Amazon Q2 earnings beat as AWS grows 37%, but AI cost overruns and Nova retreat cloud outlook

Amazon reported second-quarter revenue of $200.6 billion, surpassing analyst estimates of $196.43 billion, with diluted earnings per share of $5.75 against a forecast of $1.82. Operating income climbed 43% to $27.5 billion. Advertising revenue rose 26%, and the company highlighted record delivery speeds for Prime members, with over 40% more items delivered same-day or overnight.

The standout was Amazon Web Services, where sales jumped 37% year-on-year to $42.2 billion, the fifth straight quarter of accelerating growth. AWS now accounts for about a fifth of total revenue. Chief executive Andy Jassy told investors: "AWS is booming." In a shareholder letter, he added that the AI and chips businesses each surpassed $25 billion in annualized run rates.

AWS is booming, growing 36.7% year-over-year in Q2 — our fastest growth in 18 quarters — and our AI and Chips businesses each eclipsed run rates of more than $25 billion.

AI cost overruns

While AWS powers much of the AI industry, Amazon's own internal use of the technology has led to what senior engineers described as "catastrophically expensive" cost overruns. At a staff meeting on Tuesday, employees were told of three incidents. The largest involved spending $1.8 million on Anthropic's Claude Sonnet to match author details with product listings; the project ran 860% over budget, went undetected for five months, and the deployment failed. A financial auditing tool overshot by about $541,000, and a system to improve delivery speeds added $134,000 in unexpected costs, taking more than a fortnight to spot.

One senior employee told the Financial Times: "It's difficult to figure out how much anything [AI related] costs." Amazon said it was "experimenting, learning and improving" and that the examples were "cherry-picking small, isolated examples" that did not reflect widespread practice. The presentation noted the cases spanned only a handful of teams among roughly 300,000 corporate staff.

As with any new technology, we're experimenting, learning and improving how we use it, including how we drive cost efficiencies.

Nova models wound down

Separately, Amazon is deprecating most of its in-house Nova AI models, including the high-end Premier and Omni text models, the Reel video generator, and the Canvas image generator. The models remain supported for existing customers but are no longer a development priority. Resources are shifting to a new effort called Frontier Model Research, led by Pieter Abbeel, who joined through the 2024 acquisition of Covariant. A new flagship model is expected at the re:Invent conference this autumn.

Amazon will keep Nova 2 Lite, Nova 2 Sonic, the Nova Forge customization service, and the Nova Act agent tool. The company's San Francisco AGI site, an 80-person research group, has closed. The retreat follows earlier moves: the AGI Lab was shut, and former AI chief Rohit Prasad left at the end of 2025. Peter DeSantis, who took over the consolidated group in December, has concentrated talent on fewer frontier bets.

Financial footing and outlook

Capital expenditure across Amazon climbed to $53 billion in the quarter, with free cash flow turning negative to $7.6 billion on a trailing 12-month basis. The company expects to spend $200 billion this year, the vast majority on AI and data center infrastructure. To help fund this, Amazon has undertaken sweeping layoffs and raised $50 billion through bond issuances in March, though it faced weakening demand for long-dated debt earlier this month.

The earnings beat and AWS momentum contrast with the internal AI stumbles and the Nova pullback, illustrating the dual pressures of leading in cloud while managing the costs of the AI transition.