Amazon's $7B Decart talks: the chip portability layer gets a landlord

Amazon is in talks to buy Decart for about $7 billion, per the Wall Street Journal. Two months ago Anthropic was the suitor at $6 billion and walked away. The asset is the same, and it is not a chatbot. Decart sells software that makes AI models run faster on whatever chip you have: Nvidia GPUs, Google TPUs, or Amazon's own Trainium. When the cloud that makes one of those chips buys the layer that makes them interchangeable, ask who that layer works for after the deal closes.

What actually happened

Reuters relayed the WSJ report on October 11. The firm details:

• ~$7 billion, citing people familiar with the matter. Nothing is signed. Amazon did not comment outside business hours, and Decart did not respond to requests for comment.
• Decart was valued at nearly $4 billion in May, per PitchBook, after a $300 million round led by Radical Ventures with Nvidia as a new investor.
• Anthropic explored a ~$6 billion deal in August and dropped it last month after due diligence, according to Bloomberg. We covered those talks when they broke.
• Decart builds real-time world models (Lucy for live video, Oasis for robotics and driving simulation) and the Decart Optimization Stack, which it describes as hardware-agnostic across GPUs, TPUs, Trainium, and AMD chips.

The Amazon tie already exists. On AWS's Trainium customer page, Decart's CEO cites early tests showing up to 4x frame throughput and 2x cost efficiency versus top GPUs and says the capabilities are coming to Bedrock. Those are vendor numbers from early testing, not an audit.

Why it matters for your business

You probably do not buy chips. You do buy tokens, and the price of a token tracks how well your provider's hardware is used. Optimization layers like Decart's are how a cloud squeezes more out of the silicon it already paid for. Owning one is a cost play first.

The second-order effect is the one to plan for. A neutral optimizer helps every chip compete. An optimizer owned by a chip maker has a reason to make its own chip look best. Maybe Amazon keeps DOS open across hardware. Maybe the best tuning lands on Trainium first. Either way, the decision moves from Decart's roadmap to Amazon's.

For a small business, the practical rule has not changed. Do not let one vendor own both your model access and your exit. Keep model calls behind your own thin interface. Store prompts, evals, and outputs in your own database. Run your quality tests against a second provider every quarter, so switching is a config change you have already rehearsed, not a migration project.

Consolidation at the infrastructure layer does not hurt you today. It shrinks the number of independent escape hatches tomorrow.

Sources: Reuters via KFGO, Decart, AWS Trainium customers.