Every major leap in computing arrived as a new layer in the stack. The operating system layer created Microsoft and Apple. The intelligence layer, built on foundation models, created OpenAI and Anthropic. A startup spun out of the Chinese Academy of Sciences is betting that the next layer is organization: connecting many intelligent agents so the whole system behaves smarter than any single model.
Dunxu Technology (沌序科技), founded in April 2026, calls this Collective AGI. It does not train a bigger model. It builds the layer between models and the physical world, drones, robots, radars and edge devices, and turns how these things cooperate into a reusable product.
The thesis: intelligence you organize, not just scale
For the past few years the industry has followed one clear path: make the single model stronger. More parameters, more data, more reasoning. But real-world systems rarely run on one model. A drone swarm mixes sensors, compute and tasks. A low-altitude safety system combines cameras, radar, drones, edge boxes and different algorithms. A smart model does not natively know how these devices should divide work, when to collaborate, or who covers for a failed node.
Dunxu founder Pu Zhiqiang, a 38-year-old researcher and deputy director of the National Key Laboratory of Complex Systems Cognitive and Decision-making at the CAS Institute of Automation, has studied this question for more than a decade. His starting point was not a model but an ant colony.
A single ant can do almost nothing. An ant colony, with no commander, builds nests and bridges no individual could manage, guided by one or two simple rules. Complexity science calls this emergence: 1+1 produces more than 2. The field was recognized with the 2021 Nobel Prize in Physics. Pu's bet is that intelligence does not only come from stronger individuals; it can come from how individuals are organized. The model produces individual intelligence; organization produces collective intelligence.
The evidence: from football to drones
The team has tested this in real systems. Football is a natural swarm-intelligence lab: no player sees the whole pitch, yet the team keeps making joint decisions. Using real UEFA Champions League data, the team built a virtual coach that reasons about what happens next, including counterfactuals like what changes if a player is swapped.
In a 2025 experiment, two self-built models of around 100 million parameters each, working together, outperformed a general-purpose model with over 100 billion parameters on a specific complex decision task. The point was not small beats big, but that properly organized modest agents can produce capabilities no single model has.
The real-world proving ground is low-altitude safety. The team has accumulated roughly 50TB of multimodal drone-detection data, covering visible light, near-infrared, thermal and radar, plus 15 million labeled frames and more than 2,000 experiments. That data matters beyond scale: algorithms that work in simulation often break when the weather, background or camera changes, or when communication lags, sensors drift and nodes fail.
In front of the devices sits what the company calls a collective brain, a middle layer that fuses sensor data, makes judgments and schedules the right executor. The key product idea is organization distillation: when several devices develop an effective division of labor over many runs, the system distills it into reusable rules and eventually formulas. The next deployment, with different devices or a different mission, no longer starts from zero.
Where the organization layer sits
The clearest frame is computing history. The OS layer made Microsoft and Apple. The intelligence layer made OpenAI and Anthropic. Dunxu is building the next one.
Neighboring players are converging on the same problem from two directions. Software companies are shipping multi-agent orchestration from the model side. Physical-autonomy players like Shield AI, Anduril and Applied Intuition, with its Agentic Platform for Physical AI, are abstracting upward from drones, robots and self-driving systems. Dunxu comes from a third direction, swarm intelligence, and it is complementary rather than competitive with model makers. Organization needs models: stronger general models handle complex reasoning while small models handle cheap, low-latency tasks. The stronger the nodes, the more valuable the organization problem becomes.
What changes in the AI industry
If the organization layer becomes a real product category, the industry's cost structure shifts. Strong AI no longer requires the strongest model, a point open-model developers are already proving with a 27B open model beating larger closed ones. The bottleneck moves from training to coordination: data from real deployments, engineering tolerance for failure, and reusable organization rules.
Dunxu's path also shows a specific moat: a decade of research, tens of terabytes of physical-world data, software-hardware integration, and a revolving door with its CAS lab. It was pulled into the market, after one trade show appearance more than 30 organizations expressed intent, and demand in the tens of millions of yuan piled up before the company formally existed.
What to watch next
Three things will tell you whether the organization layer is real. First, reuse: can cooperation rules built for one deployment survive a change of devices and missions without retraining? Second, standards: whether multi-agent orchestration converges on shared protocols, much as stateless gateways are consolidating MCP for agent-to-tool traffic. Third, cost: whether a fleet of cheap agents organized well genuinely beats one expensive model on real workloads, not just in a 2025 lab experiment.
The scaling era is not over. But the next layer of value may not be a bigger brain. It may be the rules that make many small ones think like one big one.
