AI Hosting: Where to Run Models and Agents

AI hosting splits into five jobs: rent a GPU, run code on a GPU only when needed, call a model by the token, use a cloud AI platform, or host the app around the model. For open models at steady load, a rented H100 runs $2.69 to $4.29 an hour. At light load, per-token APIs cost far less. Prices were checked on 31 October 2026.

Five kinds of AI host

A common mistake is to compare across rows as if they were the same product. A GPU hour and a million tokens measure different things. The sections below show how to convert, with example volumes.

Some links on our site go through /go/ and may earn us a fee; it never changes our order. To filter hosts by price, region, GPU and data terms, use the interactive hosting compare.

Prices at a glance

All figures checked on 31 October 2026 on the vendors' own pages.

Notice the spread for the same GPU. An H100 is $2.69 on Runpod Community, $3.95 on Modal, $4.29 on Lambda and $5.49 on Replicate. The cheaper end means shared or less managed capacity. The higher end buys convenience, such as per-second billing or no setup.

How to choose

Start with the job.

I want to call a model from my app

Use an inference API. You pay per token and manage nothing. Examples from the pages above: Cloudflare Workers AI lists gpt-oss-120b at $0.35 per million input tokens and $0.75 per million output tokens. Groq lists the same model at $0.15 and $0.60. Pick on price, model list, speed and data terms. See Cloudflare Workers AI and Hugging Face Inference, and our Groq vs Cerebras page.

I want my own model, always on

Rent a GPU. A pod on Runpod or an instance on Lambda gives you the machine, and you run vLLM or Ollama on it. See vLLM and Ollama. It is cheaper than tokens only when the GPU stays busy.

I want my own model, only when requests arrive

Use serverless GPU. Modal bills by the second, with per-second rates from $0.000164 for a T4 to $0.001097 for an H100, and Runpod Serverless bills per second from worker start to stop. The trade is cold starts: the time to load a model into GPU memory. See Modal and Modal vs Runpod.

I already use a big cloud

A cloud AI platform keeps billing, identity and networking in one place. AWS Bedrock's documentation says model providers do not have access to its deployment accounts or to customer prompts and completions. See AWS Bedrock.

I need to host the agent app

The model is one piece. The code that receives requests, calls the model and holds state needs a host. Cloudflare Workers has a Free plan with 100,000 requests a day and a Paid plan from $5 a month, per its pricing page. Vercel's Pro is $20 a month. Render's paid instances start at $7 a month. See Vercel and Render.

Converting between units

Three conversions come up when you compare.

GPU hour to tokens. The honest answer is that it depends on the model, the card and the batch size. We do not give a tokens-per-second figure for a card, because the vendor pages do not publish one.. Measure on your own load.

Seconds to hours. Modal's H100 is $0.001097 per second. Multiply by 3,600 and you get $3.95 an hour. The advantage is that a 90-second job costs $0.10, and the machine is off the rest of the time.

Utilisation. A rented GPU at $3.99 an hour costs $2,873 in a 30-day month if left on. If it works 25 percent of the time, the real cost per busy hour is four times the sticker. Serverless and per-token pricing exist to remove that waste.

Three example workloads

The volumes are examples we chose. Use them as templates, with your own numbers.

A private chatbot for a ten-person team, 40 million input and 8 million output tokens a month. On Groq's gpt-oss-120b rates ($0.15 and $0.60) that is about $10.80. On Cloudflare's rates ($0.35 and $0.75) it is $14.00 plus $6.00, $20.00. A Runpod Secure A100 at $1.79 an hour for 220 hours is about $394. For this load, the API wins on cost by a wide margin. A GPU makes sense only if the data must stay under your control.

A nightly batch of 20,000 document summaries. Suppose the job takes 3 hours on one H100. On Modal that is 10,800 seconds at $0.001097, about $11.85. On Runpod Community at $2.69 an hour it is about $8.07, with the pod stopped afterwards. On Lambda at $4.29 it is about $12.87. Differences are small, so choose on setup effort and data terms.

A coding agent service for a team of eight, always available. A single always-on 80 GB card at $1.79 an hour is about $1,289 a month. Per-token APIs would cost less at moderate volume, so the decision again turns on privacy and on how many tokens agent loops really use.

Data and privacy by host

The terms we read, with sources in the frontmatter and on each host page:

None of that is a substitute for reading the contract. If the data is regulated, get the data processing terms in writing. For the strictest cases, our private AI hub covers running on your own hardware, and cloud vs self-hosted AI covers the trade.

Regions

Region affects latency, price and legal fit. Modal lets you set container regions and a routing region per Function and notes that broader regions improve availability and cold-start time. Runpod says it complies with GDPR for data processed in European regions. Lambda attaches filesystems to a region. Bedrock lists the regions where each model is offered. Check the region list on each provider's page before you commit data to it.

Limits to watch

• Spend caps. Runpod's default spend limit is $80 an hour across all resources, which you can raise by contacting support.
• Rate limits. Cloudflare Workers AI lists a default of 300 requests per minute for many tasks, with per-model exceptions.
• Cold starts. Modal's docs say a cold start can range from seconds to minutes, depending on how much has to load.
• Idle billing. Lambda and Runpod pods bill while running, used or not.
• Free plan limits. Render's free web services spin down after 15 minutes without traffic and take about a minute to start again.

Moving between hosts

Use the OpenAI-compatible API where you can. Groq, Cerebras, Ollama, vLLM and Cloudflare Workers AI each offer one, so app code changes only the base URL and the model name. Keep model weights and prompts in your own storage. Keep the container image portable, so a Modal function or a Runpod pod can be rebuilt elsewhere. See OpenRouter alternatives for gateways that put several providers behind one API.

Hosting an agent versus hosting a model

People often say "AI hosting" when they mean two different things. Keep them apart, because they have different bills.

The model. It needs GPUs, or an API that has them. Costs scale with tokens or GPU time. Everything on the GPU and inference pages of this site is about this.

The agent. It is ordinary software: a process that receives a task, calls a model, calls tools and stores some state. It needs CPU, memory, a queue or a database, and a network address. It rarely needs a GPU. Costs scale with requests and with how long each run lasts.

A typical agent service therefore uses two hosts. The agent's code lives on an app host, and it calls a model API or your own GPU. Long agent runs matter on the app side. Vercel's docs list a default and maximum function duration of 300 seconds on Hobby, and a default of 300 with a maximum of 800 seconds on Pro, with an extended maximum of 1800 seconds. A run that outlives the limit needs a queue, a worker or a different host. Render bills instances by the second and lists a paid instance from $7 a month. Cloudflare Workers charges for CPU time rather than wall-clock time, so a request that waits on a model costs little CPU, though the exact accounting is worth reading for your pattern.

For the build side of this, see how to build an AI agent.

Questions to ask any host

Use the same list for every provider, and write the answers down.

1. What is the unit of billing? Hour, second, token, request, seat or credit.
2. What is billed when idle? A pod left on, a dedicated endpoint, a reserved worker.
3. What does a cold start cost, in time and money? Ask for the typical range for your model size.
4. Where is the data processed, and for how long is it kept? Get this in writing for regulated data.
5. Is there a zero-retention option? And is it the default or a paid setting?
6. What are the rate and concurrency limits on my plan? And what happens at the limit: queue, throttle or error.
7. Which regions are available, and can I pin one?
8. What GPUs can I get, and how reliably? First-come capacity can run out at busy times. Lambda's page, for example, describes self-serve instances as first-come access.
9. What is the exit? Can I export weights, logs and configuration?
10. How do I cap my spend? A hard limit protects you from a runaway loop.

A provider that answers these plainly is easier to trust than one with a lower headline price and vague terms.

A note on price comparisons

Prices on these pages are list prices on one day. They change often, and some vendors show only part of the picture on a pricing page. Runpod's page lists several GPUs with only one visible rate in a plain fetch, and Bedrock's per-token rates load dynamically by model, region and tier. Where a price did not render for us, we say so and mark it. Before you budget, open the vendor's page and read the figure for your exact model, region and plan, and check the date on ours.

Also compare like with like. A Community-tier GPU on a peer-to-peer marketplace and a Secure-tier GPU in a certified data centre are not the same product. An on-demand price and a reserved price are not the same product. A shared per-token endpoint and a dedicated endpoint are not the same product. Our hosting compare lets you filter by those differences so that the rows you compare are actually comparable.

A short buying process

1. Write down the data class and where it may go.
2. Estimate monthly tokens or GPU hours with an example week of real use.
3. Price the two cheapest fits from the table, with the formulas above.
4. Run a one-day pilot on each. An hour on a Runpod RTX 4090 Community pod is $0.34.
5. Check the exit: can you move the weights, prompts and logs out?
6. Re-check prices each quarter, since this market moves fast.

Our take

Match the unit to the work. Light and spiky use belongs on per-token or per-second pricing. Constant use belongs on a GPU you keep busy. Private data belongs on a host whose terms you have read, or on your own hardware. We rank no host above another on price alone, and money from any provider does not change our order. These notes come from vendor pricing pages and documentation, and we did not run hands-on performance tests.

Related pages

• Hosting hub and interactive hosting compare
• Runpod vs Lambda, Modal vs Runpod, Runpod vs Vast.ai
• Private AI hub and how to self-host an LLM