Amazon in talks to acquire AI startup Decart for $7 billion

Decart's price has climbed sharply in a short window. A $300 million round in May 2026 valued the Israeli startup at $4 billion, with Nvidia among the investors ^[2]and reportedly also a prior suitor whose own acquisition talks collapsed ^[4]. Anthropic then ran due diligence on an acquisition at roughly $6 billion before walking away in September 2026 ^[1], and Amazon is now reportedly discussing a deal worth about $7 billion ^[1]. That is the fourth valuation milestone in little more than a year, and reporting suggests the talks could still collapse, just as the earlier Nvidia and Anthropic conversations reportedly did ^[4].

Before any acquisition talk, Decart was already woven into AWS's business. AWS gave Decart early access to its unreleased Trainium3 chip in 2025 after engineers met the founders and came away impressed, and Decart is listed on AWS's own Trainium customer page alongside Anthropic, Databricks, OpenAI and Uber ^[1]. Market strategist Shay Boloor of Futurum Equities framed the move as Amazon 'effectively buying software expertise' rather than a product line ^[5]- Decart's Decart Optimization Stack (DOS) compresses chip-optimization work 'from months into a couple of weeks' across Nvidia, AWS Trainium and Google TPU hardware ^[1], which is exactly the kind of tooling that could raise utilization on Amazon's own silicon.

The deal echoes Amazon's 2015 acquisition of Israeli chipmaker Annapurna Labs for roughly $350 million, which became the foundation of AWS's in-house silicon line - Graviton, Trainium and Inferentia ^[3]. Decart is also an Israeli engineering team, founded in 2023 by Dean Leitersdorf, Orian Leitersdorf and Moshe Shalev ^[4]. But the comparison only goes so far: Annapurna sold Amazon chip-design talent, while Decart would sell software that makes chips from any vendor run faster - a different, and arguably more defensible, kind of leverage over Nvidia than owning another silicon design team.

The optimistic read is that chip-agnostic optimization software erodes Nvidia's CUDA lock-in by making AWS Trainium and Google TPU hardware competitive on performance. The more skeptical read, voiced in online discussion of the reported deal, is that Trainium3 is still fabricated by TSMC just like Nvidia's chips, which limits how much a software layer alone can reshape the underlying hardware economics, and that a bidding sequence running from Nvidia to Anthropic to Amazon looks as much like a negotiating tactic to inflate Decart's price as an inevitable strategic shift. What is harder to dispute is Decart's own technical claim: at AWS's re:Invent conference, the company said its Lucy model running natively on Trainium3 reached four times the frames-per-second of comparable GPU setups at 80% tensor-core utilization.