Analysts following China's national AI infrastructure plan identified as an open question which interconnect layer the country would adopt to link data centres at national scale.
The question is technical and decisive. Filling a data centre with accelerators is a purchasing and power problem; connecting distant data centres to work as one is a latency and bandwidth problem, and no quantity of accelerators compensates for a slow connection.
Other fronts in the semester reached that conclusion independently, with foundries racing toward co-packaged optics, a network manufacturer doubling switching capacity in one generation, and a government directing public funds to technology for moving data between processors.
The choice has long-term consequences. Interconnect is a control point as relevant as chip fabrication, and whoever sets the standard for that layer sets the pace for everyone depending on it.
For operations that only rent capacity, none of this changes this month's price. It changes the shape of what will be available in a few years and at what cost per task.