Amazon’s $7B Decart Talks Show AI’s Real Prize Is Cheaper Compute
A $7 billion price tag for Decart is not a bet on another shiny AI app. It is Amazon admitting the money sits underneath the models: whoever makes compute cheaper owns the toll road.
A $7 billion price tag for Decart is not a bet on another shiny AI app. It is Amazon admitting the money sits underneath the models: whoever makes compute cheaper owns the toll road.
Reuters reported on October 11 that Amazon is in talks to buy AI startup Decart for around $7 billion, citing a Wall Street Journal report. Nothing is signed, neither company confirmed the talks, and Reuters said it could not independently verify them. Fine. That is still enough to tell you where the serious fight is heading.
It is not mainly about who has the cleverest chatbot.
It is about who can produce useful intelligence at a cost low enough that every business can afford to use it all day, every day.
Amazon is not shopping for a demo
Decart is not just another company making AI-generated videos for people who want to turn their cat into a Viking. Its real asset is a performance layer called DOS, designed to optimise AI training and inference across hardware. In plain English: it helps companies get more work from the chips they already have, and makes it easier to move workloads when a particular processor is scarce, expensive or no longer the best option.
That is properly valuable.
The AI industry has spent years behaving as though buying more chips is a strategy. It is not. It is an expensive response to an operational problem. If every useful AI task requires ever more power, memory, cooling, data centres and capital, then the winners will not be the companies with the flashiest model. They will be the ones that squeeze the most output from every watt and every dollar.
Decart has also built Lucy, a real-time video model, and Oasis, a system for generating interactive simulated environments. Those products are interesting. But the DOS infrastructure layer is likely the bit that makes Amazon sit up straight.
Amazon already has a massive incentive to make its own AI chips more attractive. Its Trainium and Graviton chip businesses had a combined annual revenue run rate above $10 billion in its latest full-year results. Amazon said Trainium2 was fully subscribed, with 1.4 million chips deployed, while nearly all Trainium3 supply was expected to be committed by mid-2026.
That is not a supply problem you solve with a nice keynote. It is a supply problem you solve by making customers confident they can run workloads efficiently on your hardware rather than defaulting to Nvidia.
The $7 billion number is the message
Decart raised $300 million in May at a valuation of roughly $4 billion, according to reporting at the time. A potential $7 billion sale only months later would be a hefty premium for a company with barely three years of public history.
That sounds mad until you look at what is being bought.
Amazon would not simply be buying engineers or an AI video product. It would be buying leverage over the cost curve of AI. If Decart’s technology lets AWS customers get better performance from Trainium, switch between chips with less pain, or run models more efficiently, the value does not sit inside Decart’s revenue line. It spreads across AWS infrastructure, Bedrock, Amazon’s own customer products and every enterprise workload Amazon can keep inside its cloud.
That is the bit founders routinely miss. The value of a piece of infrastructure is often not its subscription revenue. It is the amount of spending it redirects, protects or unlocks elsewhere.
Amazon expects to spend about $200 billion in capital expenditure during 2026, driven in large part by AI, chips, robotics and related infrastructure. Its free cash flow fell to $11.2 billion over the trailing 12 months as property and equipment investment surged. When you are spending at that scale, an optimisation tool that improves utilisation by even a small percentage can be worth an absurd amount of money.
A one percent improvement on a tiny software budget is lunch money. A one percent improvement on a giant compute estate is a board-level event.
The background: AWS is building a whole machine, not a model
Plenty of people still frame AI as a horse race between OpenAI, Anthropic, Google and Meta. That makes for simple headlines, but it misses the bigger commercial game.
Amazon is building a vertically integrated machine: cloud capacity, custom chips, models, agent tooling, distribution and customer applications. It already offers more than 20 managed models through Amazon Bedrock, including models from Amazon, Anthropic, Google, OpenAI, Nvidia and others. The obvious aim is not to guess one winning model. It is to make AWS the place where customers can use whichever model wins.
Its February partnership with OpenAI made that especially clear. Amazon committed to invest $50 billion in OpenAI, while OpenAI agreed to consume about two gigawatts of Trainium capacity through AWS infrastructure. The companies said the broader compute agreement was expanded by $100 billion over eight years.
Read that again. The cloud provider is not merely renting servers. It is investing in the model company, supplying the chips, hosting the workloads, distributing enterprise products and helping build the environment where those AI agents run.
That is what a moat looks like when it has cables attached.
A Decart acquisition would fit this picture neatly. Amazon could make Trainium more useful, lower the friction of using non-Nvidia hardware, and potentially give customers confidence that their AI stack will not be welded permanently to one supplier.
The overlooked angle: Amazon may be buying optionality, not exclusivity
Here is the contrarian view: the best reason for Amazon to buy Decart is not to lock customers into Trainium. It may be to make AWS the least painful place to run mixed hardware.
That sounds counterintuitive. Why would Amazon spend billions on software that helps customers use rival processors?
Because sophisticated customers hate being trapped. If AWS can say, “Bring your models here, run them on Nvidia, Trainium or something else, and we will help you get the economics right,” it becomes more valuable than a cloud that demands religious devotion to its own chips.
The real product is not a processor. The real product is freedom with good economics.
That is also why this deal, if it happens, matters well beyond Decart’s founders and investors. AI infrastructure is becoming a game of control points. Chips matter. Networking matters. Power matters. Data centres matter. But the software layer that decides where workloads run and how efficiently they run may become one of the most important control points of all.
The company that owns that layer can influence purchasing decisions without needing to manufacture every component.
There is a warning here too. Startups building AI wrappers, generic agents or clever front ends should pay attention. Many of those businesses will be features. Infrastructure that materially reduces cost, latency or dependence can become an asset worth fighting over.
Brutal distinction, but a useful one.
This is also a reminder that valuation is not value
A $7 billion potential price does not mean Decart is magically “worth” $7 billion in some cosmic sense. It means a buyer may see strategic value that a spreadsheet based solely on current revenue cannot capture.
Founders love this idea when it benefits them, then forget it when assessing their own businesses. They tell themselves their product is strategic because it uses AI, has a dashboard and employs clever people.
No. Strategic means your absence creates pain for a buyer with a very large existing machine.
Decart may matter because it can improve the economics of Amazon’s machine at a moment when chip capacity is tight and AI capital spending is enormous. That is a much stronger position than merely having a good product.
For investors, the lesson is equally blunt: do not confuse demand for AI with durable value in every AI company. The best businesses may be hiding in the boring bits—performance tooling, data plumbing, security, deployment, evaluation and infrastructure management. The glamour gets the headlines. The picks-and-shovels layer gets paid repeatedly.
What this means for you
If you run a business, stop asking your team whether you are “using AI enough.” That is corporate theatre. Ask three better questions tomorrow.
First: Which recurring task costs us the most labour, time or error today? Pick one with clear volume and a measurable baseline. AI without a baseline is just a demo with a budget.
Second: What does each successful AI outcome cost us? Track cost per resolved customer issue, cost per qualified lead, cost per report produced or cost per line of code reviewed. Usage numbers are vanity. Unit economics are the truth.
Third: Can we switch providers without rebuilding the business? Keep your data portable, log model performance, and avoid building a workflow so tightly around one vendor that a price increase or product shift leaves you helpless.
For founders, build where the pain is expensive and repeatable. A prettier interface will not save you when OpenAI, Google or Amazon releases the same feature. A measurable reduction in compute cost, compliance risk, latency or headcount waste might.
And for investors, quit chasing every company that says “AI-native.” Look for businesses that own a toll road: something other people must keep paying for as AI usage grows.
Amazon’s reported Decart talks are a reminder that the great AI fortunes may not be made by the loudest model makers. They may be made by the operators who make intelligence cheaper, faster and harder to leave.