Reuters: Alibaba plans a 5T to 10T parameter AI model and unveils a new chip

Reuters reports that Alibaba plans an AI model with between 5 trillion and 10 trillion parameters, and that the company has unveiled a new chip. The story ran under the headline “Alibaba plans AI model with 5 trillion to 10 trillion parameters, unveils new chip” and was carried by outlets including Yahoo Finance Singapore, with both syndicated versions describing the same development.

The parameter range is the headline number. A model at that scale would sit well beyond the size of most systems in production today, but Reuters frames it as a plan rather than a finished product.

The chip is the second half of the story, and the verb matters: the model is something Alibaba “plans,” while the chip has been unveiled. Read together, the two announcements describe not only model capability but the compute base that capability would run on.

Details remain thin. The report does not name the chip, its process node, its performance figures or a production timeline, and it does not say when the model would ship, whether it would be open-sourced, or whether it is intended mainly for training or inference. What can be stated with confidence is narrow: a parameter range for a planned model, and a chip that has now been shown.

Why it matters: pairing an ultra-large model plan with a self-designed chip places the announcement in a broader shift away from single-point model benchmarks toward a model-plus-silicon narrative. Parameter counts of this size imply enormous training and inference demand, and an in-house chip is one way to make that demand more predictable. Model scale keeps pushing into the trillion-parameter range while in-house silicon lands alongside it, which suggests China’s leading vendors are betting on both frontier capability and compute self-reliance at once.

What to watch next is official disclosure: whether Alibaba names the model, gives a release schedule or an open-source plan, whether the chip’s specifications and deployment path are published, and whether either track the company’s cloud roadmap. Until then, the parameter figure should be read as a target rather than a delivered capability.