The Twin Engine Strategy That Propels AWS Is Working Well

Like Google and Meta Platforms, Amazon knows exactly how to infuse AI into its business operations such as online retail, transportation, advertising, and even the Amazon Web Services cloud. Just like Google and IBM have been their own Customer Zero for AI efforts, Amazon has been learning how to use AI to replace or enhance business functions, not just for itself but so it can better understand how to sell such expertise to external customers and get them using the AWS cloud.

So it is not, we think, a coincidence at all that Amazon has laid off what looks like more than 60,000 people in its corporate employee base. Some of these layoffs have to do with over-hiring during coronavirus pandemic that started in 2020; some of it is just “workforce rebalancing” as IBM has called it for years. But, the fact remains that there are more corporate workers today than there were five years ago, and on top of that, there are another 1.2 million warehouse employees at Amazon (down about 100,000 from 2021 levels.

It is easy – far too easy – to make a one-to-one correlation between the 60,000 recent corporate layoffs and the rise of GenAI as a worker within Amazon. The world’s biggest online retailer, one of its biggest advertisers and media companies, and the world’s biggest cloud also wants to flatten its corporate organization – and would have needed to do this anyway. But as go the physical robots in the Amazon warehouses, so will go the AI agents in the Amazon Web Services data warehouses. Amazon had around 200,000 robots helping move and pack stuff into cardboard boxes in 2019 before, and now it is well over 1 million and probably well on its way to being much higher. You can bet that is Jeff Bezos and Andy Jassy could replace every person in a warehouse with a robot, they would. We have no idea how practical that is, but we do know that with every passing year it gets more likely.

It is harder to say that about the corporate operations running the Amazon businesses. People are still buying from people. People are still making the calls. We don’t know what the equilibrium is, and a lot depends on how good the AI agents are several years from now. The Rufus shopping agent that Amazon wants me to employ is but a first step. If money gets tight enough, people will use Rufus to do opportunistic buying. If profits get hard enough to come by, Amazon will use AI agents to do opportunistic firing. It is that simple.

All we know for sure about modeling the future is that it is important to watch what Amazon does and how AWS will benefit from the company’s deployment of AI technologies and how that experience will infuse its experience with millions of other companies. Likewise, the experience of AWS with millions of customers playing around with AI will infuse what the Amazon mothership does and does not do.

The good news for Amazon is that it has built a large advertising business that can not only buy AI services from AWS as a showcase, but, we think, is so wildly popular and profitable that it can also pay more than its fair share of the enormous capital expenses that Bezos & Co are committing to this year and beyond.

Take a look at this chart and you will see what we mean:

In this chart above, we compare revenues of what we think of as the core AWS systems business – compute, networking, and storage – against the relatively new Amazon Ads business, for which we only have five years of data based on things the company says in its quarterly reports and guesses from third party advertising market research companies that fill in the blanks in the earlier years.

We fully realize that separating out the underlying AWS systems business from the rest of the AWS stack, including platform services and software rented out, is dubious, but we have been making our estimates long before this advertising business even existed. These are two distinct datasets, across time. But look at how they twine! It is not causation, but it is correlation for sure.

The important thing is that Amazon Ads probably has much higher operating margins than the AWS core hardware business, and it can help pay for the AI buildout.

“We expect to invest about $200 billion in capital expenditures across Amazon, but predominantly in AWS because we have very high demand customers who really want AWS for core and AI workloads,” Amazon chief executive officer Jassy explained on the call with Wall Street. “And we are monetizing capacity as fast as we can install it. We have deep experience understanding demand signals in the AWS business and then turning that capacity into strong return on invested capital. We are confident this will be the case here as well.”

In the fourth quarter of 2025, Amazon spent $40.47 billion on capital expenses, an increase of 43.1 percent over the year ago period, and for all of 2025, it shelled out $134.73 billion, an increase of 60.5 percent compared to the tad bit less than $84 billion that Amazon spent on infrastructure in 2024. Our model suggests that AWS spent around $115 billion on IT infrastructure, and of this around $105 billion was for AI infrastructure. So AI was 78 percent or so of all capex, with other IT being around 7 percent and the remaining 14.5 percent being for warehouses and transportation equipment for the Amazon network of retail operations.

“We are growing at really an unprecedented rate yet,” Jassy said about the capex spending. “I think every provider would tell you, including us, that we could actually grow faster if we had all the supply that we could take. And so we are being incredibly scrappy around that.”

In the past twelve months, said Jassy, Amazon has added 3.9 gigawatts of datacenter capacity, which averages out to a very scrappy $29.5 billion per gigawatt where the big model builders like Anthropic and OpenAI are spending anywhere from $45 billion to $60 billion per gigawatt. Jassy added that Amazon added 1.2 gigawatts in Q4, and said further that only back in 2022, when AWS was at an annualized run rate of $80 billion at year end, it only had around 2 gigawatts of capacity installed. When we model the in-between years, that means AWS has around 6 gigawatts of total capacity installed as 2025 came to an end, and Jassy has said in past statements that it will double again by 2027, which will be 12 gigawatts. At the prevailing price that AWS is paying – call it $30 billion per gigawatt – that is $180 billion to add 6 gigawatts. So, the $200 billion that Amazon will spend on capex in 2026 will cover the bulk of the cost of the gear (warehouse and transportation stuff) that will be installed and ready by 2027 it looks like.

If the spending is doubling, the compute capacity of the gear is also probably doubling to quadrupling, depending on the computing precision used, as AWS moves through the Nvidia roadmap and its own Trainium roadmap. Software improvements over the two years should yield somewhere between 3X and 4X more performance if history is any guide. So the amount of compute AWS will have for AI will be vastly more than the increase in spending alone indicates.

Which is, we think, the key driver of renewed revenue growth for AWS. As inference processing gets cheaper, the elasticity of demand will increase faster and burn a lot more compute, fueling another reinvestment cycle:

It remains to be seen if growth can go as high as 30 percent or 35 percent year on year – or even higher. A lot depends on how GenAI bots and agents are adopted by enterprises and how much they customize their training.

Some interesting tidbits: Jassy said on the call that the custom compute engines at AWS – the Graviton Arm server CPU and the Trainium1 and Trainium2 AI XPU engines – ended 2025 with a $10 billion annualized run rate. That means those instanced brought in $2.5 billion in rent, and this was driven in large part by the fleet of 1.4 million Tranium2 million chips. Trainium3 installations are ramping now and all of the Trainium3 capacity for the chips it will install will be allocated by the middle of 2026. Presumably these Trainium3 systems will be installed this year and next, and probably in the millions and maybe accounting for a very large share of the $200 billion in spending this year.

What that means, in our model, is that X86 CPUs and Nvidia and sometimes AMD GPUs accounted for around $12 billion in spending for instances on AWS in Q4 2025. Some years hence, Amazon will spend more money on Graviton and Trainium than it does on external chips – when it hard to say.

In Q4 2025, AWS brought in $35.58 billion, up 23.6 percent, and operating income was under pressure a bit from chip design and manufacturing costs and only rose by 17.2 percent to $12.47 billion.

For the full year, AWS had $128.73 billion in sales, up 19.7 percent, with operating income of $45.61 billion, up 14.5 percent.

Here is how we think the breakdown of compute, storage, networking, and software revenues has broken down at AWS over the years:

As we have pointed out in the past, Amazon has never, ever given any indication of how AWS revenues break down across the four buckets we have created in that chart above, and it has never called us up to confirm or deny our model. We do this merely because we need to separate these out to understand what drives the AWS business.

For many years, the hardware was the main driver, but as AWS started building a complete platform, with networking and data services as well as development tools and sometimes full-blow applications, software came to dominate the revenue stream. But with GenAI hardware costing do damned much, and having access to it is so dear that Nvidia and AMD can charge a lot for it and AWS can turn around and charge an even higher premium to rent it, the compute part of the revenue stream has been skyrocketing at AWS and, we think, now drives more revenue than the software stack does.

And, because AWS is smart with Trainium as it has been with Graviton with core compute, homegrown chips can undercut Nvidia and AMD capacity rentals and still probably run AWS an equal or better operating profit.

AWS is in a win-win scenario here, as long as the money holds out. And until GenAI normalizes, it can invest in a twin engine approach for compute and keep the pressure on third party engine suppliers. In the long run, there is no reason to believe that the bulk of AWS DPU, CPU, and GPU/XPU compute power will come from its Annapurna Labs designers, not outside the company. And if you want to rent a third party device on the AWS cloud, you will pay a premium for that.