Anthropic is in talks to acquire Decart, an AI inference startup, for $6 billion, according to Bloomberg reporting on Wednesday. The deal would give Anthropic access to technology that runs language models at roughly 8 times the industry-average speed.
Decart's inference system processes tokens at 1,600 per second, compared to typical production systems operating at 200 tokens per second or slower. Faster inference directly reduces the compute cost per generated output, allowing Anthropic to serve more requests across its existing hardware fleet, a material advantage for a company operating at the scale Anthropic now commands after securing $5 billion from Amazon and $2 billion from Google in recent months.
The startup has raised at least $300 million to date, according to previous announcements. Its technology is built on a world-models architecture designed to optimize real-time inference performance. In the current AI market, inference efficiency has become a key cost driver: as more users call foundation models at scale, the economics of serving those requests determine profitability. Faster inference per unit of compute means lower unit costs or higher margins on the same revenue.
Anthropic has competed directly with OpenAI and other frontier labs on model capabilities, but capital efficiency in deployment has become a separate and urgent problem across the industry. OpenAI has begun rolling out its own inference optimization work. Other labs have also acquired or partnered with inference-focused companies to reduce serving costs.
The $6 billion valuation would rank Decart among the most expensive AI infrastructure acquisitions to date. If the deal closes, Anthropic would integrate Decart's team and systems into its own operations rather than operating the startup as a separate entity.
Anthropic has not commented on the talks. Bloomberg reported the discussions are ongoing and deal terms could still change. The timing of a close and the exact structure of any acquisition remain unclear.
