Hosting for an AI website must be calculated based on the entire infrastructure stack, not just disk capacity or the number of visits. When a website calls AI models, delivers AI-generated video, processes images, or serves many concurrent users, the main costs typically include servers, GPUs or CPUs, storage, outbound bandwidth, power, and cooling systems.
The safest approach is to split the architecture into two parts: the website and the AI service. The website can run on a standard CPU server, while the AI model is accessed through an API or runs on a dedicated GPU machine. You should rent a GPU continuously only when you have stable demand, low-latency requirements, or need to control your data and models.
What is AI hosting, and why is it calculated differently from regular hosting?
What is AI hosting? It is a model for hosting and operating applications with artificial intelligence components, such as chatbots, semantic search, image recognition, content generation, speech-to-text, or RAG systems. Unlike a typical WordPress website, an AI application can generate a significant computational workload with every request.
A visit to an informational page only requires processing HTML, the database, and static files. By contrast, a question sent to a chatbot may require authentication, document retrieval, embedding generation, a model call, conversation-history storage, and the return of many tokens. If images or audio are involved, the amount of data transferred also increases substantially.
Therefore, you should monitor at least six metrics:
- The number of AI requests per minute and peak periods.
- Average processing time and high-percentile response time, such as P95.
- CPU, GPU, RAM, and VRAM utilization.
- The volume of incoming data, outgoing data, and traffic between services.
- The amount of documents, images, videos, logs, and backups.
- Cache-hit rate, error rate, and the number of requests that must be retried.
There is no fixed figure for the “cost of an AI website.” Pricing varies according to the model, server region, GPU uptime, traffic policies, and redundancy level. Cloud providers typically charge separately for virtual machines, GPUs, disks, and outbound data; Google Cloud also notes that GPU pricing does not include virtual machines, disks, or networking (according to cloud.google.com).
GPU hosting or CPU hosting: Choose based on actual workload
The question GPU hosting or CPU hosting should not be answered by choosing the most expensive hardware. Choose based on the type of work, model size, and response requirements.
| Criteria | CPU hosting | GPU hosting | Recommendation |
|---|---|---|---|
| Website, API, WordPress | Suitable | Usually excessive | Use CPU for the web layer |
| Chatbot calling an external API | Suitable | Not required | Separate the AI layer and use a model provider |
| Small-scale embedding and RAG | Sufficient for low-volume and batch workloads | Useful for fast processing or many documents | Measure processing time before upgrading |
| Fine-tuning or inference with large models | Inefficient | Requires suitable GPU and VRAM | Rent by the hour or by the session |
| Image, video, and audio generation | Unsuitable for high workloads | Virtually mandatory | Prioritize GPUs, object storage, and a CDN |
| Unstable workload | Easy to keep a baseline machine running | Easy to waste money if left on 24/7 | Use a job queue, autoscaling, or on-demand GPUs |
A GPU does not only increase the cost of renting a machine. An NVIDIA H100 SXM has a maximum design power of up to 700 W, while the H100 NVL is configured in the 350–400 W range (according to nvidia.com). This is only the GPU’s power consumption; it does not include the CPU, RAM, disks, network, power supply, or cooling.
How to estimate GPU costs
The simple formula is:
Monthly GPU cost = hourly rental price × number of GPUs × actual operating hours
For example, if one GPU costs USD 1 per hour and runs 10 hours a day for 30 days, the GPU-only cost is approximately USD 300 per month. If it runs continuously, the figure rises to approximately USD 720 per month. This is why processing queues and mechanisms for shutting down GPUs when they are idle are often more important than simply finding the lowest-priced GPU.
For uneven workloads, consider three approaches: run a GPU worker on a per-job basis, use Spot or preemptible instances if the task can be interrupted, and keep a CPU machine to receive requests and place the intensive work in a queue. AWS describes the On-Demand model as paying for usage time, while committed-use or Spot options may suit different requirements (according to aws.amazon.com).
How should data-center power, bandwidth, and infrastructure costs be calculated?
Hosting users typically do not receive a separate electricity bill, but data center electricity costs are still included in the server price. Data centers must supply power to computing, storage, networking, UPS, and cooling systems. According to the IEA, servers account for approximately 60% of electricity use in modern data centers on average; cooling can range from around 7% in highly efficient data centers to over 30% in less efficient facilities (according to iea.org).
To estimate the electricity share of a dedicated server yourself, you can use:
Monthly electricity consumption in kWh = equipment power in kW × operating hours × load factor × PUE
Illustrative example: a machine with an average power draw of 1 kW, running 24 hours a day for 30 days, at an average load of 70% and an assumed PUE of 1.4, would consume approximately 706 kWh per month. If the equivalent electricity price is USD 0.10/kWh, the estimated electricity cost would be about USD 70.60. This is only an engineering calculation; the actual hosting price also includes hardware, facility space, personnel, networking, redundancy, security, and the provider’s profit.
At the industry-wide level, the IEA forecasts that global data center electricity consumption could rise from approximately 415 TWh in 2024 to nearly 945 TWh in 2030 under its base-case scenario. This is not the bill for a single website, but it shows that AI is increasing power density and the cooling requirements of infrastructure (according to iea.org).
Bandwidth calculation formula
Total monthly bandwidth = number of downloads × average response size
For example, 200,000 page views with an average download of 2 MB per view generate approximately 400 GB of data, excluding APIs, additional images, video, reloads, and bots. If a chatbot returns PDF files, AI-generated images, or audio, the size of each response may be many times larger than plain text.
Do not look only at traffic from users to the server. You must also account for:
- Traffic from the application server to the AI model provider.
- Queries between the application, database, vector store, and cache.
- Images, videos, downloadable files, and backups.
- Logs, monitoring, and data transferred between cloud regions.
For example, Google Cloud distinguishes between inbound data, data sent to the Internet, and data transferred between regions; outbound data is typically the cost item requiring particular attention (according to cloud.google.com). AWS currently states that 100 GB of data transfer out to the Internet is free each month for EC2 customers, aggregated across multiple services and regions, but pricing conditions should be checked for the specific account and region (according to aws.amazon.com).
How to optimize bandwidth for an AI website
- Serve static files through a CDN: images, CSS, JavaScript, and video should not always pass through the origin server. A CDN can store copies closer to users and reduce the load on the origin (according to developers.cloudflare.com).
- Compress and resize images: generate images at the device’s display size, use WebP or AVIF where appropriate, and do not send a 4K original image for a small display element. Cloudflare states that image optimization can produce smaller files and reduce bandwidth (according to developers.cloudflare.com).
- Limit AI output: set limits for tokens, image dimensions, video duration, and the number of attempts. A longer answer is not automatically more useful.
- Cache safely: cache common questions, embeddings, and processed documents; do not cache personal data or content that depends on access permissions.
- Use queues for heavy tasks: video generation, batch OCR, and fine-tuning should run in the background instead of keeping an HTTP connection open.
- Track egress by feature: tag traffic for images, video, APIs, backups, and AI so you can identify which items are actually increasing the bill.
Minimum preparation requirements include a tracking table for the number of AI requests, response size, processing time, CPU/GPU utilization, storage capacity, and egress. After deployment, validate the results using the cost dashboard, CDN logs, cache-hit rate, page-load speed, and API P95 latency. If traffic increases but the cache-hit rate falls, check the query string, cookies, and Cache-Control header before upgrading the machine.
Common mistakes include renting a GPU 24/7 for a chatbot receiving only a few hundred questions per day, placing the database and GPU in two regions and thereby generating inter-region traffic, storing every generated file without a deletion policy, and sending images/videos directly from the origin. The solutions are to run GPUs on a schedule or through a queue, place frequently communicating components in the same region, set retention periods, and distribute media through a CDN.
Conclusion: A small AI website should start with CPU hosting for the web layer, an external model API, object storage, and a CDN. A website with steady AI workloads should separate GPU workers from the web server to make scaling and cost control easier. Invest in a dedicated GPU cluster only when you have sufficiently long-term usage data, clear security or latency requirements, and evidence that self-operation costs less than pay-as-you-go hosting.
References
