$220 Billion Is Still Not Enough. What Amazon and Meta's Q2 Earnings Reveal About the AI Economy.

$220 billion is still not enough. That is what Amazon said, in different words, when it reported Q2 2026 earnings on July 30. AWS grew 37% year-over-year — its fastest rate since the quarter ending December 2021. Management stated that capacity would not meet all 2026 demand, and described demand in 2027 and 2028 as "substantial." Amazon then raised its full-year capex guidance from $200B to $220B, citing higher memory costs.

Meta reported Q2 results the same day and narrowed its full-year capex range upward, to $130–145B. Total annual expenses are now projected at $165–169B. Between the two companies alone, that is $350–365B in annual infrastructure commitment — a floor that didn't exist in this form three years ago. AWS's AI business and its chips business each crossed $25B in annualized revenue run rates, each growing at triple digits. The AWS backlog reached $496B, growing triple digits year-over-year. These are not rounding errors in large companies' financials. They are the primary growth vectors.

The conventional reading of hyperscaler capex at this scale confirms that AI infrastructure is the dominant capital destination of the decade. That is accurate, but incomplete. The more important insight for application builders is about competition. When AWS's backlog exceeds $496B and is still growing triple digits, it faces ongoing pressure to compete for developer loyalty — not just enterprise contracts, but the broader developer ecosystem that generates future enterprise demand. That competition translates into better pricing, more managed services, faster model access, and more aggressive developer programs over time.

The arithmetic of declining per-token costs is already visible. Enterprise AI spend grew significantly while per-token costs fell more than 80% over 18 months. That combination — higher spend, lower unit cost — describes an industry where total usage is compounding faster than anyone publicly modeled two years ago. The $220B capex is an attempt to catch supply up to demand that has already outpaced prior forecasts. It will almost certainly underestimate what materializes in 2027 and 2028.

For founders building AI-native applications in Brazil and LatAm, the hyperscaler earnings season is the clearest available signal of the infrastructure tailwind behind their work. They don't fund the capex. They inherit the economics. As compute costs continue falling and managed AI services expand, the barrier to building a sophisticated AI application in São Paulo converges toward the barrier to building the same in San Francisco. The cost of credit scoring an unbanked consumer at Pix BNPL velocity, the cost of running an agentic compliance workflow across an Open Finance data feed, the cost of serving 100 million users with AI-powered financial guidance — all of it falling, without any action required from the company deploying the capability.

There is a tension worth holding. The same capex surge that benefits application builders also validates that AI generates real economic value at scale — real enterprises are paying real money for AWS AI services now running at $25B annualized. The infrastructure floor is rising. The question the earnings season doesn't answer — and the one that actually determines who wins at the application layer — is what you build on top of infrastructure when the floor price approaches zero.