# Anthropic’s $6B Decart Deal Is Not About Video AI

> Source: <https://byteiota.com/anthropic-decart-acquisition-claude-api/>
> Published: 2026-08-14 23:09:54+00:00

Anthropic is in early talks to acquire Israeli startup Decart AI for around $6 billion — its largest acquisition ever. Most headlines are calling it a video AI deal because Decart makes some genuinely impressive real-time video generation products. That framing is wrong. This acquisition is about Anthropic paying xAI [$1.25 billion every month for compute](https://techcrunch.com/2026/05/20/anthropic-will-pay-xai-1-25-billion-per-month-for-compute/) and needing a way out before its October IPO.

## The Actual Problem: A $15 Billion Annual Compute Bill

Anthropic currently runs its Claude models on Colossus 1 — xAI’s Memphis supercomputer with 220,000+ NVIDIA GPUs and 300 megawatts of power. The contract runs through May 2029 and costs $1.25 billion per month. Against Anthropic’s $30B+ annualized revenue, that single line item is consuming close to half the company’s gross income. No frontier AI lab can sustain that ratio heading into a public offering.

Decart’s core product is DOS — the Decart Optimization Stack. It operates at the kernel and compiler level of AI hardware rather than at the application layer. Instead of accepting NVIDIA’s defaults, DOS co-designs model execution with the chip’s actual microarchitecture, hand-optimizes critical kernels, and generates chip-specific instructions. The claimed results: 1,600+ tokens per second for agentic inference versus an industry average of roughly 200, and a 100x improvement in cost efficiency. DOS also [supports NVIDIA, Google TPU, and Amazon Trainium](https://decart.ai/publications/decart-raises-300m-tech-leaders-back-the-company-as-both-customers-and-investors) — meaning it works across whatever hardware Anthropic runs on next.

If DOS delivers even a third of those claimed gains on Colossus 1, the $6 billion price tag pays for itself inside two years. The math is not complicated.

## The IPO Angle Everyone Is Missing

Anthropic [confidentially filed its S-1](https://fortune.com/2026/08/13/anthropic-said-in-talks-to-buy-startup-decart-for-6-billion/) in June and is targeting a Nasdaq listing in October at a reported $2 trillion valuation — which would make it the largest IPO in history. Goldman Sachs, JPMorgan, and Morgan Stanley are underwriting. Going public with a compute bill that absorbs half your revenue is a bad look for investors, and everyone in the IPO process knows it.

Acquiring Decart before the listing does two things. First, it signals that Anthropic is solving the problem rather than just disclosing it. Second, “we own our inference stack” is a meaningfully different investor story than “we rent compute from a competitor at $1.25 billion a month.” Google built TPUs. Amazon built Trainium. Meta has its own inference infrastructure. Anthropic was the last frontier lab still exposed at this level — as [we covered when the Colossus deal landed](https://byteiota.com/xai-rents-colossus-to-anthropic-and-google-frontier-lab-or-reit/). This is the move to close that gap.

## What About the Video AI?

Decart also builds Lucy — a real-time video model that transforms live video feeds with AI-generated changes (virtual try-on, scene restyling, real-time VFX) — and Oasis, a world model for autonomous vehicle and robotics simulation. The $300M Series B in May brought in NVIDIA, Toyota Ventures, eBay Ventures, and Andrej Karpathy as investors and customers. These are not throwaway products.

But they are not why Anthropic is paying $6 billion. If the deal closes, Decart’s team joins Anthropic’s inference and performance organization — not a video product team. The world model capabilities are upside, not the thesis.

## What Claude Developers Should Do Right Now

The honest answer is: nothing yet. The deal is early-stage and explicitly could still fall through. Both companies declined to comment. No API changes have been announced. Pricing has not moved.

What to actually watch:

- The deal close announcement, if it comes. Integration into Claude’s inference infrastructure takes 6–12 months minimum after close.
- Anthropic’s IPO prospectus. When the S-1 goes public in September or October, it will disclose Anthropic’s actual cost structure for the first time — including whether DOS is already reducing infrastructure spend.
- Claude API changelog entries referencing inference improvements or latency reductions. That is where you will feel the impact before any pricing announcement.

The acquisition has not closed. But the logic is sound. Anthropic has the most capable models on the market and the worst infrastructure cost structure among the major labs. Decart is a credible solution to the second problem. If this goes through, Claude users eventually benefit — not immediately, but the direction is toward cheaper, faster inference rather than more expensive and slower. That is worth paying attention to.

*Disclosure: Anthropic and Decart both declined to comment. The deal remains in early-stage discussions and has not closed as of August 14, 2026.*
