Anthropic's reported $6B Decart talks put a price on real-time video dependency

Anthropic is in talks to buy Decart, a live AI lab, for about $6 billion, Bloomberg reported on August 12, citing people familiar with the matter ^[1]. Nothing has been signed, the figure is still a negotiating number, and the talks could collapse ^[2]; even so, any team that has wired a real-time video or world-model API into a shipping product now has to price the possibility that its vendor becomes a component of a foundation-model company.

What is on the table is not a demo. As of August 13, 2026, Decart describes three things: DOS, an inference-and-training stack for major hardware platforms; Lucy, a real-time video model; and Oasis 3, a real-time, controllable multi-view simulation environment for physical AI ^[6]. Those models are exposed to outside developers through an API and a software development kit ^[7]. Decart says DOS can operate across Nvidia GPUs, Amazon Trainium and Google TPUs ^[8]. Radical Ventures, which led the company's May financing, describes DOS as combining hardware-aware model design, kernel tooling, proprietary compilers and inference optimization ^[9], and says Lucy responds to user input in under 30 milliseconds while naming Amazon as an early anchor API customer across multiple product lines; both of those assertions come from the lead investor and have not been independently audited ^[10].

The multi-silicon portability is the part worth tracking, because it is the part with the least obvious reason to persist. Claude has been built primarily around language, coding, agents and computer use ^[11], and the reported logic of the deal is a route into world models, live media and physical AI without Anthropic building the stack internally ^[12]. Running anywhere is a selling point for an independent infrastructure vendor. For an owner that sells its own models, keeping three silicon backends and an outside anchor customer at first-class priority becomes a business decision rather than a technical one. Nothing in the reporting says a roadmap will change ^[2]; the point is that the option now exists and should be in your contract and fallback planning.

The price indicates what the buyer thinks about timing. Calcalist reported on May 18, 2026 that Decart's Series B valued the company at approximately $4 billion ^[4], which makes about $6 billion roughly a 50 percent premium some three months later ^[5]. A $100 million round in August 2025 was reported at $3.1 billion ^[16], so the marked path is roughly 29 percent over nine months, then about 50 percent over three ^[20][5]. Against more than $450 million raised in total ^[17], $6 billion is on the order of 13 times invested capital ^[21], for a company founded in 2023 ^[3], roughly three years earlier ^[23].

By its own documentation the category is early. The first Oasis release generated output at 20 frames per second, about 50 milliseconds per frame ^[13][22], with fuzzy distant imagery, temporal inconsistencies involving uncertain objects, domain-generalization problems, limited long-horizon memory and difficulty controlling inventories and objects ^[13].

Signals worth watching: a signed agreement or a public collapse ^[2]; whether Trainium and TPU support in DOS keeps shipping after a close ^[8]; whether the Amazon API relationship described by Radical continues ^[10]; and how much of the consideration sits in retention rather than the headline price, which the reporting flags as a live variable ^[19].