How Microsoft Is Building for a World of Metered Intelligence — Mike Taylor

A dispatch from Microsoft's annual Build conference. Mike Taylor (Every's head of tech consulting, "Also True for Humans" column) argues Microsoft is the first major player to "get real" about a world where AI intelligence is on tap but constrained by the meter. Note: the email is a free preview that cuts off at the "model optimization" section behind Every's paid paywall; this note summarizes the unlocked portion only.

Why this is in the vault

RDCO runs an always-on Claude Code agent loop that bills by the token, so the economics of metered intelligence are not abstract here, they are the operating cost structure. This piece names the inflection directly: the "$5 Uber era" of subsidized LLM subscriptions is ending, and the practical disciplines Microsoft is shipping (local models, automatic model routing, cheaper small models) are the same levers RDCO already pulls (mix-and-match cheaper models, sub-agent context isolation). It also doubles as an investing-thesis signal on AI-infra cost direction.

The core argument

• The subsidy era is closing. Taylor frames it as the end of the "$5 Uber era" for LLMs: labs have been subsidizing usage at the cost of thousands of dollars per subscription, which cannot last as Anthropic, OpenAI, and xAI head toward going public and have to face their books like Uber did before its 2019 IPO.
• The trigger event: on June 1, Microsoft switched GitHub Copilot to token-based billing, sparking outrage as some users reported bills jumping from $39 to over $3,000/month. Rather than backtrack, Microsoft used the Build stage to argue for using AI more pragmatically under rising costs.
• Nadella's framing: he promised "unmetered intelligence to every desk and every home," an AI-era echo of Bill Gates's "a computer on every desk."
• Local compute as the off-meter play: the RTX Spark laptop (built with Nvidia, shipping in the fall) runs a 128-billion-parameter model locally so developers can work without paying per token, betting that budget-conscious coders will accept being a year or two behind the frontier rather than running trillion-parameter open models that won't fit on a laptop.
• Reduce switching costs: Microsoft wants to be the place to run any model, agent, or harness, with a rebuilt smart terminal app (it even adopted Mac-ecosystem terminal commands) and a new GitHub Copilot Desktop app that makes switching between OpenAI-built, Anthropic-built, and local models easy.
• Automatic model routing: delegate simpler tasks to cheaper models. GitHub's Mario Rodriguez (CPO) cited it as the affordability answer; COO Kyle Daigle observed developers reach for "the model of the day, or week, or hour" even when the task doesn't merit it, and that "the tools could" downgrade automatically even when a person won't.
• Microsoft is eating its own cost medicine: the piece notes Microsoft has been cancelling internal Claude Code licenses to cut spend. Enterprises can't access subsidized "Max" plans and pay full freight per token; one unnamed firm is rumored to have burned roughly half a billion dollars on Claude tokens in a single month.
• More cheap supply: Mustafa Suleyman's Microsoft AI research lab released a set of new, smaller, cheaper models spanning image, voice, transcription, coding, and reasoning.

Mapping against Ray Data Co

Mapping strength: strong.

• Direct operating-cost relevance. RDCO's autonomous COO loop is exactly the "full-freight per token" enterprise profile Taylor describes (no subsidized Max plan). The model-router thesis ("not every task needs a frontier model") is already RDCO doctrine: cheaper-model mix-and-match for mascot instantiation, sub-agent fan-out for context isolation, and the standing rule not to pause for per-call cost confirmation because spend is budget-controlled, not per-call-gated.
• Validates the cost-discipline patterns RDCO has been formalizing. The "tools could route, a person won't" point reinforces routing the model-selection decision into the harness rather than leaving it to in-the-moment judgment, the same logic behind RDCO's context-rot and subagent-routing discipline.
• Investing-thesis signal. Microsoft's pricing pivot, the rumored nine-figure monthly token bills, and Suleyman's cheaper-model push are demand-direction and margin-pressure data points for the AI-infra / capital-cycle thesis (memory, power, data-center capex). Pairs with the hyperscaler-capex and data-center anchor notes.
• Actionable takeaway: worth a periodic audit of whether RDCO's own loop is defaulting to frontier models on tasks a cheaper model would clear, and whether local/open small-model inference is worth piloting for low-stakes high-volume steps.

⚠️ Sponsorship

No third-party paid sponsor. The issue carries Every house-promo: a paid-subscriber paywall gating the back half of the article and a bundled-product pitch for Every's own software (Sparkle, Cora, Spiral, Monologue). Treat the subscription/product CTAs as Every's self-interest; they do not bias the reported conference observations but are disclosed here for completeness (sponsor_entity: house-promo).

Related

• [[2026-06-01-jaya-gupta-token-budget-wars]] — token-as-billing-unit economics; the metering substrate this Microsoft pivot operationalizes
• [[2026-02-23-every-chatgpt-memory-context-rot]] — same author (Mike Taylor) and column ("Also True for Humans"); cost/efficiency-of-intelligence framing
• [[2026-05-24-every-cheap-competence-new-frontier]] — the cheaper/smaller-model thesis Microsoft's MAI releases and model routing put into practice
• [[2026-05-27-not-boring-thank-god-for-data-centers]] — AI-infra capex / capital-cycle angle the pricing-pressure signals feed