LLM ROI Desk

Price three ways to run the same assistant.

These three options are three ways to host the same open-source model in private. You are choosing who owns the machines, who can see the prompts, and who keeps the servers running.

DeepSeek and Qwen stay open weights in every column. None of these is a call to a public model API.

Self-hosted

You buy the GPUs and run the model on your own network. Prompts and logs stay where you put them. You pay up front, and you own drivers, capacity, and the night shift.

Cloud GPU rent

You rent GPUs from Vultr, DigitalOcean, or Azure and still run the stack yourself. The bill is those GPUs left on for the hours in The work. It is not a price per token. Isolation follows that cloud account and region.

Azure managed compute

The same open weights, on GPU VMs that Azure runs for you inside your subscription. You pay for the VMs while the endpoint is on. Azure patches the host. You still bring the model and watch quality.

Tokens / month

256B

Base workload. Scenarios rescale it.

The work

How much the assistant is used. One request is one turn: a question, a tool call, or a drafted reply. Monthly tokens = users × turns per person per day × tokens per turn × workdays. Tokens written are the part the assistant generates. Peak speed and the GPU count follow that, not the whole document. Peak factor is the morning rush versus a normal moment: if 2 tasks are usually in flight and 6 hit at once, the factor is 3. Hours per month turns the monthly pile of tokens into an average speed. 200 hours is a work month, about 20 days × 10 hours, not a clock that runs all night.

Example. 500 employees × 100 assistant turns a day × 256,000 tokens × 20 workdays is 256 billion tokens in a month. The bill uses the cards that hold the weights. With no model selected, that is 1 card.

Active users People using the assistant this month. Requests / user / day Turns each person starts on a workday. Peak concurrency factor How much busier the rush is than a normal moment. 3 means three times as busy.

Days in month Workdays counted this month. 20 is a normal month. Hours / month Turns monthly tokens into an average speed. 200 ≈ 20 workdays × 10 hours. Use 730 only if the assistant runs all day, every day.

Tokens / request A small piece of text. About three quarters of an English word. Tokens written What the assistant writes. The GPU count follows this.

Monthly benefit

What this workload is worth each month, in HKD. Use money you earn, money you stop spending, or a number you assign because the work matters. The same benefit is compared with every hosting option.

Example. The assistant gives 10 analysts half an hour back each workday. At HK$450 an hour over 20 workdays, that is 10 × 0.5 × 450 × 20 = HK$45,000 a month.

Benefit B (HKD / month) Revenue, cost avoided, or the value you assign to this workload.

Model & capacity

Pick the smallest open model that is good enough for the task, then choose who runs the machines. The table is a quality screen. p95 latency means 95 out of 100 replies finish within that many seconds. The card count is the larger of the weights and the busy moment.

Example. A busy moment needs 250 tokens per second and one GPU does 100, so you need 3 GPUs. If replies must finish within 5 seconds for 95% of users, set p95 latency to 5.

Which open model

Card count is the larger of two numbers. Memory is the weights plus cache for replies in flight. Speed is the busy moment divided by tokens per second per GPU. A leaderboard’s tokens per second are a hosted API, not this GPU. On your own card, decode speed follows active parameters and how fast that card’s memory is.

Intelligence	Active	INT4 memory	Use when
34	27B	18 GB	Start here. One card, if this score is good enough.
40	6B active	108 GB	About two H100s. High score, and only 6B parameters are read per token.
—	13B active	170 GB	The smaller DeepSeek. About 170 GB. MIT license.
40	16B active	331 GB	Newer DeepSeek flash. Fast on hosted APIs. About 331 GB, so more cards than V4-Flash.
36	49B active	960 GB	Top DeepSeek. A full node, when the smaller model fails the task.
40	95B active	1,400 GB	Same high score as Flash-Next, but 95B active, so each token is slower. A full node.

Intelligence: a public score of how good the answers are. Higher is better on that test. Your own documents can still disagree. Active: how much of the model is read for each reply. A smaller number is usually faster. B means billion. INT4 memory: a compact way to store the model file, so it fits on fewer cards. It is a bit less precise than the full file.

Intelligence is the Artificial Analysis index on 18 Oct 2026, rounded. DeepSeek-V4-Flash has no index on that list. Memory is this desk’s INT4 rule. The tokens-per-second field is a planning guess.

GLM-5.3 leads that open-weight list at about 45, with 40B active inside 753B. That is a full GPU node. This desk does not price it. Read the license before you treat it as the production default.

Words on this page

Token: A small piece of text. About three quarters of an English word.
Weights: The model file. The cards have to be big enough to hold it.
Open weights: You can download the model file and run it on your own machines.
p95 latency (seconds): 95 out of 100 replies finish within this many seconds. The slowest 5 can take longer.

Model to self-host The same list as the table. Custom clears the memory floor so you can type your own speed. p95 latency (seconds) 95 out of 100 replies finish within this many seconds. The slowest 5 can take longer. Tokens / sec / GPU A planning guess, not an NVIDIA benchmark. With no model selected, GPU count follows tokens/sec. With a model selected, the count will not go below the cards needed to hold the weights. Type your own speed after you measure.

Self-hosted LLM

You buy this card and run the model yourself.

Hong Kong card Sets the purchase. The self-host bill is this card’s shop price times the cards that hold the model.

CapEx (HKD) The purchase price of the machines. Lifetime (years) How many years you spread that purchase over. Discount rate Annual rate. 0.08 is 8%. Annual OpEx (HKD) What it costs each year to keep them running: power, people, and licenses.

Cards to hold the weights

The model file. The cards have to be big enough to hold it.

Model	Cards	Cards only
DeepSeek-V4-Flash — 284B, INT4, about 170 GB	4	HK$1,036,000
DeepSeek-V4.1-Flash — 552B, 16B active, INT4, about 331 GB	8	HK$2,072,000
DeepSeek-V4-Pro — 1.6T, INT4, about 960 GB	16	HK$4,144,000
Qwen3.8-27B — INT4, about 18 GB	1	HK$259,000
Qwen3.8-Flash-Next — 125B, 6B active, INT4, about 108 GB	2	HK$518,000
Qwen3.8-2.4T — 95B active, INT4, about 1,400 GB	32	HK$8,288,000

Cloud rent GPU and Azure managed compute

One card for both cloud options. The cloud you pick sets the rent. Azure managed compute uses that same card on the Azure VM price.

Cloud card Sets both the rent and the Azure bill.

Which cloud Only clouds that list the cloud card. Price is on-demand, at HK$7.8 per US$1. Confirm the live meter before you buy.

Cards to hold the weights

The model file. The cards have to be big enough to hold it.

Model	Cards	Rent / month
DeepSeek-V4-Flash — 284B, INT4, about 170 GB	4	HK$14,352
DeepSeek-V4.1-Flash — 552B, 16B active, INT4, about 331 GB	8	HK$28,704
DeepSeek-V4-Pro — 1.6T, INT4, about 960 GB	16	HK$57,408
Qwen3.8-27B — INT4, about 18 GB	1	HK$3,588
Qwen3.8-Flash-Next — 125B, 6B active, INT4, about 108 GB	2	HK$7,176
Qwen3.8-2.4T — 95B active, INT4, about 1,400 GB	32	HK$114,816

Weights are INT4. Count is memory ÷ that card, rounded up to 1, 2, 4, 8…. The left amount is the Hong Kong shop price times that count. The right amount is that cloud’s hourly rate times the hours in The work. Power, a chassis, and a long context need more.

Three bills

These three monthly amounts are calculated from the two setups above. Self-host is the Hong Kong purchase spread over its life, plus the yearly running cost. Cloud rent and Azure managed compute both use the cloud card, left on for the hours in The work.

Self-hosted LLM

The Hong Kong purchase above, spread over its life, plus the yearly running cost.

HK$14,375

This bill uses 1 H100.

Cloud rent GPU

The cloud card and the cloud above, left on for the hours in The work. You still run the software.

HK$3,588

This bill uses 1 H100.

Azure managed compute

The same cloud card, on the Azure VM price, left on for those hours.

HK$14,155

This bill uses 1 H100.

Monthly economics

The scoreboard for your base workload. Each column is one way to run the same model. Monthly net is benefit minus the full monthly bill. Payback is different: it asks how many months of benefit, after cash running costs only, it takes to earn the hardware back. Renting has no big purchase, so payback does not apply. The highlighted column is the cheapest option until you click a summary card.

Example. Benefit HK$45,000 and the Vultr bill of HK$3,588 leaves HK$41,412, a 1,154% ROI. Payback uses the purchase on the form: servers of HK$259,000, benefit HK$45,000, and HK$6,000 of monthly running cost pay back in 259,000 ÷ (45,000 − 6,000) ≈ 6.6 months. The purchase spread across the years is already in the monthly bill, so payback does not subtract it again.

Live monthly cost, ROI, and payback for the three hosting options
Metric	Self-hosted LLM	Cloud rent GPU Cheapest	Azure managed compute
Monthly cost	HK$14,375	HK$3,588	HK$14,155
Usage-linked The part of the bill that grows when people use it more. These three options do not have one.	In the fixed bundle	HK$0	HK$0
Fixed monthly The part you pay even if nobody sends a question that month.	HK$14,375	HK$3,588	HK$14,155
All-in HKD / 1M tokens The monthly bill divided by how many millions of text-pieces you typed above.	HK$0.06	HK$0.01	HK$0.06
Monthly benefit	HK$45,000	HK$45,000	HK$45,000
Monthly net Benefit minus the monthly bill. What is left.	HK$30,625	HK$41,412	HK$30,845
Monthly ROI Money left over, divided by the monthly bill.	213%	1,154% Best	218%
Payback How many months of leftover money it takes to earn the purchase back.	6.6 mo	n/a	n/a

Base workload is 256B tokens/month, about 868 peak tokens/sec. The busy moment would need 4 GPUs. No model is selected, so each bill uses 1 card. The GPU count above is the busy moment, not the bill. Busy-moment GPU-hours: 323,232. A planning guess, not an NVIDIA benchmark. With no model selected, GPU count follows tokens/sec. With a model selected, the count will not go below the cards needed to hold the weights. Type your own speed after you measure. Self-host payback is CapEx ÷ (benefit − annual OpEx / 12). Cloud rent and Azure managed compute have no upfront CapEx.

Volume scenarios

Same card price, different amounts of use. Try a quiet pilot, a normal month, a busy month, and a long document of 256,000 tokens a turn. Each row pays for enough cards to hold the model and to keep up in that busy moment, so a heavier month costs more.

Capacity and monthly cost by workload scenario
Scenario	Tokens/mo	Peak tok/s	GPUs	GPU-hours	Self-host	Cloud GPU	Azure
Low	20.5M	36	1	200	HK$14,375	HK$3,588	HK$14,155
Med	4B	1,389	7	1,400	HK$100,625	HK$25,116	HK$99,085
High	3.1B	5,333	25	5,000	HK$359,376	HK$89,700	HK$353,875
Long document	256B	2,604	12	2,400	HK$172,501	HK$43,056	HK$169,860

Tokens: a small piece of text, about three quarters of an English word. Written: how many of those tokens the assistant generates. Peak tok/s follows Written, spread over a day and raised by the peak factor. It does not treat the whole document as text the model must write. GPUs: how many cards that writing speed needs, and at least enough to hold the model file. GPU-hours: those cards times the hours they stay on. The teal figure is the cheapest bill in that row.

What the numbers leave out

A lower bill is not always the right choice. This table is words, not math: where the data lives, who gets called at 2am, and how hard it is to switch later.

Qualitative comparison of self-hosting, renting GPUs, and Azure managed compute
Dimension	Self-hosted LLM	Cloud rent GPU	Azure managed compute
Where the data sits	The model file, what people type, and the logs stay on your network. You decide how long to keep them.	The traffic and files sit in a rented account. How separate that account is depends on that cloud.	The model file runs in your Azure account. You choose the region. Azure runs the computer.
When the money leaves	You pay for the machines up front, spread the cost over several years, and still pay to run them each year.	No machine purchase. You pay for the rented cards while they are switched on.	No machine purchase. You pay for the rented computers while the service is switched on.
How cost moves with volume	Mostly a fixed bill. The same machines serving more people makes each piece of text cheaper.	Mostly a fixed bill. More text on the same rented cards does not raise it.	Mostly a fixed bill. More text on the same rented computers does not raise it.
If usage suddenly grows	You can only use the cards you already bought. A busy morning needs spare cards you paid for earlier.	You can rent more or fewer cards. What you avoid is paying for machines that sit idle.	You can switch computers on or off. Each one is billed the whole time it stays on.
Who is on call	Drivers, scheduling, spare capacity, and the night shift are yours.	The machines belong to the cloud. You still keep the software running.	Azure patches the computer and keeps the service up. You still supply the model and watch answer quality.
How steady the speed is	You choose the hardware and how many replies go through at once. The occasional slow reply is yours to fix.	The hardware is good. Unless you reserve it, you share it with other customers.	You choose the computer size. The occasional slow reply is still yours to fix.
Rules about where data may live	The simplest story when files cannot leave a named building.	Depends on the region, whether data may leave the country, and whether that card is allowed.	Region and data residency follow the Azure account. A private network link is extra.
How hard it is to leave	The machines still have resale value. The model file is yours only if the license says so.	Stop when the invoice stops making sense. Little money is stuck in hardware.	Switch the service off and the computer bill stops. The model file is yours only if the license says so.