Cloud AI Revenue Turns Real | AI Field Notes #77

AnalysisCoding agents are moving from writing throwaway scripts to speeding up serious scientific software. OpenAI published a field report on July 29 documenting eight scientific-computing projects where its coding agents cut program runtimes, the kind of performance-tuning work that usually needs a specialist who understands both the science and the machine underneath. The report frames agents as collaborators on real research code, well beyond autocomplete. Performance optimization has been a senior, expensive skill, the sort of thing a lab hires one person to do well. If an agent can find the same speedups, that seat gets harder to justify.

AnalysisBuying software that negotiates with your suppliers just pulled in serious money. Freehand raised a $75 million Series B on July 29, led by Battery Ventures and NewRoad Capital, for autonomous agents that handle procurement and supplier contracts. Customers running the agents recovered 5% to 10% of their spend and cut cycle times more than 70%, with deployments at Meta, Unilever, and Johnson & Johnson. Sourcing has been a quiet, human-heavy function inside big companies. An agent that claws back a tenth of company spend is a direct threat to the people who used to do that work by hand.

AnalysisMeta is now on pace to spend more building AI than it earned in the quarter. On July 30 it raised the top of its 2026 capital-spending range to as much as $145 billion, most of it data centers and chips, against second-quarter revenue of $60.8 billion, up 28%. Mark Zuckerberg paired the numbers with a pitch for 'personal superintelligence,' AI built for individuals rather than institutions. The revenue is healthy and growing. The spending is growing faster, and the return on tens of billions in compute is still a promise investors keep funding on the strength of the story.

AnalysisThe clearest sign yet that AI spending is turning into AI revenue landed in the cloud results. Amazon reported second-quarter numbers on July 30 showing AWS at $42.2 billion, up 37% year over year, its fastest growth in eighteen quarters, with chief executive Andy Jassy saying the AI and custom-chip businesses each passed a $25 billion annual run rate. A day later Microsoft's Azure crossed $100 billion in annual revenue for the first time, growing 43%. For years the open question was whether all the AI buildout would ever pay for itself. The demand side finally has real numbers behind it.

AnalysisDemand for Nvidia's China-legal AI chip is running roughly three times ahead of supply. Chinese firms ordered more than two million H200 processors in 2026 while Nvidia held about 700,000 in stock, pushing the company into emergency talks with TSMC (the Taiwanese foundry that makes the chips) about restarting older Hopper-generation production. Washington has approved roughly ten Chinese buyers, including Alibaba and ByteDance, for up to 75,000 chips each, yet a Commerce official told Congress on July 14 that actual shipments stay 'trivial.' The choke point is advanced packaging and high-bandwidth memory, the stacked chips that feed data to the processor fast enough.