Amazon Is Paying $7 Billion for a Startup Anthropic Wouldn't Buy at $6 Billion

In August, Anthropic looked at buying an Israeli AI startup called Decart for roughly $6 billion. In September, after due diligence, it decided not to. On Sunday, the Wall Street Journal reported that Amazon is now in talks to pay about $7 billion for the same company.

That is a $1 billion premium over the price a frontier lab was willing to explore, paid by a buyer that has a specific problem to solve. The reported valuation is also about 75% above where Decart stood five months ago. Nothing is signed, so the useful work is separating what is reported from what is inferred.

What Actually Happened

Per Reuters, which relayed the Journal's report, Amazon is in talks to acquire Decart in a deal that could be valued at around $7 billion. The Journal cited people familiar with the matter. Reuters says a deal could happen soon if talks do not fall apart or another suitor does not emerge. Amazon declined to comment outside regular business hours, and Decart did not respond to requests for comment.

Globes, the Israeli business paper, reports the same core facts and adds that the negotiations have not been finalized, with a possibility that they could fail or that another potential buyer could enter. Globes describes the deal, if signed at the reported valuation, as one of the biggest exits in the history of Israel's technology industry and one of the biggest acquisitions by Amazon in recent years.

The price history is where the story gets its edge. The Journal, citing PitchBook data via Reuters, said Decart was valued at nearly $4 billion in May. That month, Reuters notes, Decart raised $300 million in a round led by Radical Ventures, with Nvidia joining as a new investor. Globes puts the arithmetic plainly: $7 billion would be a 75% increase within five months. Because May's figure was "nearly" $4 billion, the real step-up is at least that.

Then there is Anthropic. Reuters, citing Bloomberg, says Anthropic explored a deal in August for potentially $6 billion and decided in September not to pursue it after conducting due diligence. Globes goes further, reporting that Decart's founders had previously rejected an offer from Nvidia that was higher than Anthropic's, and that Anthropic's proposal was a share deal that anticipated an Anthropic IPO. That second point is Globes' reporting alone, and we have not seen it confirmed by a second outlet.

As for what Decart makes, Reuters describes the Lucy model, which is backed by Nvidia and can edit live video in real time, and Oasis, a model that generates simulated environments for training and testing robotics and autonomous driving systems. Globes says the company was founded in 2023 by Dean and Orian Leitersdorf and Moshe Shalev and has raised more than $400 million from investors that include Sequoia Capital, Benchmark and Radical Ventures.

Reading a Price Against a Walk-Away

A completed acquisition tells you what one buyer would pay. A walk-away tells you where another buyer stopped. Here we have both, and they sit close enough together to be informative.

Start with what is not known. No outlet we reviewed reports why Anthropic passed. Due diligence can end a deal over price, product fit, retention of key staff, legal exposure, or a dozen other reasons. The reporting says only that Anthropic looked, and decided against it. Reading that as a verdict on Decart's technology would be a guess.

What the numbers do support is narrower. Two large strategic buyers, plus Nvidia per Globes, have been reported as interested in a company that was valued at nearly $4 billion this spring. The founders reportedly turned down a higher Nvidia offer while an Anthropic deal was live, and Globes ties the Anthropic offer to an IPO-dependent share structure. A price that depends on a buyer's future public listing is a different instrument from a price paid in Amazon's cash and stock, and a founder weighing the two could reasonably favor the more certain one. That is inference, not reporting, and it is worth holding loosely.

The market-signal reading is simpler. Private AI valuations are moving in large steps over short periods, and strategic acquirers are willing to pay them. A 75% mark-up in five months on a company with a little over $400 million raised implies that investors who came in at $4 billion are about to book a quick gain, and that the next funding round for a comparable startup will be priced off this one.

The Trainium Question

Why Amazon? The most developed answer comes from Tech Times, and it comes with caveats.

Tech Times argues that Decart's most strategic asset is not the video models Reuters describes but an inference optimization product the company announced in May, which it says runs across Nvidia GPUs, Google TPUs and Amazon's own Trainium chips. The thesis is that the hard part of moving AI workloads to non-Nvidia silicon is porting the software. Nvidia's CUDA ecosystem has years of accumulated libraries and tooling, and Tech Times describes Amazon's competing Neuron SDK as still catching up. A tool that automates the porting work would, in that reading, lower the cost of choosing Trainium over Nvidia hardware.

Tech Times also reports that Amazon stated Decart's Lucy2 model reached 80% Model FLOPS Utilization on Trainium3 chips, and that Decart claims more than 1,600 tokens per second for agentic inference against an industry average of roughly 200, along with 8x speed and 100x cost-efficiency gains. The outlet itself says the Decart figures are company-stated and not independently verified. We are repeating them as claims, not as measured results.

Two cautions apply. First, the Trainium rationale is Tech Times' analysis. Reuters and Globes report the talks and the price and do not state Amazon's intent. Second, Tech Times also describes a $300 million strategic partnership between Amazon and Decart formed in May, while Reuters describes a $300 million funding round that month led by Radical Ventures. Those may be different things or one thing described two ways, and we could not reconcile them from the sources in hand.

What holds up regardless: if Amazon's goal is to make customers comfortable running inference on Trainium, then the price is a bet that software friction, not silicon, is what keeps developers on Nvidia. That bet is plausible. It is also unproven until a deal closes and customers actually move workloads.

What the Deal Could Change for Inference Costs

Whether or not the optimization claims survive independent testing, the direction of travel matters for anyone paying inference bills. Tech Times cites industry research that inference can account for 80% to 90% of an AI system's lifetime operating cost. That figure is the outlet's, not something we verified, but it matches how most AI products behave once they have real usage: training is a line item, serving is a utility bill.

If a hyperscaler owns the tooling that moves workloads between chips, the pricing conversation for cloud AI changes. Today a builder choosing between Nvidia-based instances and Trainium-based instances prices in engineering time to port and tune. A tool that shrinks that time lowers the switching cost, which gives the buyer room to discount the cheaper silicon and gives customers a stronger hand in negotiation. If the tool stays cross-platform, as Tech Times says it currently is, builders could use it to compare chips more evenly. If Amazon folds it into AWS, the cross-vendor part could narrow. Nobody has reported which way this goes, and it should be watched.

What Builders Should Take From It

• Treat the $7 billion as a report, not a closing. Reuters and Globes both flag that talks could fail or another bidder could appear. Do not build vendor decisions on a signed deal that does not exist yet.
• If you depend on Decart's APIs or models, ask now. Find out what an Amazon ownership scenario would mean for pricing, availability on non-AWS clouds, and data handling. Nothing has been announced, but acquisitions of dependencies are the case where contract terms matter most.
• Benchmark inference on more than one chip. Whether the claims about porting speed hold or not, the cost spread between Nvidia and alternative silicon is the variable a hyperscaler is trying to move. Run your own workload on at least two backends and record cost per million tokens, not vendor throughput numbers.
• Price your own exit or raise off real comparables. A 75% valuation jump in five months is a data point for founders and investors in adjacent categories, but it is one transaction involving a company with several bidders. It sets a ceiling more than a median.
• Read walk-aways as signal, but do not over-read them. Anthropic's diligence outcome tells you a sophisticated buyer stopped at about $6 billion. It does not tell you why.
• Watch for the second-order filings. If Amazon announces, expect disclosure on deal structure, key-employee retention and any Israeli regulatory review. Those documents will say more about strategy than the press reports do.

Developer312 covers the AI business signals builders actually need to act on. Get the weekday briefing at developer312.com.