A Company from Nanjing University's Orbit Bets on a Fourth Driver of AI Growth

For the past several decades, AI capabilities have grown along three axes: compute, algorithms and data. The Nanjing team covered by the Chinese outlet Jiqizhixin proposes adding a fourth: connectivity among the models themselves. The announcement frames the bet this way: strong models are already abundant and each has been studied well, so they can be linked into a network and made to work together. It is a logical move for a market where scaling single models is getting more expensive faster than results are improving. But the available excerpt of the article contains no product name, no benchmarks and no figures, and that is important to keep in mind.

Three axes and what they cost

Large language models grew on a simple formula. Clusters got bigger, parameter counts went up, and text corpora swelled to hundreds of billions and trillions of tokens. The result was a set of "super-unit" models, each strong on its own and trained on its own infrastructure.

This trajectory has two limits. The first is economic: every next step in compute and energy costs noticeably more than the last. The second is informational: high-quality open data is getting scarcer, and synthetic data does not always replace it. Against this backdrop, the idea of "assembling" capabilities from models that are already trained looks like a rational alternative to further scaling of individual models.

The fourth element: what exactly is being proposed

The introductory excerpt points to one main thesis: models whose capabilities are sufficiently well known can be connected, and their interaction can be organized. The author of the article phrases this as a question about what will emerge once there are enough strong models. The word "emerge" hints at an expected emergent effect: the network's output should exceed the sum of its parts.

A single introductory paragraph does not reveal how the team plans to pull this off. Here is what is known and what remains unknown:

• The original three axes: compute, algorithms, data.
• The proposed fourth axis: connectivity and cooperation among models that are "well understood" by one another.
• The organization behind it: a company from the Nanjing University ecosystem ("南大系" in the original).
• Not disclosed: the product name, the interaction architecture, protocols, benchmarks and the list of models that can be connected.

Similar approaches already exist. Multi-agent frameworks, request routing between models and ensembles of specialized LLMs have been used in practice for some time. Open protocols such as MCP and A2A have emerged for communication between agents. If the Nanjing pitch is novel, the novelty must lie in how exactly the models "learn" each other's capabilities and divide up tasks. The source does not show that yet.

What to check when details emerge

The bet on connectivity will be convincing if the team demonstrates three things. First: a gain on open tasks compared with the best single model in the same price category. Second: the cost of coordination, meaning how many extra tokens, how much latency and how many errors each "hop" between models adds. Third: how robust the system is when one of the connected models changes or degrades after an update.

For developers and teams building agent systems, it is worth measuring two things on your own tasks right now: where a single model hits its quality ceiling, and how much you are willing to pay for routing across several. If the gap is large, the idea of a network of models is worth an experiment. If not, it is cheaper to improve prompts, data and tooling. Keep an eye on the original Jiqizhixin publication and on whether the company discloses the product name and test results. For now, this is a concept pitch, not a confirmed technical result.