AI scale is an infrastructure decision

Alibaba said on 22 September that it was developing an artificial-intelligence model with five trillion to ten trillion parameters and unveiled its Zhenwu V900 processor. Reuters reported that the company expects the chip to enter mass production in early 2027 and has set a target for Alibaba Cloud's global data-centre capacity to exceed 20 gigawatts by 2032.

Those announcements connect three layers that are often discussed separately: models, semiconductors and data centres. A larger model may require more computing, but commercial performance also depends on memory, networking, cooling, software efficiency and how heavily the equipment is used. The infrastructure plan therefore matters as much as the headline model size.

Dunes Bureau interpretation: providers should express expansion as a chain of assumptions. How much usable computing will each site deliver? What power and cooling capacity is secured? What utilisation is required for acceptable returns? Which workloads will pay for training capacity when it is not fully occupied? A weak answer at any link can turn technical ambition into stranded capital.

Customers face a different decision. They should test whether a provider's scale improves the cost, speed, reliability or data controls of their own workload. Contract reviews should include unit pricing, capacity commitments, service levels, data movement costs and a realistic path to another model or provider if economics change.

The announced targets are forward-looking and do not establish future market share, delivery dates or financial returns. Parameter count is also an incomplete measure of capability. The useful signal is broader: leading AI companies are committing across the full infrastructure stack, which raises both the cost of competition and the importance of disciplined capacity planning.