits on an embedded board, size it first: our LLM VRAM calculator does that arithmetic for you, and the GPU buy-vs-rent calculator is worth running before anyone talks you into buying a workstation.
Renting compute in hourly blocks and shutting it down is almost always the right call at learning stage. Idle owned hardware is the most expensive thing in a small team's budget.
🌐 Why this reaches Sri Lanka faster than the robots do
No humanoid is walking into a Colombo warehouse soon, and I'm not going to pretend otherwise. The transmission runs through work, not through units sold:
- Simulation, annotation and evaluation work is remote-friendly. It needs judgement and English, not a factory next door. Sri Lankan teams already sell that kind of work.
- Cheap bodies mean cheap parts. A price war among eight automakers over actuators and sensor modules eventually shows up as affordable components for a university lab or a hardware startup here.
- The software is where the durable jobs sit. If the hardware layer commoditises and the AI layer doesn't, the wage premium sits in the layer you can learn without capital.
- The claims will outrun the results. A market where several companies are raising against the same story will produce demos that don't survive contact with a real workplace. Being the person who can tell the difference is itself a career.
What this means for you 💡
Take this as a signal about where skill will be scarce, not as a reason to buy anyone's stock or believe anyone's demo timeline. The story here is that a group of large, capable manufacturers just admitted their core business is running out of margin, and bet the next one on a control problem they have not yet solved.
Three things I'd act on:
- Don't chase the hardware. Competing with a Chinese automaker's supply chain from Sri Lanka is not a plan.
- Do build the software half. Simulation, perception, planning, evaluation. All of it is learnable on a laptop plus rented compute, and all of it transfers to non-robotics work if humanoids underdeliver.
- Keep your scepticism calibrated. Funding rounds measure belief, not capability. A $6.3B valuation tells you what investors expect, not what the robot can do.
Bottom line: the bodies are being solved by people with factories. The brains are still open. Go where the problem is still open.