Key concepts

"The model" has at least three parts: a family (Sonnet, gpt-6.1-sol), a snapshot ID (the exact version), and the host or deployment that serves it. On Azure you call a deployment name you created, and on Bedrock the same Claude model has its own ID and its own retirement date. Open-weight models add the precision and serving stack, which is why OpenRouter lets you filter hosts by quantization. The billing identity is a tree of organization, workspace or project, and API key, and rate limits attach to the organization (Anthropic, OpenAI) or the project (Google), per model, so teams in one organization share its headroom.

The law names its own entities. Under the EU AI Act the "provider" places a model on the market and the "deployer" uses it, and a company that fine-tunes with more than a third of the original training compute becomes the provider of the result. An IP indemnity keys on plan, service and model together, so one API key can be covered for one model and not another. US export rules now follow a company's ultimate parent, wherever its subsidiary is registered (guidance of May 31, 2026).

A request goes from admitted to queued, prefilling, decoding and streaming, and ends complete, incomplete or failed. A model goes from preview to generally available, legacy, deprecated (still works, retirement date set) and retired. A prompt cache is written, stays warm while used, and dies when it expires or when the prefix, tools or reasoning settings change. An account climbs tiers with spend history and pauses at its spend cap until the next month. Capacity goes from contracted to built, energized and allocated.

Retirement clocks differ by vendor and are getting shorter at some:

The same model can retire on different dates at the lab and at each cloud. Products retire too: OpenAI ends its Assistants API, hosted evals and self-serve fine-tuning jobs between August 2026 and January 2027.

On cloud marketplaces Anthropic converts usage into consumption units of $0.01 and reports them to AWS or Azure, so a discount shows up as fewer units. A private lab's "run-rate" is a recent month times twelve. Your own gateway log is the only record of errors from your side, and it's the evidence an SLA claim needs.

Access is rationed by rules the provider writes and can change:

The top model gets the tightest limit, and Anthropic calls its limits "maximum allowed usage, not guaranteed minimums". New organizations start lower "to prevent fraud and abuse", and fast ramps hit acceleration limits. Gateways add the customer's own policy (route by price or speed, cap the price per token, use only zero-retention hosts, set budgets per team), and the labs' usage policies restrict high-stakes uses without human review.

Dates I'd keep on a calendar as of October 2026:

Some dates float. Retirement dates are "not sooner than" (Anthropic, AWS) or "earliest possible" (Google), and introductory prices can end or be made permanent: Anthropic kept Sonnet 5's and cancelled a rise planned for September 1, 2026. DeepSeek even prices by time of day.

OpenAI's Chat Completions and Responses formats are the de facto interface: Bedrock exposes them ("change the base URL and API key"), most inference providers do, and Anthropic runs an OpenAI-compatible endpoint next to its own Messages API. That makes switching mechanically cheap. Everything around the call is unstandardized: usage fields, rate-limit headers, deprecation metadata and reasoning controls differ per provider. Tokenizers differ too, so the same text can be about 30-35% more tokens on Anthropic's newest models than on its older ones.

Measurement is its own layer. Artificial Analysis measures time to the first token and output speed from one server over 72 hours, MLPerf audits hardware results that vendors submit, and SemiAnalysis's InferenceX turns benchmarks into cost per million tokens. Model makers have started checking hosts: Moonshot publishes a Kimi Vendor Verifier and OpenAI added a compatibility test for gpt-oss. Below the token the units change again: GPU-hours for renting chips, gigawatts of power capacity for compute deals, and megawatt-days in power markets.

Money flows down the chain and equity flows back up it: Nvidia invests in OpenAI and in GPU clouds and providers, Amazon and Google in Anthropic, Microsoft owns about 27% of OpenAI and has invested in Anthropic, and AMD gave OpenAI warrants for up to 160 million of its shares. Amazon sells Claude on Bedrock, holds Anthropic notes, supplies it Trainium and sells its own Nova models, all at once.

The same model reached through the lab, Bedrock or Vertex can come with different retention, residency, SLA and indemnity terms, so the route is a legal choice too. For most builders the law arrives through vendor terms and procurement questionnaires.

Reversals run both ways. OpenAI's "$1.4 trillion" of compute ambitions became about $600 billion through 2030 in reporting by February 2026, and the expansion of the Abilene site beyond 1.2 GW was dropped. Export rules went from ban to unban to a revenue share that was never codified, all within 2025.

Builders carry price and capacity risk, labs carry volume risk, GPU clouds carry financing and obsolescence risk, and household electricity customers carry part of the grid cost. Liability caps in the labs' and clouds' terms are typically 12 months of fees.