Pick the isolation controls a multi-tenant LLM API must enforce

The keyword that sorts this question is **per-tenant**. Real isolation gives each customer its own limits and its own private data, never a shared pool. The three correct controls all do that. Per-tenant rate limits stop one customer's burst from eating all the shared capacity. Per-tenant cost caps with real-time accounting stop a runaway customer from running up an unbounded bill and let you bill each one accurately. Tenant-scoped context, caches, and logs keep one customer's data from ever appearing in another's response. The two wrong options look helpful but break isolation. A single shared semantic cache keyed only on query text can hand one tenant another tenant's cached answer — a data leak. A single global rate limit caps total traffic but isolates nobody: one noisy customer can consume the whole limit and starve everyone else. Whenever an option lumps all tenants together, it isn't isolation.

Multi-tenant isolation means each customer is protected from every other, which requires controls scoped to the individual tenant — that's the test for every option. The three correct picks cover the three threats. **Per-tenant rate limits** prevent a noisy neighbor from exhausting shared GPU capacity. Per-tenant cost caps and token budgets, accounted in real time, prevent unbounded spend and enable accurate chargeback. Tenant-scoped context, caches, and logs prevent confidential data from leaking across customers. The two distractors are not just suboptimal — they actively break isolation. A semantic cache keyed only on query text ignores tenant identity, so a similar question from tenant B can return tenant A's cached answer, leaking data. A single global rate limit caps the aggregate but distributes nothing. One customer can consume the entire budget, starving the rest — exactly the noisy-neighbor failure you're trying to prevent. The pattern to internalize: an aggregate or shared control might cap totals or save cost, but it does not isolate one customer from another. Isolation lives at the per-tenant boundary, on every axis — capacity, cost, and data.