Amazon's Q2 was great, but the earnings release is packed with baloney

Amazon reported its earnings today, and because I am professionally depressed I read the thing in full [PDF]. "How long can I go before the red haze of rage sets in" is a fun game, and today I made it all the way to the bottom of the second page when I encountered a bullet point touting how AWS "made its spec-drive [sic] coding agent, Kiro, available on iOS." Yes, I was in the room when they announced it at the New York summit, six weeks ago. As of this writing, their website, which I have screenshotted says I can "request early access" because "We'll invite a limited number of people to try the app via Apple's TestFlight, and we'll send everyone a link when it's ready." So Kiro is "available" in the same way as I am available to play in the NBA. You can twist yourself into a pretzel and assert that this claim is technically true, but for all practical readings it's what we'd colloquially term "a lie." You need to be explicitly invited to Apple's developer beta testing tool, where a limited number of users can try out an unpublished version. You cannot download it on your phone, and there is no page in the App Store that showcases the product. The delay is almost certainly due to Apple's byzantine App Store policies, which I have some sympathy for — but this is an earnings statement. If they're going to "shade the truth" like this, what else are they not being forthcoming about? Once you notice it becomes hard to stop There are a lot of other statements that one suspects might not stand up to scrutiny. Graviton boasts "up to 30 to 40% better price-performance," which I only accept because I have seen the numbers myself on customer workloads. The express statement that their AI business and chips business are each exceeding $25B run rates in consecutive bullets, with no word on whether those dollars overlap (we will come back to this point shortly). And their Bedrock statement: "customers spent more in Q2 than all prior quarters combined," which makes it sound like a rocket until you realize that they're saying the past 90 days exceeded the other 10 quarters for which Bedrock has been available. Without actual numbers tied to these, that makes it sound like for the first couple of years Bedrock was showing up wearing a party hat but no pants. Then there's the AWS operating margin of 39.4%, which came in above every published analyst estimate and which everyone will invariably cite as cherry-picked proof the AI buildout is printing money. On the call, CFO Brian Olsavsky disclosed that it includes roughly $600 million of mark-to-market gains on energy derivative contracts. By his math, AWS margins were up 650 basis points year over year, or 520 "if you exclude the derivative accounting gain." Strip that gain out yourself (behold the power of arithmetic!) and the blowout margin goes right back inside the range analysts had modeled. Amazon now hedges electricity the way an airline hedges jet fuel, and this quarter the hedges paid off directly. Olsavsky noted these adjustments "have not been significant in prior quarters." The first quarter they are significant, they land in AWS margin, and their Q3 guidance already assumes no impact from these remeasurements going forward. Amazon knows it's noise, but clearly saw no reason to turn down claiming the win. The chips business that sells no chips Back to those dueling $25B run rates I touched on; describing their "AI chips business" that way struck me as an incredibly odd thing to say. That business has revenue, growth, a triple-digit trajectory, sarcastic numbers of happy customers — but what it doesn't have is a product that you can buy. There is no Trainium price list, they will not ship you a socketed Graviton chip to put in your next desktop build, there isn't even an external part number. What Amazon books as "chips revenue" is EC2 instance rental (possibly filtered through higher level services like Bedrock, SageMaker, the half-baked agents that fail to properly explain your AWS bill to you, etc.), and an EC2 instance is not a chip. It's the chip, plus the nVME, plus the NICs (themselves built on Nitro, which uses Amazon's own silicon), plus some aspects of the data transfer that somehow aren't directly billed, plus the building the whole mess lives in—and then with AWS's margin layered on top. The silicon itself is a minority line item in the internal bill of materials that constitutes its business. You don't have to take my word for it; Amazon CEO and AI Marketing Manager Jassy spent last quarter's call lamenting that the cost of components, "particularly memory, has skyrocketed," so by his own testimony a growing slice of the "chips business" is memory revenue. Cynically, the category exists so that headline writers will talk about it in the same breath as Nvidia's data center numbers, which they of course will. But Nvidia's $25 billion is silicon sold in the form of physical packaged chips, shoveled out their loading dock. Amazon's is fully-loaded infrastructure rental. This is a hotel comparing its revenue to a mattress company's. But wait, there's one more layer of inanity here. Olsavsky has said that the majority of Bedrock's workloads run on Trainium. So if you follow one Anthropic dollar through the earnings release it's AI-business revenue, it's chips-business revenue, and it's AWS segment revenue. It's nice when you can get a triple-brag for the same thing. You don't have to take my word on the "sells no chips" part either. On today's call, Morgan Stanley's Brian Nowak asked when Amazon might start selling Trainium to third parties. I want one too; I get it. Jassy answered that customers are increasingly interested in getting Trainium "separate from our cloud," that Amazon is "actively having those conversations," and that "there's a real chance we'll do that in the future." IN THE FUTURE. "Yeah, we have yet to sell a single chip" is quite something to hear from the CEO about his purported $25 billion chips business. Why does this matter? The way they talk about this matters deeply, because the numbers they're draped around serve as the justification for the largest capex program in corporate history. On the call, Andy Jassy raised the year's spending to $220 billion and announced backlog hit $496 billion; up $132 billion in a quarter, during the same quarter Anthropic signed its $100 billion-over-a-decade commitment. Amazon booked $53.4 billion in gains on its Anthropic stake this quarter, which is most of why "net income" septupled, while free cash flow went $26 billion in the angry direction and the company sold $25 billion in bonds. Jassy himself described the AI demand curve on the call as "very barbelled": AI labs consuming "gobs and gobs of compute" on one end, enterprises doing cost-avoidance on the other, and in the middle you've got the stuff that actually seems durable if AI is to have a future: the enterprise production workloads running inference at scale, "most of which aren't" doing so yet. He went on to admit he doesn't know whether that middle will follow the same "wildly steep trajectory" as the labs have. That's the CEO stating that the demand underwriting $220 billion is concentrated today in a handful of AI labs, one of which Amazon happens to own a meaningful piece of, while the broader enterprise adoption wave remains a forecast. We're hoping for sunshine! None of this is fraud, and all of this is real infrastructure, but the entire shape of it all is being told in the same sitting that described a waitlist as "available." That is what's at stake here, and why a bullet about Kiro matters more than Kiro itself does. When the music inevitably stops and the bill comes due, how will these statements look through the clarifying lens of hindsight? This too shall pass Kiro will presumably ship on iOS – months after folks gave the slightest toss about it. The run rates AWS said are probably close to real; they grew 37% YoY and that's no small thing at their scale. Their business is firing on all cylinders and they've got a lot to be proud of, which makes their overstating things just that much weirder. The company that posts these kinds of numbers doesn't need to inflate the software bullets. But when everything's "available," then nothing is. ®