Microsoft: AI costs more than humans; Uber blinks on spend

UNIT ECONOMICS

Microsoft reports AI is more expensive than paying human employees — internal Microsoft analysis finds running agents costs more per task than the humans they replace. The labor arbitrage thesis underpinning half the agent startup pitches of the last two years just took a body shot from the company best positioned to know.

If you're modelling automation ROI, rerun the spreadsheet with current token prices, not the ones in your deck from January.

Uber president says AI spending is getting 'harder to justify' — a second data point in the same direction: Uber is publicly reassessing AI spend because the returns aren't landing. When a company that built its entire business on algorithmic dispatch starts squinting at the AI line item, the enterprise budget conversation has shifted.

Anthropic ships Claude Opus 4.8 with dynamic workflows and cheaper fast mode — adds dynamic workflows capped at 1,000 subagents per run, plus a cheaper fast-inference tier. The 1K ceiling is a real number to design around — fan-out orchestration just got an explicit budget — and the fast-mode price cut directly changes the math for latency-sensitive agent loops.

Google Pay preps for AI agents with Universal Commerce Protocol — introduces a Universal Commerce Protocol and a server redesign explicitly for transactions initiated by agents rather than people. Google is betting agentic checkout is a near-term load, not a 2028 problem — useful primitive if you're building anything that needs to actually spend money on a user's behalf.

The internet is being rebuilt for machines — AWS, Cloudflare and friends are quietly retuning cloud infra for agent-to-agent traffic patterns rather than human browsers. Connection pooling, rate limits, bot detection, caching — all the assumptions baked in for human users start to invert when 80% of your requests are headless.

DEBATE

The AI bubble isn't like the internet bubble — Doctorow argues today's AI has measurable utility in narrow domains — code, search, moderation — so a correction looks more like overvaluation and consolidation than a 2001-style smoking crater. Pairs nicely with the Microsoft and Uber items above: real utility, real cost, real reckoning.

Just like gold and oil, we'll soon be able to trade AI token futures — exchanges are drafting futures contracts for AI tokens, treating inference output as a commodity input like electricity or bandwidth. Conceptually clean; standardization and liquidity are entirely unproven. If it works, inference-heavy products get a hedging instrument they currently lack.

'AI washing': firms scramble to rebrand as tech-focused — the 2026 version of putting ".com" in your name. Worth a skim before your next vendor evaluation — the gap between AI in the pitch deck and AI in the codebase is widening fast.

How long until AI automates all cognitive labor? — a tracker aggregating expert AGI timelines, currently spanning 2030 to 2060+. Read it as a calibration exercise for your own priors, not a roadmap — concrete capability milestones still beat vibes-based forecasting.