Amazon’s $7B Decart Acquisition Deal to Reduce Reliance on Nvidia GPUs

Amazon is in advanced talks to acquire AI startup Decart for approximately $7 billion, a deal that would give the company full ownership of software that lets developers run AI models across competing chip architectures. The Amazon Decart acquisition is not primarily about generative video. It is about owning the software layer that makes Amazon’s own Trainium chips a credible alternative to Nvidia GPUs, solving a migration problem that has blocked most AWS customers from switching away from Nvidia in the first place.

Decart’s software makes it easier for developers to move between different AI chips, a capability that has grown more important as companies scramble to acquire additional computing power. If the deal closes, it would surpass Amazon’s $6.5 billion purchase of MGM in 2022, making it the largest acquisition in the company’s history.

What Does Decart Actually Do, and Why Is It Worth $7 Billion?

Decart’s core value is its chip-portability software, not the AI models it has built on top of it. Founded in 2023 by Israeli engineers Dean Leitersdorf, Orian Leitersdorf, and Moshe Shalev, Decart developed software that enables AI developers to switch between processors from Nvidia, Amazon, Google, and other manufacturers more easily. The company’s Decart Optimization Stack (DOS) eliminates the need for custom programming libraries required by different chip types.

That sounds like a plumbing problem, and it is. Existing AI models are built for specific chips and require performance tuning and other work to run elsewhere, a burden DOS is designed to remove. The technology has gained importance as surging AI demand has made Nvidia GPUs both scarce and expensive. Relying on a single chip type leaves companies exposed to supply shortages and price increases, so the ability to switch chips flexibly is key to securing computing power reliably and holding down costs.

Nvidia’s enduring dominance has always stemmed from its deeply entrenched software ecosystem built around CUDA. Customers’ models, code, and engineering teams are deeply integrated into this ecosystem; simply offering a cheaper chip isn’t enough to convince them to switch. DOS directly attacks that lock-in. It adjusts how training and inference workloads execute on each hardware architecture, reducing the engineering work required when a customer changes chip families.

On a dollar-per-result basis, the technology has already proven itself. When AWS launched Trainium3 UltraServers at re:Invent 2025, Decart was the headline proof point. Decart, an AI lab specializing in efficient, optimized generative AI video and image models, was leveraging Trainium3’s capabilities for demanding workloads like real-time generative video, achieving 4x faster frame generation at half the cost of GPUs. Decart CEO Dean Leitersdorf confirmed those numbers directly: “Trainium3’s next-generation architecture delivers higher throughput, lower latency, and greater memory efficiency, allowing us to achieve up to 4x faster frame generation at half the cost of GPUs,” he said.

Decart employs more than 60 people and has raised over $450 million. The $7 billion price would mark a big jump from the company’s last funding round in May 2026, which valued it at $4 billion.

Why Did Nvidia, Anthropic, and Then Amazon All Try to Buy Decart?

The Amazon Decart deal latest development follows a bidding sequence that reveals more about the asset’s strategicvalue than any analyst note could. Nvidia approached Decart first. The logic was straightforward: own the tool that helps customers migrate away from your chips, and you can shape the terms on which that migration happens, or slow it down entirely. Nvidia’s interest alone confirmed that Decart’s software was not aniche product but a genuine threat to the GPU lock-in model.

Anthropic then entered talks at a reported $6 billion valuation. The AI safety company’s motive was different. Anthropic runs its largest models on massive GPU clusters and faces the same cost exposure as any hyperscaler. Owning Decart’s portability stack would let Anthropic shift workloads onto cheaper hardware, reducing its operating costs and dependence on third-party infrastructure. Those talks ultimately failed, with sources indicating the two sides could not agree on terms.

Amazon’s bid at $7 billion represents a $1 billion premium over Anthropic’s offer. That premium is essentially a statement: Amazon Decart $7 billion news reflects what AWS is willing to pay to prevent a competitor from locking up the one piece of software that could determine whether Trainium ever achieves mass adoption. Radical Ventures, one of Decart’s backers, helped bring the company to prominence in North American venture circles, and that investor network likely kept Amazon informed as Anthropic’s talks collapsed.

The bidding sequence matters for investors tracking this story. Three separate sophisticated buyers, including Nvidia itself, evaluated Decart and decided it was worth pursuing. That convergence of interest is a stronger validation of the technology than any benchmark.

How Does Decart’s Software Fit Into Amazon’s Chip Strategy?

Amazon has spent years building Trainium as a direct answer to Nvidia’s data center dominance, but adoption has been slow. The Amazon Trainium vs Nvidia GPU performance gap has narrowed with each generation, yet most AWS customers still default to GPU instances because their existing code runs there without modification. That inertia is the real competitive problem, and Decart’s chip-portability software is the most direct solution Amazon could buy.

The Trainium3 UltraServer, unveiled at re:Invent 2025, is already capable hardware. The bottleneck is software friction, not silicon. By acquiring Decart and integrating DOS into the AWS stack, Amazon could offer developers a migration path that does not require them to rewrite large portions of their training or inference code. An AWS customer running a large language model on Nvidia H100 instances could, in theory, move that workload to Trainium3 with far less engineering effort than is currently required.

That changes the competitive calculus for enterprises evaluating their AI infrastructure. Cost is already a factor. Many companies running sustained AI workloads at scale pay meaningfully more for Nvidia GPU hours than they would for equivalent Trainium capacity. The barrier has always been migration cost, not hardware price. If Amazon can remove that barrier through the DOS software layer, the economic case for switching becomes hard to ignore.

There is a counterargument worth taking seriously. Nvidia’s CUDA ecosystem is 17 years old and touches every major AI framework. A portability layer can reduce switching friction, but it cannot eliminate the performance tuning required to get optimal results on a new chip architecture. Some workloads, particularly those that rely on very specific CUDA primitives, may never run as efficiently on Trainium regardless of what abstraction layer sits on top. AWS customers with heavily customized GPU pipelines should not expect Decart to make their code chip-agnostic overnight.

What Are the Lucy and Oasis Models, and Do They Matter Here?

Decart has released two AI models that gained public attention. Oasis, a real-time generative world model, can produce interactive video environments and attracted significant interest when it launched in 2024. Lucy is a multimodal AI model capable of handling text, images, and video tasks. Both products have been impressive as demonstrations, but they are not the reason Amazon is paying $7 billion.

Oasis and Lucy function as proof-of-concept benchmarks that show what Decart’s optimization stack can do on real hardware. Running a real-time generative video model at high frame rates is one of the most compute-intensive tasks in AI, which makes it an ideal stress test for chip-switching software. If DOS can handle Oasis across multiple chip architectures without performance degradation, it can handle most commercial workloads as well. The models are a portfolio window, not the acquisition target.

That distinction matters for AWS developers wondering what changes if the deal closes. Decart’s generative video capabilities are unlikely to be launched as a standalone AWS consumer product. The more probable outcome is that DOS gets absorbed into the AWSNeuron SDK, which currently handles compiler and runtime support for Trainium and Inferentia chips. Decart’s team would expand that toolchain and focus on reducing the porting effort for enterprise customers.

What Does the Amazon Decart Acquisition Mean for Nvidia?

Nvidia is not about to lose its market position over one acquisition. The company held approximately 70 to 80 percent of the AI accelerator market in 2024, and its Blackwell GPU architecture has continued to generate record data center revenue into 2025. A single software acquisition by a competitor does not erase that lead.

What the Amazon Decart deal does is narrow the moat. Nvidia’s durability comes from CUDA lock-in as much as from chip performance. If DOS makes it meaningfully easier to move workloads off Nvidia hardware, the marginal cost of switching drops, and enterprise procurement teams begin treating Nvidia as one option rather than the only option. That is a slow erosion, not a sudden cliff. Nvidia’s stock has already priced in some risk from hyperscaler chip programs; this deal adds another data point to that thesis without changing the near-term revenue picture.

AWS itself has an incentive to keep some Nvidia capacity available regardless of how DOS performs. Enterprise customers often run mixed workloads, some of which will remain on GPU instances by preference or contractual obligation. Amazon’s goal is chip flexibility for its infrastructure, not the elimination of Nvidia from its catalog. The Amazon Decart deal latest intelligence suggests the two companies will continue to be significant business partners even as Amazon works to reduce its dependency at the margin.

Is the Amazon Decart Deal Done?

No. As of October 2026, the deal is described as advanced talks, not a signed agreement. M&A processes at this price point involve regulatory review, extended due diligence, and board approvals on both sides. The Amazon acquire Decart update most widely reported places the discussions in a late stage, but multiple factors could still disrupt a close.

Antitrust scrutiny is a real consideration. Regulators in the United States and European Union have shown increased willingness to examine large technology acquisitions, particularly those that could affect competition in AI infrastructure markets. A deal that gives Amazon exclusive control of chip-portability software used by competitors could attract that attention. The FTC’s approach to AI-adjacent acquisitions has been cautious since 2023, and a $7 billion price tag will not escape notice.

There is also the question of retention. Decart’s value sits in roughly 60 engineers who built DOS and the underlying optimization research. Acquisition agreements of this type typically include multi-year retention packages, but engineering talent in competitive AI markets does not always stay through long integration cycles. If key members of the Decart team depart post-close, Amazon would own the software but lose the team best positioned to extend it.

For AWS developers and enterprise customers watching the Decart AI startup acquisition talks: the deal’s strategic logic is sound, but the timeline to any product-level impact is likely 18 to 24 months after close, assuming regulatory approval and successful integration. Plan your chip migration strategy on your current options. Watch the AWS Neuron SDK release notes in 2027 for early signals that Decart’s work has been absorbed.

Frequently asked questions

What is Decart and what does it do?

Decart is an AI startup founded in 2023 by Dean Leitersdorf, Orian Leitersdorf, and Moshe Shalev. Its primary product is the Decart Optimization Stack (DOS), software that allows AI developers to run and migrate workloads across different chip architectures, including processors from Nvidia, Amazon, and Google, without needing to rewrite large portions of their code for each hardware type. The company has also built AI models, including Oasis, a real-time generative world model, and Lucy, a multimodal model, which serve as public demonstrations of what DOS can do on demanding compute tasks.

Why is Amazon willing to pay $7 billion for Decart?

Amazon’s Trainium chips are capable hardware, but most AWS customers have not switched to them because their existing AI code is built for Nvidia GPUs and would require significant engineering work to migrate. Decart’s DOS software directly addresses that friction. Byowning DOS, Amazon could lower the migration barrier enough to make Trainium a realistic default for enterprise workloads, which would reduce AWS’s dependence on Nvidia GPU supply and pricing. The $7 billion also reflects competitive pressure: both Nvidia and Anthropic pursued Decart before Amazon, and Amazon’s premium over Anthropic’s reported $6 billion offer signals how much it values keeping this software out of a rival’s hands.

Has the Amazon Decart acquisition been finalized?

No. As of October 2026, the two sides are in advanced talks but have not signed a definitive agreement. Regulatory review, extended due diligence, and board approvals on both sides remain ahead. Antitrust scrutiny from the FTC and EU regulators is a realistic risk given the deal’s size and the competitive implications of one company owning chip-portability software used across the AI industry.

Why did Nvidia try to buy Decart?

Nvidia’s CUDA software ecosystem is the primary reason customers stay on its GPUs even when alternatives exist. Decart’s DOS software reduces that lock-in by making it easier to move workloads to competing chips. Nvidia’s interest in acquiring Decart was a defensive move: owning the tool that enables migration away from CUDA would give Nvidia control over the terms and pace of that migration. That Nvidia pursued the company at all confirms that DOS poses a genuine structural threat to its software moat.

What happens to Decart’s Oasis and Lucy models if Amazon closes the deal?

The models are unlikely to become standalone AWS consumer products. Their most probable future is as internal benchmarks and proof-of-concept assets that demonstrate DOS’s capabilities on demanding workloads. The core engineering effort post-acquisition would focus on integrating DOS into the AWS Neuron SDK to reduce porting effort for enterprise customers moving workloads from Nvidia hardware to Trainium instances.

How does Decart’s software compare to Nvidia’s CUDA?

CUDA is a proprietary programming platform that Nvidia has developed over 17 years and that underpins virtually every major AI framework. It is not a portability layer; it is a deep, chip-specific ecosystem. DOS sits above that level and is designed to abstract away chip-specific requirements so that developers do not need to write separate optimized code for each architecture. DOS can reduce migration friction but cannot fully replicate the performance of code that has been tuned specifically for CUDA primitives. For most commercial AI workloads, that gap is manageable. For highly customized GPU pipelines, it may not be.

What does the Decart acquisition mean for AWS customers currently on Nvidia GPU instances?

In the short term, nothing changes. The deal has not closed, and even after a close, integrating DOS into AWS infrastructure and validating it against enterprise workloads will take time. A realistic window for product-level impact is 18 to 24 months after regulatory approval and deal close. Customers should monitor the AWS Neuron SDK release notes through 2027 for early signals that Decart’s tooling has been incorporated. In the meantime, chip migration decisions should be based on current Trainium and Neuron SDK capabilities, not on anticipated future tooling.