What is a Concurrent User?
In this pricing model, the term user refers to concurrent access, not the number of actual people using the service.
For example, a single real user could run 3 AI chats at the same time. Those 3 simultaneous generations would count as 3 concurrent users. If your server has a capacity of 2 users, two generations can run at the same time. The third request will be queued until one of the first two generations finishes. This means you can have several real users sharing a 2-user server—or even a 1-user server. If more requests arrive than the server can process concurrently, they are placed in a queue based on arrival time and processed as capacity becomes available.

A note about tools such as VS Code

AI harnesses and tools such as VS Code may make multiple requests to the endpoint from a single user prompt.
For example, if two real users (humans or agents) share a 1-user endpoint, requests generated by their tools may become interleaved. From the users' perspective, both may appear to be working with the AI at the same time, but the server is actually processing one generation at a time and queuing the others. The main tradeoff is latency: as more users and requests share the same concurrency capacity, requests may spend more time waiting in the queue. So, if increased latency is acceptable for your workload, a server with capacity for 1 concurrent user can still serve multiple real users.
"Concurrent users" = concurrent in-flight generations (max-num-seqs), not unique humans or agents. You get a dedicated server: zero data retention, auto-stop, per-second billing.

* Starter SWA (1M ctx) uses TNT-SWA attention — a 90K sliding window: identical to Qwen's full attention up to 90K tokens, then a small quality tradeoff in exchange for a giant, affordable context.
All other servers use Qwen's original full attention (YaRN-scaled on Blackwell) — full quality at every context size.