The ledgers and the courts moved

This week AI reset three things a business treats as fixed background: what it costs, how it shows up in the accounts, and what it is allowed to do. The largest software company began rationing the AI usage it sells others on maximizing. Paper gains on private-AI stakes inflated Big Tech's reported profits. And a US appeals court ruled that when an AI agent shops for you, the shopper reaching the store is you.

Maximum usage stopped being the point

For two years the pitch, Microsoft's most of all, was that AI value scales with usage: put an assistant in every developer's hands, let the agents run, and output climbs with the tokens consumed. Engineering leaders budgeted AI as a growth input to feed rather than a cost to cap. The harder the agents worked, the more you were winning, and buying more inference was buying more productivity. A rising meter was a sign of adoption, not a problem to manage.

On August 4 Microsoft told its own engineers to ease off. Executive vice-president Jay Parikh set division-level "AI token budget targets" to cap AI usage. He also made OpenAI's cheaper GPT-5.6 the default in Microsoft's internal coding assistant, which had been quietly routing most work to Anthropic's pricier Claude models. Microsoft's engineers, in other words, had been writing code on expensive Claude tokens billed to Microsoft. "Tokenmaxxing is not what we are optimizing for," Parikh wrote; the aim was "more impact per token." Internal guidance noted engineers were each spending hundreds to a few thousand dollars a month, per The Next Web and CNBC that week.

The uncomfortable party is anyone whose AI plan, or AI product, treats consumption as the metric that matters. When the company whose whole external pitch is that every developer should run its assistant caps its own developers and routes them to a cheaper model, the "buy all the inference" reflex is finished, even as inference keeps getting cheaper. OpenAI cut its mid-tier GPT-5.6 prices by up to 80% the week before. Cost-per-task is the number now. Meter each agentic workflow the way you already meter cloud compute, budget it per task, and read a climbing token bill as something to govern rather than proof the tool is working.

Reported profit stopped tracking the business

The settled way to read a Big Tech earnings report was that the profit line reflected the operating business: advertising, cloud, software, devices. Investors and operators used those numbers to judge who was winning the AI buildout, on the logic that rising profit meant the enormous AI spending was paying off in the core. A record quarter counted as evidence the strategy worked. The minority stakes these companies held in private AI labs were a footnote to the story, not the story itself.

The second-quarter numbers broke that. Big Tech's stakes in Anthropic and OpenAI, each now valued near $1 trillion in private markets, are marked to that valuation every quarter, and this quarter the markup swamped the operating results. Close to half of Alphabet's record quarterly profit came from revaluing its private-AI holdings rather than from selling advertising or cloud, per CNBC on August 3 and Axios on August 4. Strip those paper gains out of the US giants' results and reported second-quarter earnings growth for the S&P 500 falls from roughly 48% to under 30%. The gains are unrealized: no cash moved, and a down round in the private labs would reverse them.

The exposed party is anyone sizing the AI trade off headline tech earnings: investors indexing to reported profit, operators benchmarking rivals, boards judging their own AI spend against the giants'. Reported profit now moves with private valuations a single financing round could cut. This is the accounting cousin of the circular financing flagged when the contest left the screen: the same handful of AI valuations now sit inside a chip vendor's debt guarantees and the hyperscalers' income statements at once. Read the AI names on operating cash flow rather than net income, and know how much of any "record" quarter is a mark you cannot spend.

The agent shopping for you is legally you

The defensible read on controlling AI agents was that a platform could keep them out with the same anti-hacking law it already used against scrapers. In the US that law is the Computer Fraud and Abuse Act, the 1986 statute that makes unauthorized access to a computer a federal offence, long invoked to block automated traffic a site did not want. When an AI agent logged into a site and acted, the assumption was that the company that built the agent was the one "accessing" the site, and could be sued and blocked for it. Control the developer, control the access.

On August 4 the US Ninth Circuit Court of Appeals rejected that. It vacated an injunction Amazon had won against Perplexity's Comet, an AI browser that logs into Amazon and buys on a user's instruction. The court held that it is the user, not Perplexity, who "accesses" Amazon's servers when the agent acts on that instruction, per Engadget and Courthouse News that day. Noting there is almost no case law applying the statute to AI agents, it read the ambiguous criminal law narrowly. It is the first US federal appeals ruling on whether AI agents may act across the open web for the people directing them.

The uncomfortable party is any platform whose plan to keep third-party agents out rested on anti-hacking law, and any business whose moat is a walled garden an agent now enters carrying a customer's credentials. For a retailer, a marketplace, a booking site: if an agent acting for your customer is legally your customer, you cannot treat it as an intruder. The fight is far from over, this being one circuit on a preliminary question, but the direction is set, and Europe's rulebook moved the same week, as the EU AI Act's transparency duties became enforceable on August 2 with fines reaching 3% of worldwide turnover. Decide now whether agents are traffic you serve, gate, or charge for, because the law is drifting toward treating them as your users.

Read the three together

For a year this category argued that capable models are commoditizing and that value drains off the model into the scarce things around it. This week the breaks sat somewhere quieter, and for an operator closer to home: the systems used to pay for AI, to count it, and to govern it. Microsoft turned AI cost from a usage it maximized into a budget it rations. Big Tech's earnings showed how much reported profit is now a mark on private-AI stakes that a down round could erase. A US court decided that an AI agent acting for a user carries the user's legal identity. None of the three turned on a benchmark. The operator question for the rest of 2026 is no longer only which model you run, or even what you own around it. It is whether your ledgers and your contracts are drawn for a world where AI itself is the thing being metered, booked, and adjudicated, because this week the meter, the ledger, and the court all moved without waiting for you.