Amazon Says AWS Revenue Rose 39% in Third Quarter

Washington has spent years auditioning for the role of Amazon’s probation officer. The quarter just reported should embarrass the audition.

Amazon Web Services did $45.98 billion in revenue in three months, up 39 percent from a year earlier — an acceleration, and the fastest growth since late 2021. Operating income rose 55 percent, to $17.77 billion. Chief Executive Andy Jassy says capacity is largely sold out for next year and heavily committed through 2028. The company raised its 2026 capital-spending target to $220 billion.

Sit with that last figure. It is not a pilot program. It is not a strategy memo from a deputy assistant secretary. It is private capital, risked by people who have to earn a return, at a scale most cabinet departments could not appropriate without a national panic. China is building power plants and data halls as instruments of state. For the moment we are answering with customers. That is a better answer than the industrial-policy romance now fashionable in both parties — but only if we stop treating the builders as the defendant class.

The fashionable objection is scale. AWS is big, therefore AWS is a problem. The quarter argues the opposite. Scale is the product. Training runs, inference, storage, and the unglamorous work of keeping other companies’ software alive all reward the firm that already has the buildings, the chips, and the network. Jassy says ordinary computing and storage, not just the glossy AI line, grew at their fastest pace in years. A division that is nearly a quarter of Amazon’s revenue and throwing off margins like these is not a charity the retail business subsidizes. It is the profit engine. Breaking it to satisfy a seminar on market structure would be strategic self-harm in a consumer-protection costume.

None of this means a cloud outage is a private matter, or that a household ratepayer should quietly finance someone else’s substation. Cost allocation is a legitimate fight. Regulators who socialize the wires and privatize the upside deserve to be thrown out. That is a different argument from the one that says the firm writing the checks should be dismantled, delayed, or lectured into a European pace. Pricing power in a sold-out market is the signal that summons the next $220 billion. Microsoft, Google, and Oracle are not hypothetical rivals invented for a footnote. They are why the monopoly brief keeps failing its audit.

The binding constraint is not a shortage of theories. It is a shortage of electrons and permission. You cannot regulate a sold-out order book into existence. Interconnection queues, multi-year reviews, and a political class that still treats reliable power as a moral failing are the industrial policy we actually have — a policy of delay. The hyperscalers are hunting nuclear restarts, gas plants, and long-term power contracts because the grid they are told to “transition” cannot carry the load they have already sold. States that can permit a plant and a substation will get the jobs, the tax base, and the strategic capacity. States that can only produce an environmental impact statement will get a press release about their values.

A certain kind of conservative hears “Amazon” and reaches for the culture-war indictment. Some of that indictment was earned. These firms spent years playing internal politics, policing speech, and lecturing the customers who pay them. Discipline that. Stop the viewpoint discrimination. Do not confuse a real grievance about corporate politics with a plan to handicap the compute layer of the American economy while Beijing does not share the grievance.

There is a matching fraud on the left. Abundance has been rediscovered just in time to collide with the permitting constituencies, the antitrust instinct, and the suspicion of any firm large enough to matter. They want the AI future, the green future, the national-strategy future, and the breakup future, simultaneously. This quarter is what it looks like when those futures cannot all be true. You do not get capacity committed through 2028 in a country that cannot decide whether a power plant is allowed to exist.

Trainium, which Amazon says it will offer outside its own facilities beginning in late 2027, is a smaller tell and a useful one. No committee designated a national training-chip champion. A customer with a brutal cost and power problem started building an alternative to a supplier whose margins invited one. That is the American habit worth defending: rivals, capital, and time.

The country that has the data centers has the models, the military applications, the productivity gains, and the leverage. Those decisions are being made now, in chip contracts and substations, not in the next white paper. Washington can slow that down. It has gotten very good at slowing things down. It cannot replace it. The Right’s job here is not to cheer a ticker. It is to clear the physical path — power, permits, export controls that actually bite — and to refuse both populist resentment and progressive suspicion the chance to dismantle an advantage we are visibly compounding. The order book is full. The politics are not.