Unitree, the Chinese humanoid robot maker known for backflips, dancing machines and now a robot called "Superman", raised RMB 6.1 billion in Shanghai in an IPO reportedly oversubscribed more than 8,000 times by retail investors. On debut, the stock surged as much as 629%.
There is a reason the enthusiasm is building now. Beijing hosted the second World Humanoid Robot Games, a showcase built around exactly the sort of spectacle that makes a young industry feel inevitable: robots running, dancing, competing and performing practical tasks.
China increasingly looks like the place trying to industrialise humanoid robots. The United States still has many of the ingredients that matter most in AI: frontier models, leading chips, top research and deep software capability. But China has been assembling a different advantage. It has manufacturing scale, a deep hardware supply chain, intense domestic competition and a political system willing to support sectors it sees as strategically important.
Building a convincing demo is one thing. Building thousands of units cheaply enough for commercial adoption is another. China has already shown in solar, batteries and electric vehicles that it can move from curiosity to cost advantage very quickly. The robot race may follow a similar path.
Humanoid robots still face three obvious problems.
First, reliability. A backflip in a controlled environment is impressive. Replacing a worker in a warehouse, factory, hospital or hotel is harder. Real work involves repetition, judgment, error handling and operating safely around humans.
Second, data. Robot brains need vast amounts of real-world experience, not just simulations. That is one reason embodied AI is attracting so much attention. The more robots move through factories, shops and homes, the more usable data they generate. Scale itself becomes an advantage.
Third, economics. The robot only becomes truly interesting when it is cheaper, more flexible or easier to deploy than the alternative. Until then, it remains a demo with a valuation attached.
The United States may still shape the "brains" of the category. China may be better positioned to win the "bodies" race, taking components, manufacturing know-how and falling costs and turning them into commercial deployment. If so, the eventual winners may not be whoever builds the most charismatic robot. They may be whoever most efficiently closes the gap between prototype and product.
The robot IPO did not just show that investors love a story. It showed that humanoid robotics is starting to look investable enough for markets, governments and industry to organise around it. That does not mean adoption is imminent, it means the race is now real.
Macro
Public markets
M&A
Staying Diligent
Things we are watching this week: 24–30 August:
Unhedged Commentary
AI is making knowledge cheaper and verification more expensive.
AI is doing something extraordinary to knowledge work. It is making the production of plausible text, summaries, explanations and even research-looking output dramatically cheaper. The cost of producing something that sounds informed is collapsing. More people can draft faster, analyse more material and get to a first answer in seconds rather than hours. The productivity gain is real.
The problem is that the cost of verification is not falling at the same speed. This week, Nature reported that around 90% of biomedical papers published in December 2025 and indexed in PubMed showed signs of AI-assisted writing. Separately, a California appeals court sanctioned a lawyer after AI-generated fake citations ended up in a legal filing. These are very different settings, but they point to the same underlying issue. Knowledge is becoming abundant, trust is not.
That distinction matters because most professional work does not fail at the point of drafting. It fails at the point where someone assumes the draft is true. In a world where machines can instantly produce convincing prose, the scarce asset is no longer expression, it is provenance.
When text is cheap, bad text multiplies. When summarising becomes trivial, unsupported summarising becomes common. When everyone can generate something that looks polished, polish stops being a signal of quality.
In many high-trust settings, the answer itself is not the final product. The final product is a defensible answer. That requires citations, context, traceability and the ability to inspect what sits underneath the surface. In other words, the premium is moving from content to confidence.
Media businesses will need to show why their information is reliable, not just fast. Scientific publishing will need stronger norms around disclosure and attribution. Legal and financial professionals will need workflows built around checking, not merely generating. And companies building AI tools will increasingly be judged not only on what they can produce, but on how well they can ground, track and audit what they produce.
This is also where a quieter opportunity sits. In a world of infinite answers, tools that preserve evidence start to matter much more. Not because they slow work down, but because they make fast work usable. The more AI compresses the cost of saying something, the more valuable it becomes to know exactly where that something came from.
AI is making knowledge cheaper, but it is simultaneously making verification more valuable. The winners will not simply be the platforms that generate the most text. They will be the systems that make text trustworthy enough to use when the consequences matter.
The future of knowledge work may still be automated, but it will have to be auditable.
In Other News
Your meteorologist has teeth.
The latest tool in hurricane forecasting is not a satellite, a drone or a new buoy. It is a shark.
Scientists in the United States are attaching small sensors to sharks to collect real-time ocean data, particularly temperature and depth information from the upper ocean layer that helps fuel hurricanes. In effect, the sharks become moving data platforms, swimming through areas where better information can improve forecasting.
The story is unusual enough to be memorable, but the business lesson is surprisingly familiar. When organisations think about innovation, they often start by imagining new infrastructure. But sometimes the most useful asset already exists.
The sharks are already swimming. The challenge is not creating the network, it is finding a way to make the network useful. Many businesses overlook opportunities because they focus on building from scratch rather than reusing what is already there. A company may already have a customer base that can double as a distribution channel, a product that can generate valuable data, a service team that can become a sales engine or a piece of infrastructure that can support an adjacent business line.
The best innovations are often the ones that fit naturally into the real world rather than requiring the world to rearrange itself around them. A shark does not need to be persuaded to swim. A customer does not need to be taught a new habit if your product improves the one they already have. A team adopts new tools more readily when those tools sit inside existing workflows.
The lesson: before investing heavily in building a new network, ask whether one already exists, quietly moving through the world, waiting to become useful. Sometimes the smartest sensor is the one with fins.
The Thinking Corner
When one country leads in software and another in manufacturing, which advantage ultimately matters more in determining who captures the economic value from the next technology cycle?
Latest from the team
Market commentary and field notes from the team, on The Felix View.