Microsoft, Uber Confront Real AI Costs as Subsidy Era Ends
Enterprise AI's low-cost era is ending in real purchasing decisions. Microsoft canceled internal Claude Code licenses this week because token-based billing proved untenable even with effectively unlimited cloud resources; Uber's CTO warned internally the company burned through its entire 2026 AI budget in four months. Three independent source batches corroborated this shift — not isolated churn but a structural repricing hitting every AI-heavy enterprise at once.
What the Source Actually Says
Market analyst @HedgieMarkets — whose post was amplified by Yann LeCun — frames the dynamic precisely: "The AI subsidy era is ending in real time." Microsoft, which invested $13 billion in OpenAI and built the Azure infrastructure powering most of Anthropic's compute, concluded that a competitor's coding tool wasn't worth paying for at token-level billing. That is not a productivity verdict — it is a unit-economics verdict. GitHub (also Microsoft-owned) is simultaneously dropping flat-rate plans across its products in favor of usage-based billing, and American AI software prices have jumped 20–37% across the major labs over the past six months.
The mechanism driving this is structural. @emollick identified it explicitly: complex agentic workflows consume "thousands of times more tokens" than single-turn chatbots, creating a divergence the flat-rate era never exposed. Nate Herk's 100-hour Claude Code vs. Codex comparison quantified the asymmetry in practice — Claude's output-token volume per task runs 2–5× higher than Codex on identical prompts, which is why enterprise users are hitting monthly limits faster than before. Box CEO Aaron Levie described token economics as the most heated CIO-dinner topic in the post-Opus 4.7 world; CIOs are now actively mixing model-routing strategies, per-team caps, and capability-tiered access to manage the exposure.
The market is locked between two unappealing paths. As HedgieMarkets lays out: either enterprises scale back AI usage to fit budgets — throttling the revenue ramp the labs need to justify pre-IPO valuations — or the labs cut prices and absorb losses at exactly the wrong moment for unit economics. Both trajectories "land in the same place, the numbers stop working, and somebody has to take the writedown."
Strategic Take
Flat-rate assumptions are gone. Model routing by task type — cheap workhorse models for volume, frontier models only for high-stakes work — is no longer optional cost optimization, it is baseline operating discipline. Teams building AI-dependent workflows must budget for token-based billing from day one; the subsidy era that made experimentation cheap is structurally closed.
