Amazon Q2 earnings put AWS margin and AI run-rate claims under scrutiny

Amazon Q2 earnings drew attention beyond the top-line results after the company reported a 39.4% AWS operating margin and framed several AI-related metrics without giving the underlying numbers. The release said AWS made its Kiro coding agent available on iOS, described both its AI business and chips business as exceeding $25 billion run rates, and said Bedrock spending in Q2 topped all previous quarters combined.

The Kiro claim is the simplest example of the broader disclosure issue. Amazon’s earnings materials said AWS made the spec-driven coding agent available on iOS. AWS’s public Kiro page, however, said users could request early access and that a limited number of people would be invited to test the app through Apple’s TestFlight, with a link to be sent when the app is ready. That means the product was not generally downloadable through the App Store based on the information AWS itself posted.

What did Amazon say about AWS margin in Q2?

AWS reported a 39.4% operating margin, a figure above published analyst estimates. On the earnings call, CFO Brian Olsavsky said the result included roughly $600 million in mark-to-market gains from energy derivative contracts.

Olsavsky said AWS margin increased 650 basis points year over year, or 520 basis points excluding the derivative accounting gain. He also said those adjustments had not been significant in earlier quarters, and Amazon’s Q3 guidance assumes no effect from similar remeasurements. The disclosure matters because the headline margin is likely to be cited as evidence that AWS can fund AI infrastructure expansion profitably, even though part of the quarter’s outperformance came from electricity hedging.

Amazon’s AI and chips run-rate claims overlap

Amazon said its AI business and its chips business each exceeded a $25 billion run rate. The company did not say whether those figures count some of the same spending more than once.

The chips label also needs context. Amazon’s Trainium and Graviton are not sold broadly as standalone chips with public price lists. Revenue tied to those products is booked through AWS services such as EC2 instances and possibly higher-level services including Bedrock and SageMaker. An EC2 rental includes more than silicon: storage, networking, data center infrastructure and AWS margin are all part of the commercial product.

On the call, Morgan Stanley’s Brian Nowak asked when Amazon might sell Trainium to outside customers. CEO Andy Jassy said customers are showing more interest in getting Trainium outside Amazon’s cloud, that Amazon is discussing it, and that there is “a real chance” it could happen in the future. That answer indicates Amazon is not yet treating third-party chip sales as the core of the reported run-rate business.

Amazon also said Bedrock customers spent more in Q2 than in all previous quarters combined. The company did not disclose Bedrock revenue, so the statement shows acceleration but gives no baseline for judging scale. Bedrock had been available for 10 prior quarters, according to the discussion around the release.

AI capex depends on demand that is still concentrated

Jassy raised Amazon’s annual spending plan to $220 billion and said backlog reached $496 billion, up $132 billion in the quarter. The same quarter included Anthropic’s $100 billion commitment over 10 years, according to the call discussion. Amazon also booked $53.4 billion in gains on its Anthropic stake, while free cash flow moved unfavorably by $26 billion and the company sold $25 billion in bonds.

Jassy described AI demand as “very barbelled,” with AI labs consuming large amounts of compute on one side and enterprises focused on cost avoidance on the other. He said the middle, enterprise production workloads running inference at scale, has not yet broadly arrived and that he does not know whether it will follow the same steep path as lab demand.

The quarter was strong by Amazon’s own disclosed figures: AWS grew 37% year over year, and the company reported sizable backlog. The unanswered questions sit in the presentation of the numbers: how much AI and chips revenue overlaps, how durable non-lab demand will be, and how much of the AWS margin story investors should assign to operations rather than one-time derivative gains.

This story draws on original reporting from The Register.