Andy Jassy is using cloud revenue to justify an AI buildout that deepens Amazon’s monopoly. It is true that Amazon reported a strong quarter — profit more than tripled year-over-year, cloud revenue accelerated, and the stock jumped nearly fifteen percent. In the narrow sense in which quarterly earnings ever prove anything, the numbers were real. The trouble is what analysts concluded from them: that Amazon’s massive artificial-intelligence investments are “beginning to pay off,” as though last quarter’s cloud growth were evidence that the tens of billions committed to AI-specific infrastructure had produced AI-specific returns. It is not what the numbers show. What Amazon reported was a cloud-computing business — a mature, dominant, extraordinarily profitable monopoly — continuing to grow on the strength of its existing market position. The AI narrative is an overlay, and a convenient one.
To be precise about what “cloud revenue” actually means, because the public discourse treats it as synonymous with AI and it is not: Amazon Web Services sells compute time, storage, and a growing menu of managed services to millions of customers. The overwhelming majority of that revenue comes from running conventional workloads — databases, websites, application backends, enterprise software — that predate the current generation of AI models and have nothing to do with large-language-model inference. AWS has been the dominant cloud provider since before ChatGPT existed. Its margins are among the highest in corporate America because it occupies what Cory Doctorow would recognize as a textbook chokepoint: once an organization has migrated its infrastructure to AWS, the switching costs are enormous, and Amazon’s pricing reflects that knowledge.
The analytical move analysts are making — cloud revenue grew, therefore AI investments are paying off — is a post-hoc attribution error dressed up as a thesis. Amazon does not break out AI-specific revenue from its cloud segment, and the reason matters: the company has no incentive to reveal what fraction of its cloud growth comes from the pre-AI workload migration that was already its core business. The overall acceleration in cloud revenue is consistent with a half-dozen explanations — enterprise digital transformation, post-pandemic infrastructure refresh, the general shift from on-premise to managed services — that were driving AWS growth well before anyone outside a research lab had heard of a transformer model. Conflating AI capex with cloud revenue is precisely the kind of loose accounting that flatters a narrative investors already want to hear.
The mechanism is a bezzle. John Kenneth Galbraith’s term, which Doctorow has usefully revived, names the interval between the commission of a financial confidence trick and its discovery — when the perpetrator feels enriched and the victim has not yet registered the loss. Amazon’s capital-expenditure commitment is real and enormous — it shows up on the balance sheet, it pulls actual labor and materials out of the economy, it raises the electrical load on grids that were not built for this. The return, in contrast, is entirely narrative-based at this stage. The company is selling tomorrow’s promised efficiency gains to justify yesterday’s spending, and the market is accepting the promise at face value because every other hyperscaler is doing the same thing. The consensus is self-reinforcing — which is exactly how a bezzle inflates. The “payoff” analysts are describing is not a return on capital in any durable sense. It is the market interpreting more spending as evidence that the spending was a good idea. That is circular, and it is the signature of a bezzle.
What the capex buildout is actually doing is something more consequential than “paying off.” Amazon is constructing the physical infrastructure layer of the AI economy — data centers at hyperscale, the submarine cables that carry the traffic (TeleGeography tracks the rising share of international capacity controlled by the four hyperscalers, with their share of total international bandwidth surging from negligible levels in 2010 to roughly 75% today), the power-purchase agreements for nuclear and renewable baseload that make the electricity bill bearable — on top of an already-dominant cloud position. This is not a company building a new product line. This is a company extending its control into every adjacent layer of the computing stack, using the AI boom as the financing narrative for an infrastructure consolidation that was underway before the word “GPT” entered the investor vocabulary.
The same attribution play showed up in Microsoft’s earnings the day before — cloud revenue growth credited to AI, massive capex increases justified by that credit — and in both cases the companies are spending to build infrastructure that, once built, makes their market positions structurally harder to challenge. The four forces that Doctorow identifies as the historical constraints on platform monopoly — competition, regulation, self-help through interoperability, and organized labor — are each weakened by vertical integration at this scale. Consider competition alone: a would-be cloud rival would need to replicate not just Amazon’s software stack but its power-purchase agreements, its submarine cable leases, its custom silicon investment — capital requirements that exceed the entire market capitalization of every cloud startup combined. That is not a competitive disadvantage. It is a structural wall. A firm that controls the cloud, the AI training infrastructure, the energy supply, and the network fabric is not merely outcompeting rivals. It is removing the preconditions for competition itself — credible entry becomes a question of whether a competitor can duplicate not just a software product but an entire energy-and-connectivity apparatus. That is a different problem than building a better search engine, and it is the problem these quarterly earnings reports keep burying under the “AI is paying off” headline.
Rising oil prices, which added to Friday’s market jitters, underscore the physical reality behind the narrative. The compute that powers large language models runs on electricity, and while Amazon claims to have matched 100 percent of its electricity consumption with renewable energy for three consecutive years, that figure reflects market-based accounting — renewable energy certificates and long-term PPAs — not the physical grid mix serving its data centers. In regions of the United States hosting more than 70 percent of Amazon’s data centers, electricity still comes primarily from natural gas or coal. Each increase in oil-driven electricity prices hits a rival considering new construction harder than it hits Amazon, whose long-term PPAs lock in below-market rates for years — the new entrant faces spot-market energy costs that track commodity prices while the incumbent’s costs are fixed. The oil-price worry that rattled bond markets on Friday is a short-term trading concern; the infrastructure lock-in it underscores is a decades-long structural one.
My father’s mill in Selkirk was bought by Gerdau in 1995. The new owners came in talking about modernization and global competitiveness — the same vocabulary the hyperscalers are using today. What actually happened was that they extracted the value of the installed workforce, laid off a third of the line, and ran the rest at higher speed until the mill could not sustain it. The talk about the future was real enough in its way; the extraction was what funded it. The vocabulary changed; the extraction logic did not. The workers whose labour was extracted saw their families’ security turned into shareholders’ yield. That is also what is happening now, at scale: the AI buildout is being financed by the extraction of surplus from the rest of Amazon’s operations — the marketplace, the logistics network, the cloud customers who have nowhere else to go — and the promise of future returns is what keeps the borrowing window open.
The bezzle will last as long as the next quarterly report supports it. When the capex comes due and the revenue growth does not materialize at the promised multiple — which is what happened to every prior cycle of technology-infrastructure overbuild — the victims will not be the analysts who wrote the bullish notes. They will be the workers whose wages were held flat, the small businesses paying higher fulfillment fees, the taxpayers whose grids are being upgraded for hyperscaler draw, and the investors who bought at the top. The utility commission in Northern Virginia — where the densest concentration of hyperscaler data centers on the continent draws on a grid Dominion Energy is already upgrading at ratepayer expense — is the live example of that future arriving. The public-comment period is still open. That deadline is the only part of this machinery that the people running the bezzle actually respect.