On October 11, 2026, The Wall Street Journal reported that Amazon is in negotiations to acquire AI startup Decart at a valuation of about $7 billion. Sources said that unless talks break down or a new bidder emerges, an agreement could be reached soon. This is the most representative M&A deal in the AI industry this year—not because it is the largest by amount, but because it reveals a previously overlooked dimension of competition: whoever controls the layer of software between chips and models will hold the next source of leverage.
Five Months, Valuation Doubles from $4 Billion to $7 Billion
Decart was co-founded in 2023 by three Israeli engineers—brothers Dean Leitersdorf and Orian Leitersdorf, and Moshe Shalev—and is headquartered in San Francisco. According to PitchBook data, the company has raised more than $450 million cumulatively from well-known venture firms including Sequoia Capital, Benchmark, and Radical Ventures. In its latest funding round in May of this year, it raised $300 million at a valuation of about $4 billion.
From $4 billion to the $7 billion in these talks, the time span is less than five months. Driving this jump was a dense sequence of competing bids.
According to Bloomberg, Nvidia was the earliest to be interested in acquiring Decart. Anthropic then became the preferred bidder around August this year at about $6 billion. Bloomberg reported on September 8 that Anthropic withdrew from the deal after completing due diligence. Amazon then entered with a higher offer and pushed the valuation to $7 billion. Three rounds of bidding, each higher than the last—that itself is a signal.
DOS: Turning Chip Switching from "Months" into "Weeks"
Decart's core product is the Decart Optimization Stack, or DOS—a software adaptation layer that allows AI models to run efficiently on different hardware such as Nvidia GPUs, Google TPUs, and Amazon Trainium.
The real-world problem is this: the vast majority of AI models are optimized by default for Nvidia GPUs during development, and switching chips means a great deal of re-debugging work, often lasting months. DOS claims it can compress this process to weeks. According to industry estimates, its acceleration effect can reach up to 8x in heterogeneous-chip scenarios. Amazon's AWS has already listed Decart as a customer and in 2025 gave it early testing access to Trainium3 chips.
Beyond DOS, Decart has also developed two application-layer products: Lucy handles real-time video editing, while Oasis provides virtual simulation environments for robotics and autonomous driving systems. But for Amazon, the appeal is almost entirely concentrated in DOS.
Why Anthropic Withdrew: The Logical Cracks After Due Diligence
Anthropic's withdrawal is the most information-rich detail in the entire M&A narrative, and also the least deeply explored.
On the surface, Anthropic is a model company, and acquiring a chip-adaptation software layer has a plausible strategic intent—if Decart's DOS can let the Claude series of models run more efficiently on more hardware, it would directly lower inference costs. But the problem is that Anthropic is also on the eve of an IPO. According to multiple media reports, Anthropic is planning to launch an IPO roadshow in mid-October, with a target valuation as high as $2 trillion. At this juncture, spending $6 billion in cash on a large acquisition makes the financial logic extremely complex.
The deeper contradiction is this: Decart's commercial value is built on "chip neutrality"—it serves customers of Nvidia, Google, and Amazon simultaneously, and it is precisely this neutrality that makes it infrastructure. Once Anthropic brought it in-house, Decart's appeal to customers of other chipmakers would immediately decline, and the "chip neutrality" moat would collapse. Judging from the due diligence results, this strategic contradiction may well have been the core reason Anthropic ultimately withdrew.
Amazon's Calculation: Making Trainium a "Low-Risk Option"
Amazon faces a very different situation from Anthropic. AWS has accumulated more than $225 billion in customer committed revenue on Trainium3, but Trainium remains a niche option to date. The fundamental reason is migration cost—for enterprises to move AI workloads from Nvidia to Trainium, they must bear enormous software adaptation friction.
DOS happens to be the friction remover. By bringing DOS into the AWS ecosystem, Amazon's logic is: make Trainium go from "requiring huge migration investment" to "switchable within weeks," so that more customers are willing to place part of their workloads on Trainium. This is a more direct strategic synergy than Anthropic's—using software to pry open hardware sales.
Yet there is a deep paradox here: by acquiring DOS, Amazon wants "Trainium's best partner," but Decart's greatest value to customers is precisely that "it runs on anyone's chips." If Amazon binds DOS deeply into AWS, will customers of other chipmakers still choose it? Once chip neutrality is damaged, a considerable portion of DOS's market value will evaporate with it. How to balance "strategic bundling" and "maintaining an open ecosystem" will be the hardest question to answer after this acquisition closes.
Reference Point Among Chipmakers: AMD Bought Fei-Fei Li's Company for $8.2 Billion
Decart's acquisition talks are not an isolated event but part of a larger wave. Just earlier this month, AMD announced an all-stock acquisition of World Labs for $8.2 billion—the spatial intelligence AI company founded by Fei-Fei Li, focused on generating interactive 3D environments from images and video. Fei-Fei Li will become AMD's chief scientist.
According to CNBC, AMD's move aims to integrate AI model research more closely with chip hardware development and to pre-position for next-generation AI workloads (inference, robotics, physical AI). This is strikingly similar to Amazon's logic in acquiring Decart: chipmakers and cloud service providers are both penetrating the software layer, trying to build software moats for their own hardware.
By comparison, the two deals point to the same structural judgment: in AI infrastructure competition, the way to win compute is no longer just to stack faster chips, but to make developers' work easier on your chips—this is a war in which software determines the fate of hardware.
Valuation Anchor and Risks
What does $7 billion mean for a company founded only three years ago whose revenue scale has not yet been publicly disclosed? Judging from its funding record, Decart has raised $450 million cumulatively, and the valuation in these talks is nearly double its May funding valuation. This premium comes largely from bidding competition—three rounds of bidding by Nvidia, Anthropic, and Amazon pushed up the floor price, rather than a fundamental change in Decart's business data.
There is currently no sufficient independent third-party validation of DOS's core technical barrier. "Compressing the optimization cycle from months to weeks" and "up to 8x acceleration" both come from the company's own statements, and no peer-reviewed benchmark data has been publicly released. Amazon is paying a bidding premium to buy a technical promise, not a mature product that has been repeatedly validated in large-scale production environments.
Independent Judgment
The strategic logic of Amazon's deal is clear, but the execution risk is underestimated. Decart's greatest asset—chip neutrality—faces structural damage the moment it is acquired by a cloud giant. If Amazon cannot convince the market that AWS will continue to maintain a truly open DOS ecosystem, Decart's value will degrade from "general-purpose infrastructure" to "a Trainium-specific tool," directly weakening the return on this acquisition.
Conversely, if Amazon is smart enough to operate DOS as a kind of "neutral platform" (allowing Nvidia and Google customers to continue using it), it could genuinely build a software ecosystem that is more open than CUDA but equally sticky. This is a harder path to execute, but one with multiplied value.
What this acquisition ultimately proves is that the chip-adaptation software layer has upgraded from tooling infrastructure to a strategic asset. When Decart was founded three years ago, no one treated it as a $7 billion target. Now that Nvidia, Anthropic, and Amazon have successively made bids, the importance of this niche track needs no one to endorse it.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接