Overview: Dissecting Amazon’s Earnings
Amazon announced its earnings today, and in a state of professional discontent, I read the entire document [PDF]. “How long can I last before the red mist of anger descends” is an entertaining challenge, and today I managed to reach the end of the second page when I came across a bullet point claiming that AWS “released its spec-driven [sic] coding agent, Kiro, for iOS.”
The Kiro Deception
“Available” But Not Quite
Indeed, I was present when they revealed it at the New York summit six weeks ago. As of now, their website, which I’ve captured, indicates I can “request early access,” since “We’ll invite a limited number of users to test the app via Apple’s TestFlight, and a link will be sent to everyone when it’s prepared.” Thus, Kiro is “available” similarly to how I am available to join the NBA. You can contort your logic and argue that this assertion is technically accurate, but in practical terms, it’s what we typically refer to as “a falsehood.”
You must receive a specific invitation for Apple’s developer beta testing tool, which allows a limited number of users to experience an unreleased version. It cannot be downloaded to your phone, and there’s no page in the App Store showcasing the product. The delay is likely a result of Apple’s complex App Store policies, which I somewhat sympathize with — but this is an earnings report. If they’re going to “shade the truth” in this manner, what else are they hiding?
Once You See It, It’s Everywhere
There are numerous other claims that might not withstand scrutiny. Graviton boasts “up to 30 to 40% improved price-performance,” which I only accept since I’ve witnessed the figures on customer workloads. The explicit mention that their AI and chips divisions are each surpassing $25 billion run rates in successive bullet points, with no indication whether those sums overlap (this point will be explored soon). And their Bedrock assertion: “customers spent more in Q2 than all previous quarters combined,” which sounds impressive until you realize they mean the last 90 days exceeded the previous 10 quarters in which Bedrock has been available.
The Profit Conundrum
Without real numbers attached to these claims, it sounds like for the initial years Bedrock was appearing at a party with a hat but no pants. Then there’s the AWS operating margin of 39.4%, which exceeded all published analyst forecasts and which everyone will inevitably reference as selected proof that the AI expansion is generating profits. During the call, CFO Brian Olsavsky revealed that it includes about $600 million of mark-to-market profits from energy derivative contracts. According to him, AWS margins increased by 650 basis points year over year, or 520 “if you disregard the derivative accounting gain.” Remove that gain yourself (witness the power of arithmetic!) and the inflated margin goes right back into the range that analysts had predicted.
Amazon now hedges electricity like an airline hedges jet fuel, and this quarter’s hedges were successful. Olsavsky mentioned these adjustments “haven’t been significant in earlier quarters.” The first quarter they are significant, they affect AWS margin, and their Q3 guidance already anticipates no effect from these re-evaluations going forward. Amazon recognizes it’s merely noise but evidently saw no reason not to claim victory.
The Chips that Aren’t Physical Products
Revenue Without Tangible Goods
Returning to those conflicting $25 billion run rates I mentioned; labeling their “AI chips business” this way felt peculiar. That division has revenue, growth, and an impressive trajectory, along with numerous satisfied customers — but it lacks an actual product that can be purchased.
There’s no pricing for Trainium, you won’t receive a socketed Graviton chip for your next desktop build, and there isn’t even an external part number. What Amazon categorizes as “chips revenue” is the rental of EC2 instances (perhaps filtered through higher-level services like Bedrock, SageMaker, the half-baked agents that misinterpret your AWS bill, etc.), and an EC2 instance is not a chip. It encompasses the chip, plus the nVME, plus the NICs (which are built on Nitro, utilizing Amazon’s silicon), plus certain aspects of data transfer that somehow aren’t directly charged, plus the infrastructure where all of this resides — and then AWS’s margin is added on top.
The silicon itself is a minor line item in the internal bill of materials underlying its business. You don’t have to take my word for it; Amazon CEO and AI Marketing Manager Jassy lamented during last quarter’s call that the cost of components, “especially memory, has surged,” indicating that a substantial portion of the “chips business” revenue is derived from memory. The cynicism here is that the category exists to allow headline writers to mention it alongside Nvidia’s data center figures, which they inevitably will. But Nvidia’s $25 billion represents silicon sold as physical packaged chips, distributed from their loading area. Amazon’s is fully-loaded infrastructure rental. This is akin to a hotel comparing its income to that of a mattress company.
But there’s one more layer of absurdity. Olsavsky stated that the majority of Bedrock’s workloads rely on Trainium. So if you trace one Anthropic dollar through the earnings release, it counts as AI-business revenue, chips-business revenue, and AWS segment revenue. It’s nice to get a triple-acknowledgment for the same thing.
Future Commitments
You also don’t have to accept my assertion about the “no chips sold” aspect. During today’s call, Morgan Stanley’s Brian Nowak inquired when Amazon might start selling Trainium to third parties. I want one as well; I understand. Jassy responded that customers are increasingly eager to obtain Trainium “separate from our cloud,” that Amazon is “actively pursuing those discussions,” and that “there’s a real probability we’ll do that in the future.” IN THE FUTURE. “Yes, we have yet to sell a single chip” is quite a statement to hear from the CEO regarding his alleged $25 billion chips business.
What’s the Significance?
The manner in which they communicate this is critically important, as the figures they present serve as justification for the largest capex initiative in corporate history. On the call, Andy Jassy raised the year’s expenditures to $220 billion and revealed that backlog reached $496 billion; an increase of $132 billion in a single quarter, during the same quarter, Anthropic committed to its $100 billion-over-a-decade agreement. Amazon reported $53.4 billion in gains from its Anthropic stake this quarter, which is the primary reason “net income” increased sevenfold, while free cash flow dropped $26 billion in an unfavorable direction, and the company issued $25 billion in bonds.
Jassy himself characterized the AI demand curve on the call as “strongly barbelled”: AI labs consuming “vast amounts of compute” on one end, businesses focused on cost-avoidance on the other, and in between you find what seems durable if AI is to thrive: enterprise production workloads running inference at scale, “most of which aren’t” currently doing so. He further admitted he doesn’t know if that middle sector will follow the same “steep trajectory” as the labs have. That’s the CEO acknowledging that the demand supporting $220 billion is currently concentrated in a handful of AI labs, one of which Amazon possesses a significant stake in, while the broader wave of enterprise adoption is still a prediction. We’re all hoping for sunshine!
None of this constitutes fraud, and all of this is tangible infrastructure, but the entirety of its presentation is delivered alongside a description of a waitlist as “available.” That is what’s at stake here, and why a mention about Kiro carries more weight than Kiro itself. When the music ceaselessly ceases and the bill arrives, how will these claims appear through the clarifying lens of hindsight?
This Too Shall End
Kiro will presumably launch on iOS – months after anyone cares about it. The run rates AWS reported are likely close to accurate; they grew 37% YoY and that’s significant at their scale. Their business is performing exceptionally, and they have much to be proud of, which only makes their exaggerations more perplexing. The company that releases these kinds of figures doesn’t need to inflate software claims. But when everything’s declared “available,” then nothing truly is.
Conclusion: Amazon Earnings – A Deck of Cards?
Amazon’s recent earnings are a blend of success and stretchers. While their growth is commendable, the exaggerations regarding product availability and chip sales create a narrative as credible as a reality TV show. Until they follow through on their declarations, treat these claims with skepticism.
