Anthropic’s $6 billion Decart talks put AI compute efficiency in focus

70

TL;DR:

Anthropic is reportedly in early talks to acquire Decart AI for about $6 billion, a move that would add software designed to squeeze more performance from AI chips as compute costs and capacity become strategic constraints.

Article:

Anthropic is in early talks to acquire Nvidia-backed Decart AI in a deal that could be worth about $6 billion, according to Reuters, citing a source familiar with the matter. If completed, the acquisition would give the Claude maker software and engineering talent focused on making AI chips work harder, a growing priority as frontier model developers spend heavily on data centres and confront capacity constraints.

The timing matters because Anthropic is expanding compute on multiple fronts. In April, it announced an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity from 2027; in July, AMD said Anthropic would deploy up to 2 gigawatts of Instinct MI450-series systems. Buying optimisation technology could complement that hardware spending by raising utilisation and lowering the cost of training and inference.

Decart’s core asset is the Decart Optimisation Stack, software designed to run AI workloads across different processor architectures. Reuters said Decart could help Anthropic absorb more demand, while the startup’s team would join Anthropic’s inference and performance organization if a deal closes. Decart raised $300 million in May in a round led by Radical Ventures, with Nvidia joining as an investor.

“Moving models between different hardwares is incredibly hard,” Decart co-founder and CEO Dean Leitersdorf told The Wall Street Journal in June, describing a problem that can make switching chips costly and slow. That complexity matters as AI labs increasingly spread workloads across Nvidia GPUs, Google TPUs, Amazon Trainium and AMD accelerators.

The talks remain preliminary and could still fail. But the competitive edge is shifting toward extracting more useful compute from every installed system. Better chip utilisation can improve AI inference efficiency, stretch existing infrastructure and delay the need for another expensive data-centre expansion.