Skip to content
S SHANE AI GUY AI. AUTOMATION. WEBSITES. MARKETING. REQUEST CONSULTATION

AI Server Earnings

What the server earns

Billable hours in, running costs out. Every default below is a measured number from the ShaneWebGuy LLM Server, not an estimate.

Concurrency is limited by memory first, not by the GPU

Every concurrent stream needs its own KV cache. At 16K context with an 8-bit cache that is about 1.5 GB per slot, on top of the 17 GB the model itself occupies.

Which explains the measured 4090 curve. 44.3 / 52.3 / 36.0 / 31.8 tok/s at 1 / 2 / 4 / 8 looks like the card giving up, but 24 GB minus the model leaves room for roughly three slots — so four and eight never fit, and throughput fell because the context was being squeezed. The aggregate 52.3 is also far below what that card's bandwidth allows, which says the same thing: memory-starved, not compute-bound.

So the thread limit below is computed from the hardware you pick, and anything past Shane's four measured points is marked modelled, never presented as measurement.

What it earns

Rate per billable thread hour$6

What one hour of one stream is worth.

Billable threads2

Running11% of the day

The hardware decides the threads

Context per stream

Each stream holds its own cache at this size. 16K now — halve it and twice as many fit.

Per month

$0Revenue
$0Cost
$0Net

At other thread counts

ThreadsAggregateRevenueCostNet

Measured inputs: 4090 throughput 44.3 / 52.3 / 36.0 / 31.8 tok/s at 1 / 2 / 4 / 8 concurrent; RunPod rate $0.89/hr; network volume $3.50/month; M5 Ultra prices as quoted. A month is 730 hours. This prices capacity, not demand — it says what the machine can earn if the hours are sold, never that they will be.

Want a tool like this for your business?

We build custom tools, automations, and websites. Book a free consultation and we’ll scope it with you.

Book a free consultation
Shane Clark

Before we chat, tell us a little about yourself.