For a small 7B–8B quantized model, plan on roughly $5–$15 per month for a CPU VPS with 8–16 GB of RAM. A rented GPU is far more expensive when left running: commonly about $190–$475 per month from hourly rates of $0.26–$0.65. Your model size, context length, traffic and whether the machine runs continuously determine the real bill.
Antihost offers Linux VPS plans with full root access, unlimited traffic on a 1 Gbps port and a 14-day free trial.
What you are paying for
Ollama itself does not create a special VPS fee. The bill comes from the virtual machine, its memory, storage, data transfer and, if selected, GPU time. Model files are downloaded to that machine and inference consumes its RAM or VRAM. For most self-hosted installations, memory is the first limit; extra CPU cores improve concurrency and responsiveness only after the model fits.
As an Amazon Associate I earn from qualifying purchases.
Prices below are market figures reported by comparison guides and hourly rental listings rather than verified quotations from every provider. Catalogs, regions, taxes and promotions change, so confirm the current checkout price before ordering. A June 2026 pricing notice summarized for one provider conflicts sharply with older guide figures, illustrating why a published number is not a guarantee.
CPU-only VPS cost by model size
CPU inference is practical for development, private experiments and low request volumes. It avoids GPU rental and normally uses a predictable monthly plan. The trade-off is latency: one guide reports about 3–10 seconds per response for a 7B model on a modern VPS CPU, while another reports 12–18 tokens per second for Llama 3.1 8B on a particular CPU configuration. The benchmark method was not documented, so treat those speeds as indicative rather than a promise.
#1 Best Overall
- AI Performance: 767 AI TOPS
- OC mode: 2632 MHz (OC mode)/ 2602 MHz (Default mode)
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Axial-tech fan design features a smaller fan hub that facilitates longer blades and a barrier ring that increases downward air pressure
- A 2.5-slot design maximizes compatibility and cooling efficiency for superior performance in small chassis
| Model class | Memory guidance | Typical VPS choice | What the evidence suggests |
|---|---|---|---|
| 7B–8B quantized | At least 8 GB RAM; 8–16 GB leaves useful headroom | 4–8 virtual CPUs, 8–16 GB RAM | About $5–$15/month in aggregator quotes; 1–2 GB plans commonly run out of memory |
| 13B | At least 16 GB RAM | 8 or more virtual CPUs, 16–32 GB RAM | Usually costs more than an entry plan; exact current prices vary by provider and region |
| 70B, CPU inference | About 64 GB RAM | High-memory instance with substantial CPU allocation | Well beyond the inexpensive VPS tier; verify the provider’s memory and sustained-CPU policy |
The memory figures come from an Ollama library guidance page: approximately 8 GB for 7B, 16 GB for 13B and 64 GB for 70B. They are minimum-scale guidance, not a guarantee that every quantization, context window or workload will fit.
Examples of quoted flat-rate plans
Aggregator guides have quoted the following examples, but the figures conflict and should not be treated as current offers:
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
- A 4-vCPU, 8-GB plan from Hetzner at $7.49 per month in one guide and roughly €8.30 in another; an 8-vCPU, 16-GB plan at $14.49 in one guide.
- Contabo examples of $5–$8 for 8 GB, about $9 for 16 GB and about $15 for 32 GB, with guides disagreeing on the exact amounts.
- A 16-GB general cloud VPS from Vultr quoted from $24 per month.
- A 16-GB memory-optimized DigitalOcean plan quoted at $84 per month.
A June 2026 pricing notice, summarized in search results, listed a 2-vCPU, 4-GB CX23 at €5.49 (previously €3.99), an 8-vCPU, 16-GB CPX42 at €69.49 (previously €25.49), and a dedicated-vCPU CCX13 at €42.99 (previously €15.99), excluding VAT. The changes apply to new orders and resizes from 15 June 2026. Those amounts conflict with the older aggregator quotes, and the underlying primary pricing page was not opened, so use them as a warning to check the live regional price rather than as a definitive tariff.
Rank #2
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5070 Ti
- Integrated with 16GB GDDR7 256bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
When a GPU VPS is worth the premium
A GPU reduces generation latency and is important for larger models or sustained interactive traffic. GPU capacity is usually rented by the hour rather than bundled into a cheap, flat-rate VPS. Examples in the collected listings include an RTX 4090 at about $0.39 per hour, an A6000 with 48 GB at about $0.44 per hour on a secure rental and $0.26 on a community rental, and a 48-GB A40 at about $0.65 per hour. One fixed GPU plan was quoted at $90 per month for Llama 3.1 8B, but plan availability and included resources must be checked directly.
| Hourly rate | Approximate 100-hour month | Approximate 300-hour month | Continuous 730-hour month |
|---|---|---|---|
| $0.26/hour | $26 | $78 | $190 |
| $0.39/hour | $39 | $117 | $285 |
| $0.44/hour | $44 | $132 | $321 |
| $0.65/hour | $65 | $195 | $475 |
The monthly column assumes 730 hours and is arithmetic, not a provider-published monthly price. An hourly instance that is stopped outside working hours can therefore cost much less than a permanently running GPU. Check whether storage, stopped-instance fees, minimum rental periods, data transfer and a monthly spending cap are billed separately; hourly billing with a cap is common, but the cap rules differ.
VRAM requirements for larger Llama models
A third-party estimate places a 7B model’s Q4 weights at about 4 GB and a 70B model’s Q4 weights at about 40 GB. Weight size is not the whole requirement: the runtime, operating system, buffers, tokenizer and context cache consume additional memory. For Llama 3.3 70B, the same estimate puts Q4 use near 40 GB of VRAM and FP8 near 70 GB. A 24-GB card therefore cannot be assumed to run a 70B model simply because a compressed file appears smaller.
Rank #3
- Powered by the NVIDIA Blackwell architecture and DLSS 4. System Requirements: Minimum 850W PSU with 16-pin 12V-2x6 (12VHPWR) connector required. Verify before purchasing.
- Military-grade components deliver rock-solid power and longer lifespan for ultimate durability. Compatibility: 348mm (13.7") length, 3.6 slots, 4.3 lbs. Confirm case clearance and slot spacing. GPU bracket included.
- Protective PCB coating helps protect against short circuits caused by moisture, dust, or debris
- 3.6-slot design with massive fin array optimized for airflow from three Axial-tech fans
- Phase-change GPU thermal pad helps ensure optimal thermal performance and longevity, outlasting traditional thermal paste for graphics cards under heavy loads
How context length changes the bill
Long prompts and conversation history allocate a key-value cache in memory. Ollama documentation describes default context lengths that rise with available VRAM: 4,096 tokens below 24 GiB, 32,000 tokens at 24–48 GiB and 256,000 tokens at 48 GiB or more. Ollama recommends at least 64,000 tokens for agent-style tasks. Those defaults are capacity-dependent, not free performance: increasing context can consume enough memory to make an otherwise fitting model fail or force slower operation.
Budget headroom for your target context, concurrent requests and the largest prompt you will send. A VPS that barely fits model weights may work for a short single-user chat and fail when an API client submits a long document.
A practical sizing and budgeting procedure
- Choose the model and quantization. Record whether you need a 7B/8B, 13B or 70B model and whether it is Q4, FP8 or another format. Do not size from parameter count alone.
- Set a context target. Decide whether ordinary chat, document prompts or agent workflows are required. Include the context cache in the memory budget instead of treating it as an optional extra.
- Add operating headroom. Start above the stated 8-, 16- or 64-GB guidance so Linux, Ollama, caches and simultaneous requests have room. Leave additional disk space for model downloads and updates.
- Decide whether latency justifies a GPU. For occasional or development use, begin with CPU-only. For interactive production traffic, calculate expected GPU hours rather than assuming 24/7 operation.
- Compare the complete monthly total. Add the VPS or GPU rental, persistent storage, backups, public IPv4 charges, transfer fees and taxes. A low headline hourly rate can become expensive when storage and an always-on schedule are added.
- Set an automatic shutdown or cap. Stop an hourly GPU when it is idle and configure billing alerts. Verify that stopping releases compute charges and whether storage continues to accrue fees.
- Test the exact model before committing. Download the intended quantization, send representative prompts and observe memory use, first-token latency and sustained generation. A different context size or model revision can change the result.
Installing and operating Ollama on a VPS
After creating the instance, use the current Ollama installer for the VPS operating system, start the Ollama service, and pull the model you intend to run. Keep the API bound to a private interface or protect it with a reverse proxy, firewall rules and authentication; exposing an unauthenticated inference endpoint can let strangers consume your CPU, GPU and bandwidth.
Rank #4
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5060
- Integrated with 8GB GDDR7 128bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
- Choose a current 64-bit Linux image and confirm the advertised RAM, virtual CPU type, disk size and location.
- Install Ollama using its current official installation method, then verify that the service starts after reboot.
- Download a small test model first and run a short prompt. Watch system memory and swap while the model loads.
- Run the production model with the context setting you actually need, then test two or more simultaneous requests if your application will be concurrent.
- Record response latency, tokens per second, peak memory and GPU utilization. These measurements are specific to your provider, model revision and prompt pattern.
- Enable automatic security updates where appropriate, restrict inbound ports, and keep model files on persistent storage so a compute instance can be replaced without downloading everything again.
Nvidia compatibility check
Ollama’s GPU guidance states that Nvidia cards need compute capability 5.0 or newer and driver version 550 or newer. Confirm both before renting a card. A listed VRAM amount is not enough if the driver is too old or the image lacks the required runtime libraries.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsCommon failure modes and fixes
“Model requires more system memory”
For example, one reported error required 164.8 GiB while only 13.4 GiB was available on a 16-GB ARM board. The fix is to move to a larger-memory instance, use a smaller model or lower-memory quantization, reduce context length, and stop other memory-heavy services. Swap may prevent an immediate crash but generally makes generation dramatically slower; it is not a substitute for sufficient RAM.
The model loads but responses are unusably slow
Check whether the model spilled from VRAM into system RAM, whether CPU throttling is occurring, and whether context history has grown. Reduce context, lower concurrency or select a GPU with enough VRAM. For light use, accepting several seconds per response on CPU may be cheaper than paying for an idle GPU.
Best Value
- Powered by the NVIDIA Blackwell architecture and DLSS 4 OC mode: 2640MHz/Default mode: 2610MHz (Boost Clock)
- Military-grade components deliver rock-solid power and longer lifespan for ultimate durability
- Protective PCB coating helps protect against short circuits caused by moisture, dust, or debris
- 3.125-slot design with massive fin array optimized for airflow from three Axial-tech fans
- Phase-change GPU thermal pad helps ensure optimal thermal performance and longevity, outlasting traditional thermal paste for graphics cards under heavy loads
The GPU is unavailable to Ollama
Verify the driver and compute capability, confirm that the VPS image exposes the device, and check that another process has not consumed the VRAM. A marketplace or community rental can also have different reliability and isolation characteristics from a dedicated allocation.
The invoice is higher than expected
Inspect hourly runtime, attached volumes, snapshots, public addresses, egress and tax. Reconcile the provider’s billing timezone with your shutdown schedule. Do not assume that stopping a VM deletes its disk or that a quoted price includes VAT.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Which setup fits each usage pattern?
| Use case | Starting point | Cost logic | Main compromise |
|---|---|---|---|
| Learning, development or occasional personal chat | 8–16 GB CPU VPS for a 7B–8B quantized model | Flat monthly cost in the commonly quoted $5–$15 range | Several-second responses and limited concurrency |
| Small internal API used during business hours | 16 GB CPU VPS or an hourly GPU scheduled for working hours | Compare a predictable monthly VM with perhaps 100–300 GPU hours | Scheduling complexity versus latency |
| Interactive public service | GPU with enough VRAM for the model and context | Budget for sustained hourly use, monitoring and security | Continuous operation can reach roughly $190–$475 per month before extras |
| 70B experimentation | About 64 GB system RAM for CPU or roughly 40 GB VRAM at Q4 | High-memory or GPU rental; inexpensive entry VPS plans are unsuitable | Large memory footprint, slower downloads and higher failure cost |
Bottom line for your budget
If you are asking how much to set aside, use $5–$15 per month as a starting estimate for a small 7B–8B CPU deployment with 8–16 GB RAM, then verify the live plan and add storage, backup, transfer and tax. A 13B model generally moves you into the 16-GB tier, while a 70B model requires a high-memory CPU machine or a GPU with roughly 40 GB of Q4 VRAM. GPU rental is economical when usage is intermittent and the instance is stopped promptly; leaving a $0.26–$0.65-per-hour card running all month turns it into an approximately $190–$475 compute bill. Memory headroom and context length matter as much as the advertised model size, so test the exact workload before committing to a long-term VPS.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
