Chung仲

China has the first approved implantable brain-computer interface (bci), the fastest trial timelines in the field, and more money going in than anywhere outside the US. RMB 5bn+ in the first half of 2026, more than 2024 and 2025 put together.
Most people read that as a regulatory story. Lighter ethics review, hospitals that say yes, a state that decided to move. That's part of it. It leaves out the more interesting half.
Look at what actually got built.
Most of the Chinese programmes that reached patients sit outside cortex. Electrodes above the dura or resting on it, not pushed into brain tissue. The device that won the world's first approval never touches the brain, and it runs on wireless power through the skull instead of a battery, so there's nothing to go back in and replace.
Everyone explains this as convenience. Less invasive, faster to clear, easier to staff. That's true. I've made the argument myself.
It's also a hedge on duration, which is the problem this whole field hasn't solved.
The brain treats an electrode as a foreign object. Glia migrate in, wrap it, and the signal degrades. That's the known failure mode for anything sitting in cortex, and it plays out over one to three years, well past the end of most trials.
Then look at what the data covers after that. Registration endpoints at three and six months. One Beijing programme reporting its longest implant at fifteen months. Neuralink's first patient a bit over two years in. The only human datasets spanning a decade come from rigid academic arrays built in a different era, and they don't transfer to flexible devices.
No flexible implant anywhere has five years of human data at high channel counts. Not one.
These devices are sold as restoration. Permanent. The datasets behind them are measured in months.
Staying out of cortex hedges that directly. You give up bandwidth to buy a device that might still be working in ten years.
So the Chinese field has quietly split into two bets. The epidural camp takes durability first and accepts a lower ceiling. The intracortical camp takes the signal now and expects materials to fix durability later, with stretchable electrodes that move with tissue instead of resisting it, aimed at the electrode retraction that showed up in Neuralink's first patient.
That split is what I'd be watching, rather than who has the most channels.
It's also why the channel count race misreads things. A thousand channels for eighteen months is not obviously better than a hundred for a decade, especially when the patient is thirty and the alternative is another craniotomy.
Some scale is worth holding onto here. Cumulative human implants worldwide, every company and every trial combined, are in the low hundreds. China's most advanced programmes count patients in the tens. The field is young enough that the failure mode hasn't had time to show up.
So here's where I land. China's lead is real, but it's a lead in a race where nobody has agreed on the finish line. Approvals and throughput get scored now. Durability gets scored around 2030, and that's what decides which of the two bets was right.
Bandwidth is an engineering problem with an obvious gradient, and everyone gets there eventually. Duration is biological. It answers slowly, and you can't iterate your way through it faster.
So when a company walks you through its numbers, ask what its longest human implant is and what the signal looked like at the end.

Chung Research Note #8 - robotics moat
The moat in robotics isn't the model. It's data collection.
Everyone's watching the hardware race and the benchmark race. But those might not be where the real story is.
Think about how GPT got trained. The text was already there. 20 years of the internet, trillions of tokens, sitting there long before anyone set out to train on it.
Robots get no such head start. You can't scrape "how to fill a glass of water without spilling it" off a webpage. Every motion a robot learns has to be produced in the physical world, one repetition at a time.
That's the real bottleneck this generation of embodied AI is stuck on, and it's the part everyone skips while they stare at the hardware and the models.
So the winner won't be whoever's robot looks most human, or whose model posts the best benchmark. It'll be whoever can collect real-world interaction data at scale AND at quality. The data doesn't exist yet, and that scarcity is structural.
Right now you get it through teleoperation, people driving robots through demos. But it's a trilemma of cost, scale, quality, and you get two at most. Clean demos are slow and expensive. Cheap and high-volume, and quality collapses.
"Just use simulation" only gets you so far. A model that's flawless in sim falls apart on real hardware. Sim amplifies real data. It doesn't replace it.
Put it together and you get a flywheel. More robots deployed, more real data, stronger models, more deployment. It's a first-mover's game, and the lead snowballs.
Tesla isn't the strongest hand because Optimus looks good. It's because 1M+ cars generate road data every day, and almost nobody can match that stream. Same story with Symbotic in Walmart's warehouses, Serve on real sidewalks, Intuitive with millions of surgical procedures.
The moat was never the hardware. It's the proprietary data stream nobody else can reach.
And this is a multi-year game. General-purpose usefulness is at least 3 years out, and that lag IS the moat. A flywheel that takes 3+ years to spin up means the leader is ahead by a window that's very hard to close.
So here's where I land. The data layer may be worth more than the model layer. Models commoditize the second open weights drop. Hardware keeps getting cheaper. But you can't collect data inside someone else's warehouse, roads, or operating room.
Don't bet on the flashiest robot or the best-benchmarking model. Bet on whoever holds the data collection flywheel.
One question tells you if a company is worth it. Is it generating real-world data nobody else can get, every day, with volume still climbing?
If yes, that's the moat.