Amazon targets $7B Decart AI deal to boost multi chip compute

If Amazon closes a $7 billion deal for Decart, it would be a very direct play on one of the hardest bottlenecks in the AGI race: efficiently mapping large models onto whatever accelerators you can actually get. Decart’s DOS platform promises to shrink the months long process of retuning models for new chips down to weeks, and that fits neatly with AWS’s strategy of pushing its Trainium3 silicon alongside Nvidia hardware rather than staying dependent on one vendor.

Strategically, this is Amazon trying to turn chip heterogeneity from a liability into an advantage. If DOS works as advertised at scale, AWS can tell customers that they do not have to care as much whether they are on Trainium, Blackwell or some future accelerator; the platform can flex workloads to whatever is available or cost efficient. That is a powerful answer to Microsoft’s deep Nvidia pipeline and Google’s TPU centric stack. For Decart, being embedded inside AWS would immediately broaden distribution for its world models and agent tooling.

In the broader race to AGI, this is less about inventing a radically new model and more about industrializing the infrastructure that existing and future models need. But that kind of plumbing is exactly what determines which labs can keep scaling context, tokens and agents without hitting a wall on cost or supply.