RentServers.comSelf-Hosted LLM on a Dedicated Server, Managed for You
Private LLM hosting for teams that want an open-weight model on hardware no one else touches. We provision the server, load the model, expose an OpenAI-compatible API and run it day to day, for one fixed monthly price. Bigger than one server? Providers bid for the GPUs.
Private LLM hosting plans
Fixed monthly prices, sold and billed by Wells Software Consulting LLC. Every plan shows the part of the price that is the server lease at our cost.
Private LLM Core
$649/month
A dedicated CPU server with 128 GB or more of RAM running a quantized 7B to 14B model.
- Runs: Llama 3.1 8B, Qwen 2.5 7B and 14B, Mistral 7B and Gemma 2 9B at 4-bit or 8-bit
- Load: 1 to 3 users at once; about 10 to 20 tokens per second per user on an 8B model at 4-bit
- Best for: Internal assistants, document Q&A and batch jobs for a small team, where a few seconds of latency is fine.
In stock October 22, 2026
- Tampa, FL: 2 x E5-2683 v4 2.1GHz Intel Xeon CPU (32 cores), 128 GB RAM, 960GB SSD, 20TB / 10Gbps
- Los Angeles, CA: 7313P 3.7Ghz EPYC AMD (16 cores), 256 GB RAM, 2TB NVMe SSD + 1TB NVMe SSD + 480GB, 20TB / 10Gbps
Disclosed margin: of the $649, $114 to $270 is the server lease at our cost and $379 to $535 is our management fee and margin.
Questions first? Request this serverPrivate LLM GPU
$1,399/month
A dedicated server with an NVIDIA L4 24 GB GPU and 128 GB or more of system RAM, serving through vLLM.
- Runs: Llama 3.1 8B at FP8 or FP16, Qwen 2.5 14B at FP8, and models up to about 32B at 4-bit with shorter context
- Load: About 20 to 40 concurrent streams on an 8B model; about 25 to 35 tokens per second per stream on an 8B model at FP8
- Best for: Production chat, RAG, extraction and agents for a team or an internal product, with streaming responses.
In stock October 22, 2026
- Tampa, FL: 2 x 6148 2.4GHz Intel Gold Xeon CPU (40 cores), 512 GB RAM, 1.92TB SSD + 960.2GB SSD, NVIDIA L4 24GB, 20TB / 10Gbps
- Tampa, FL: 7302P 3.3GHz EPYC AMD (16 cores), 128 GB RAM, 480.1GB SSD + 3.84TB SSD, NVIDIA L4 24GB, 20TB / 1Gbps
- Tampa, FL: 7302P 3.3GHz EPYC AMD (16 cores), 128 GB RAM, 2TB NVMe SSD + 480.1GB SSD, NVIDIA L4 24GB, 20TB / 1Gbps
Disclosed margin: of the $1,399, $932 to $999 is the server lease at our cost and $400 to $467 is our management fee and margin.
Questions first? Request this serverPrivate LLM 70B
$7,499/month
Month-to-month. Lower on a signed term:
| 12 months | $6,400/mo |
| 24 months | $5,995/mo |
| 36 months | $4,995/mo |
A dedicated accelerator slice with 128 GB of accelerator memory in a US cloud region, serving one 70B-class model at 8-bit through vLLM.
- Runs: Llama 3.3 70B, Qwen 2.5 72B and 70B DeepSeek R1 distills at 8-bit
- Speed: Throughput is benchmarked on your own prompts during setup and shared before go-live
- Best for: Teams that need 70B quality on private hardware: regulated data, internal copilots and agents that smaller models answer poorly.
Disclosed margin: of the $7,499 month-to-month price, $7,008 is accelerator capacity at our cost and $491 is our management fee and margin. On terms: 12 months, $4,906 cost and $1,494 ours; 24 months, $4,906 cost and $1,089 ours; 36 months, $3,154 cost and $1,841 ours.
Sign a 12, 24 or 36-month order form Larger than 70B, or many replicas: enterprise inferenceMonth-to-month plans have a $499 setup fee. No setup fee on a 12-month term. Hardware stock changes daily; we confirm the exact server and delivery date before you pay.
Pay by invoice
Same plans, same prices, billed by invoice. We open your subscription and take you to a Stripe invoice you can pay in USDC from your crypto wallet (Ethereum, Solana, Polygon or Base), by card or by ACH bank transfer. Each later month arrives as an emailed invoice you pay the same way.
- Setup starts once the payment clears: card right away, USDC when it settles, and ACH in up to 4 business days. An engineer then confirms your server and delivery date within one business day.
- The invoice is due within 3 days. If we cannot deliver the server, you get a full refund to the way you paid.
Which model fits which server
Approximate memory for the model weights alone. Add room for the KV cache, which grows with context length and concurrent users.
| Model size | 4-bit | 8-bit | 16-bit | Plan |
|---|---|---|---|---|
| 7B to 9B | about 5 GB | about 9 GB | about 16 GB | Core (4-bit or 8-bit) or GPU |
| 13B to 14B | about 9 GB | about 15 GB | about 28 GB | Core (4-bit) or GPU (8-bit) |
| 27B to 32B | about 18 GB | about 33 GB | about 64 GB | GPU (4-bit, shorter context) |
| 70B to 72B | about 40 GB | about 72 GB | about 140 GB | Scale-up quote: two or more 48 to 80 GB GPUs |
What managed means
Everything between the bare server and a working private API is ours to run.
- Single-tenant server. Your model runs on a dedicated server in a US data center. No other customer shares the machine, and prompts never leave it.
- OpenAI-compatible endpoint. vLLM on GPU servers, llama.cpp on CPU servers, behind TLS with your own API keys, so existing OpenAI client code works by changing the base URL.
- Model setup and updates. We download, quantize and load the open-weight model you choose, and swap or upgrade it on request.
- Locked-down network. Firewall closed by default, access limited to your IP addresses or a private WireGuard tunnel.
- Monitoring and patching. Health checks and alerts around the clock. OS and driver updates run in the pre-authorized window, Sundays 02:00 to 06:00 UTC.
- Engineer support. Questions and changes go to the engineers who run the server, not a ticket queue.
Self-hosted LLM or an API?
An API is usually cheaper per token. A dedicated server is about privacy, a fixed bill and control. Here is a typical month at a 3:1 input to output ratio, against an open 70B model API1.
| Output tokens per day | API per month | Smallest plan that carries it |
|---|---|---|
| 1M | $125 | Private LLM Core, $649 |
| 10M | $1,248 | Private LLM GPU, $1,399 |
| 30M | $3,744 | Private LLM GPU, $1,399 |
1 Open 70B model API (Llama 3.3 70B) at $1.04 input and $1.04 output per million tokens, Together AI serverless pricing, as summarized by MorphLLM, checked October 11, 2026. The API runs a larger model than the small plans; compare on the quality you need.
Outgrowing one server? Make providers bid for the GPUs
The fixed-price plans cover one model on one server. Past these limits the request goes to our quote desk, and you choose who runs it.
| Fixed-price plan | Scale-up quote | |
|---|---|---|
| Model size | Up to about 32B parameters | 70B-class and larger (Llama 3.3 70B, Qwen 2.5 72B, DeepSeek-class) |
| Concurrency | Up to about 40 streams at once on Private LLM GPU | More than 40 streams, or a hard latency target at peak |
| Daily volume | Up to about 30M output tokens a day (up to 14B), about 11M (27B to 32B) | More than that, or traffic that has to survive a server failure |
| Footprint | One server, one US region | Several servers, failover across regions, or a region outside the US |
Counter-quotes for the GPU servers
At least three providers bid blind on your spec: they see the hardware, term and region, never your name or company. Every quote is normalized into one 36-month cost, setup fees and egress included, and you sign directly with the one you pick at the provider's own price.
Get a scale-up quoteWe run it on that hardware
Model, vLLM endpoint, autoscaling, monitoring and on-call, on the servers you chose or on our reserved H100 and H200 pools, with the margin shown.
Dedicated managed inferenceBilling, cancellation and service terms
The short version. The Private LLM Hosting terms are the full agreement.
Billing
Monthly in advance in US dollars through Stripe: card, Link or ACH bank debit at checkout, or an emailed invoice payable in USDC, by card or by ACH. Stripe emails a receipt for every payment.
Setup
$499 once on month-to-month plans. No setup fee on a 12-month term.
Cancellation
Month-to-month: cancel any time by email; service runs to the end of the month you paid for. 12-month term: billed monthly for 12 months.
Delivery and refunds
An engineer confirms your server and delivery date within one business day after your payment clears. ACH bank debits take up to 4 business days to clear, and setup starts once yours clears. If we cannot deliver it, or you decline the date we offer, we refund everything you paid.
Service level
99.5% monthly uptime target for your endpoint outside the maintenance window (Sundays 02:00 to 06:00 UTC), with credits of 10% below 99.5% and 25% below 99.0%.
Taxes
Prices exclude sales tax. Where tax applies, it is added at checkout and shown before you pay.
Seller
Wells Software Consulting LLC, 1 E Erie St, Suite 525-2098, Chicago, IL 60611. Questions: hello@rentservers.com.
Questions
Something else? Email hello@rentservers.com.
What does self-hosted LLM mean here?
You get your own server running an open-weight model such as Llama, Qwen, Mistral or Gemma, reachable through a private API. Prompts and outputs are processed only on that server and are not used to train anyone's model. We operate the server; you control the model choice and who can call it.
Is a self-hosted LLM cheaper than an API?
Usually not on price alone. Open-model APIs charge around $1 per million tokens for a 70B model and less for small models, so a dedicated server only wins on cost at high, steady volume. Teams self-host for privacy, a fixed monthly bill and control over the model. The self-hosted LLM cost calculator shows the break-even for your volume.
Which model should we run?
For most internal assistants and RAG, a current 7B to 14B instruction model is enough. The Private LLM Core tier runs these at 4-bit or 8-bit for a few users; the Private LLM GPU tier serves them to dozens of concurrent users with streaming. If you need 70B quality, we quote a multi-GPU deployment.
What if we outgrow one server?
Private LLM GPU carries about 30M output tokens a day on a model up to 14B, or about 11M on a 27B to 32B model, at the 50% planning load. Past that, or for a 70B-class model, the same request form routes you to our quote desk: at least three providers bid blind on the GPU servers, each quote is normalized to one 36-month cost, and we price running it for you if you want that. Moving between plans keeps your API keys and endpoint.
How fast is it?
About 10 to 20 tokens per second per user on an 8B model at 4-bit on Private LLM Core. About 25 to 35 tokens per second per stream on an 8B model at FP8 on Private LLM GPU. These are planning estimates; before you sign we benchmark your own prompts on the server you will get.
Can we fine-tune or use our own weights?
Yes. We load your fine-tuned or merged weights in safetensors or GGUF format. Fine-tuning runs themselves are quoted separately, because they need different hardware for a short time.
What does the price include, and what is your margin?
The monthly price covers the server lease, bandwidth, the model endpoint, monitoring, patching and engineer support. Each tier states how much of the price is the server at our cost and how much is our management fee and margin. Month-to-month plans have a $499 setup fee; there is no setup fee on a 12-month term.
Where are the servers?
In US data centers. The tier cards list the cities with matching servers in stock today. If you need a specific region, tell us and we quote it.
Can we run the server ourselves instead?
Yes. Send us your spec and we get counter-quotes for GPU or high-RAM servers from several providers, free on standard sourcing. Get counter-quotes.
Tell us the model and the load
One form for every size. If one server carries it, we confirm the server, benchmark your prompts on it and reply within one business day. If it is bigger than one server, our quote desk replies with GPU counter-quotes and a managed price.
- Nothing is charged here. You pay by card, ACH or USDC only after we confirm the server and delivery date in writing.
- Sized for you. If the plan you pick is too small for the load you describe, we say so and size it up.
- Bigger needs, blind bids. On a scale-up quote, providers see the technical spec only. Your name, email and company stay with us until you pick a winner.
Private LLM Hosting is sold, billed and supported by Wells Software Consulting LLC. Scale-up quotes for GPU servers are free referrals: you sign with the provider at its own price. How we make money. Llama is a trademark of Meta Platforms, Inc. NVIDIA and L4 are trademarks of NVIDIA Corporation. Wells Software Consulting LLC is not affiliated with these companies.