TL;DR

Anthropic is discussing an acquisition of Decart, an Nvidia-backed startup working on AI infrastructure and optimisation, according to a person familiar with the situation. Bloomberg, which broke the story, put the possible price near $6bn. Talks are early, neither company would comment, and the stated motivation is capacity rather than capability.

Buying throughput, not models

The logic on offer is unusually plain. Decart builds tooling that squeezes more out of existing hardware, and its team would reportedly land inside the group at Anthropic responsible for inference and performance — the part of the business that determines how much customer demand can actually be served, and at what cost per request. Decart also ships models of its own: Lucy, which edits video as it plays, and Oasis, which generates simulated worlds for training robotics and self-driving systems. Neither looks like the reason for the conversation.

That reading fits what Anthropic said a week ago, when it began recruiting engineers who can work across hardware and software to help design custom silicon alongside the models meant to run on it. A company that felt comfortable about its serving capacity would not be doing both.

Looking forward

The context is a listing. Anthropic is preparing to float, its president Brad Lightcap left this week, and the constraint it keeps describing is not research talent but compute it can bring online. Paying several billion for an optimisation team is a wager that efficiency gains arrive sooner than new data centres — a reasonable bet when capacity is the thing rationing revenue.

For UK customers the relevance is prosaic and immediate. Enterprise buyers here have spent the past year negotiating around rate limits and unpredictable throughput on frontier models, and this is the clearest signal yet that the labs regard those limits as a structural problem rather than a queue that clears. Radical Ventures led a $300m round into Decart back in May, Nvidia coming in as a new backer; a $6bn exit three months later says something about what serving capacity is currently worth.