How many people can share one AI server? Named users, active users and simultaneous requests are three different numbers.
Short answer. There is no single number, because it depends on how many people are waiting for an answer at the same moment, not how many have an account. A server can hold as many conversations as its spare memory allows, since each needs its own KV cache, and it can write for several at once by batching them, which raises total output while each person’s share falls. As a worked example, 96 GB of graphics memory holding a 4-bit 27B model leaves room for about 33 conversations of 32,768 tokens on Qwen3.8-27B, though speed will usually limit you first. ANVE has not published concurrency figures.
01
Three different counts
Named users are accounts. Active users have sent a request recently. Simultaneous requests are the ones being answered at this instant, and only those compete for the machine. ANVE’s launch tiers are described in named users (1 to 3 on One, up to 25 on Office, up to 100 on Office Pro, 100 or more on Estate), which counts accounts, not simultaneous requests.
How many of your people will be generating at once is a fact about your office, not about the hardware. The model-call log is the way to find out: ANVE logs every call with the user, app and model, and any serving software can count requests per minute.
02
Two limits: memory and speed
Memory first. Each conversation keeps a KV cache that grows with its length (2 × layers × key/value heads × head size × bytes, per token, in NVIDIA’s formula). The vLLM paper found that older serving systems used only 20.4 to 38.2 per cent of their cache memory for actual tokens, and that throughput is memory-bound, which is why modern servers manage the cache in blocks (PagedAttention) and batch requests continuously. With Ollama the documentation says a default of 1 parallel request per model, and that raising it multiplies the context memory.
Speed second. When several conversations are batched, the weights are read once per step for all of them, so total output rises; but each conversation also adds cache to read, so each person’s answer slows as the batch grows. Past a point the machine is simply busy and requests wait in a queue; Ollama queues up to 512 by default, then refuses further requests.
03
A worked example: memory only
These are upper limits from memory alone. Weights use parameters × bits per weight ÷ 8 (Q4_K_M about 4.8, Q8_0 about 8.5). Qwen3.8-27B costs 0.69 GB per 8,192-token conversation and 2.30 GB per 32,768-token one, including about 0.15 GB of fixed state, my estimate from its configuration. Qwen3-32B costs 2.15 GB and 8.59 GB. Speed, not memory, will usually bind first.
Memory budget and model
Left after weights and 2 GB overhead
Conversations at 8,192 tokens
Conversations at 32,768 tokens
24 GB card, Qwen3.8-27B at Q4_K_M
5.8 GB
8
2
48 GB of a 64 GB machine, Qwen3.8-27B at Q8_0
17.3 GB
25
7
96 GB of a 128 GB machine, Qwen3.8-27B at Q8_0
65.3 GB
95
28
96 GB of a 128 GB machine, Qwen3-32B at Q4_K_M
74.3 GB
34
8
04
What to measure before you buy
Peak simultaneous requests in a busy minute, not the daily average.
Typical and longest prompt: a one-line question and a 100-page PDF are very different loads.
How long a wait people accept: a few seconds for the first word, or minutes for a batch job such as minutes or a long report.
Which jobs can run overnight or in the background instead of competing with live chat.
A paid pilot is the way to measure this on your own workload; the fee is agreed in writing before installation.
05
Questions people ask
How many people can share an ANVE One?
ANVE One is specified for 1 to 3 named users, ANVE Office for up to 25 and ANVE Office Pro for up to 100. Those are accounts. ANVE has not published how many requests each tier handles at once, because no sealed unit has been benchmarked.
What happens when too many people ask at once?
Requests wait in a queue and each answer takes longer. Ollama queues up to 512 requests by default and then returns an error, so a serving setup should show people that their request is waiting.
Would a faster graphics card let more people share the server?
Often, because higher memory bandwidth lets the server read the weights and caches faster for each batch. The card’s memory still limits how many conversations it can hold, so capacity and speed both matter.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
