Amazon in talks to buy Decart for about $7 billion, maker of software that optimizes AI models across NVIDIA, Google and Amazon chips

Both ratios here are ceilings, because the source gives only floors for headcount and funding. Decart employs more than 60 people and has raised over $450 million ^[4], so Bloomberg's reported $7 billion ^[10] works out to under $117 million a head ^[9] and under 15.6 times the capital raised ^[8]. The reports do not include revenue.

We think most of that price is for DOS, the software layer that runs models on Nvidia GPUs, Google TPUs and Amazon's Trainium ^[3]. A buyer with its own chip among those three has an obvious use for software that makes the other two easier to swap in. In our view the purchase adds no chips; it raises what each existing chip can run, the only part of the industry's chip scarcity ^[15] that software reaches. The counter-thesis is that the models are the asset: interactive world models like Oasis are drawing growing attention from AI labs exploring simulation and gaming ^[7], and a buyer could be paying for that.

Anthropic reportedly weighed $6 billion before leaving the process ^[12], so the reported Amazon price is $1 billion higher, about 17% ^[14]. The reports do not say why Anthropic withdrew. Our hypothesis is that a chip owner values a portability layer differently from a model lab, since a workload that moves to Trainium is business for Amazon.

The deal can close with DOS kept on all three chips, close with support drifting toward Trainium (the tilt developers will be watching for ^[17]), or fail in diligence as Anthropic's reportedly did ^[12]. We would call the thesis wrong if the announcement leads with Lucy and Oasis ^[6] and does not mention DOS on Nvidia and Google hardware.