Indicator CapitalChina Deep Tech Expedition 2026
Indicator CapitalChina Deep Tech Expedition 2026
The Dragon and the Data-Hungry Monster
The Dragon and the Data-Hungry Monster
Six days across the full Smart & Connected stack — sensing, connectivity, compute, reasoning, and actuation. Fourteen companies in Shenzhen, Dongguan, and Guangzhou.
Expedition I — inside the room, not outside the building
Morgan Stanley (Jun 2026) — IFR preliminary 2025 data — 15th Five-Year Plan recommendations
The thesis
Intelligence is moving out of software and into machines.
For four years we have described artificial intelligence as a data-hungry monster. Every model better than the last is better because it ate more, and the supply of text on the internet is finite in a way the physical world is not.
So we drew a flywheel with three beats. What we found on the ground in China was a fourth. The most valuable data is not observed, it is provoked — a machine attempting something, failing by two millimetres, and recording what the world did about it.
A sensor tells you what happened. An actuator tells you what happens the next time you try.
And it compounds. When one machine works out how to grip a wet component, the correction is validated and pushed to every other machine on the fleet before the next shift starts.
One robot learns. All robots know.
It moves the interesting question away from who has the best model, and toward who owns the machines that generate the data the model cannot get anywhere else. In Brazil those machines are already in the ground, in the field, and on the road.
For four years we have described artificial intelligence to our LPs as a data-hungry monster. The description holds. Every model that is better than the last one is better because it ate more, and the supply of text on the internet is finite in a way the physical world is not.
So we drew a flywheel. AI improves, which creates demand for more data, which puts more sensors and connected machines into the world, which improves AI again. Three beats. We drew it on whiteboards in São Paulo for two years and it was, we now think, incomplete.
What we found on the ground in China was a fourth beat. The most valuable data is not observed, it is provoked. It comes from a machine attempting something, failing by two millimetres, and recording what the world did about it. That is not sensing. That is actuation, and it is the layer where the loop finally closes.
A sensor tells you what happened. An actuator tells you what happens the next time you try.
Then the second thing, which is why this compounds instead of merely accumulating. A fleet does not learn the way a person learns. When one machine works out how to grip a wet component, the correction is uploaded, validated, and pushed to every other machine on the fleet before the next shift starts.
One robot learns. All robots know.
This is the upgrade to our own framework, and it changes where we look for companies. It moves the interesting question away from who has the best model and toward who owns the machines that generate the data the model cannot get anywhere else. In Brazil, those machines are already in the ground, in the field, and on the road. They are simply not yet instrumented, and almost nobody is being paid for what they know.
The flywheel
The week
Five program days. Fourteen companies. One thesis.
The week traverses the entire chain: who makes the machine see, who moves the data, who makes it think, who makes it move, and who is already operating it at commercial scale.
Proposed agenda, subject to confirmation.
Why this lineup
Not a list of famous companies. The entire chain, in order.
This is the framework we use to classify every company we invest in. The week runs the length of it.
We were in the room before the inflection
On our first expedition we sat with Unitree when robot dogs were the business and humanoids were research and entertainment. Our host was blunt: the robots were still following a path a human had drawn for them.
Since that conversation, Unitree's STAR Market IPO application was accepted, and humanoid revenue overtook quadruped revenue for the first time — crossing 51% of the total.
The second leg overtook the first. We watched it from inside the room.
Expedition I — Lenovo Tianjin, Tuya Smart, Inspur, Alibaba Cloud, Unitree, MEGVII, Jieshun, TRS
Why Brazil
Brazil is not going to win the race to build the frontier model, and it does not need to. What Brazil has is the thing the frontier model cannot buy: physical operations at a scale almost nowhere else can match, running in conditions no dataset currently describes.
Four sectors carry this directly. Agribusiness, where the machinery already drives itself down a row it has never seen. Energy and natural resources, where the assets are enormous and remote. Advanced manufacturing and logistics, where margin is measured in seconds of cycle time. And smart cities and infrastructure, where a municipality operates a fleet it can barely see.
Every one of those is a machine acting in a place the model has never been. That is not a market for imported software. It is a position.
Not the country that builds the robot. The country where the robot learns what the world is actually like.
Brazil is not going to win the race to build the frontier model, and it does not need to. What Brazil has is the thing the frontier model cannot buy: physical operations at a scale almost nowhere else can match, running in conditions that no dataset currently describes.
Four sectors carry this directly. Agribusiness, where the machinery already drives itself down a row it has never seen before. Energy and natural resources, where the assets are enormous, remote, and inspected by people who have to travel to reach them. Advanced manufacturing and logistics, where margin is measured in seconds of cycle time. And smart cities and infrastructure, where a municipality operates a fleet it can barely see.
Every one of those is a machine acting in a place the model has never been. That is not a market for imported software. It is a position — and the companies that hold it will be the ones that instrumented their own operations first, which is a decision somebody makes in a boardroom this year, not a technology anyone waits for.
Not the country that builds the robot. The country where the robot learns what the world is actually like.
Who's leading it
Five people, on the bus with you all week.
Smart and connected thesis. Around 30 portfolio companies. LPs include BNDES, Qualcomm Ventures, and Lenovo.
FCGI has organized China roadshows for more than 600 Brazilian investors and entrepreneurs.
Who it's for
Logistics
Request an invitation
Applications are reviewed by the Indicator team.
Apply for a seatApplications close September 30, 2026, or when all 15 seats are filledRequest an invitation
Applications are reviewed by the Indicator team. Seats are limited to 15, and we prioritize participants with an operating or investment mandate in the sectors on the agenda.