Infrastructure

AI and cloud GPUs for e-commerce

GPU pricing, private LLMs, image models and agents on your own servers.

10 guides10 in-depth83 min of reading

Running models yourself instead of calling a hosted API comes down to two questions: will the model fit, and will the GPU be busy? Fit is about VRAM. An 8B-parameter model needs about 6 GB at 4-bit quantization and roughly 16 GB or more at 16-bit, so the model and precision decide the smallest card you can use. Utilization decides the bill. A GPU rented by the hour costs the same whether it serves one request or thousands, so batch jobs such as rewriting catalog descriptions or classifying reviews suit per-second billing that stops when the job ends. A service that answers customers all day needs a machine that stays up, with the networking and regions of a normal cloud.

RunPod vs Vultr follows that split. RunPod is usually cheaper per GPU hour, bills per second and offers the widest range of consumer and data center cards, which fits experiments and batch inference. Vultr costs more per hour but gives full VMs, private networking and 30+ locations for production services. Deploying a private LLM covers serving. Use Ollama for simple single-user setups, and vLLM for high-throughput serving with an OpenAI-compatible API. In both cases, keep the model's port closed and reach it through an SSH tunnel or an authenticated reverse proxy. An open inference port is the most common mistake on a rented GPU.

Most store automation doesn't need a GPU at all. Self-hosted n8n on a small VPS runs workflows with Docker Compose, Postgres and HTTPS through Caddy on about 2 GB of RAM. It also covers what must be backed up (the database and the encryption key) and what the licence allows (internal business use, not reselling n8n as a hosted service). The older guides in this topic go further: fine-tuning with QLoRA, self-hosting FLUX and ComfyUI for product images, and running agent frameworks on a Linux server. Read them with the same two questions in mind: what has to fit in memory, and how many hours a day the hardware will be doing paid work.

Start here

AI & cloud GPUs7 min read

RunPod vs Vultr for running AI models (2026): hourly GPU cost, hardware availability and infrastructure compared

RunPod is usually cheaper per GPU hour, bills per second and has the widest range of consumer and data center GPUs, which makes it the better fit for experiments, batch jobs and inference on open models. Vultr costs more per hour for comparable cards but gives you full cloud VMs, private networking and 30+ global locations, which suits production services that need a conventional cloud setup.

AI & cloud GPUs10 min read

How to deploy a private LLM (Llama 3.1, Qwen) on a cloud GPU with Ollama or vLLM: VRAM sizing, setup and lockdown

Size the GPU from the model's config.json, not from its name: weights take parameters times bytes per parameter (about 15 GiB for Llama 3.1 8B in BF16, about 35 GiB for a 70B model in 4-bit AWQ), and every concurrent sequence adds KV cache (128 KiB per token for Llama 3.1 8B, so 1 GiB per 8k-token conversation). Use Ollama for one user and quick tests, and vLLM for concurrent or batch traffic through an OpenAI-compatible API. Bind the server to 127.0.0.1, reach it through an SSH tunnel or an authenticated proxy, and stop the GPU when the queue is empty.

AI & cloud GPUs8 min read

Self-hosted n8n on a VPS: setup, security, backups, daily health reports and private AI models

Run n8n with Docker Compose on a Linux VPS (2 GB RAM for light use), backed by Postgres and served over HTTPS through Caddy. Back up the database and the encryption key, turn on 2FA, keep port 5678 closed, and pin the image version. Then watch it: a daily report from n8n's public API that flags workflows with a high error rate, slow runs, executions stuck in running and active workflows that didn't run at all catches the silent failures that cost orders. The license allows internal business use, not reselling n8n itself as a hosted service.

Common questions

How much GPU memory do I need to run an LLM?

Roughly the parameter count times the bytes per parameter, plus room for the KV cache. An 8B model needs about 16 GB at 16-bit and around 5 GB at 4-bit quantization; a 70B model needs about 40 GB or more even at 4-bit. Longer context windows grow the KV cache, so leave headroom.

How to deploy a private LLM (Llama 3.1, Qwen) on a cloud GPU with Ollama or vLLM: VRAM sizing, setup and lockdown

Ollama or vLLM for serving a model?

Ollama is the easy option for one user or a small team: it runs quantized models with a simple API. vLLM is built for throughput under many concurrent requests, using continuous batching and PagedAttention on NVIDIA GPUs. Start with Ollama for prototypes and move to vLLM when many people or services hit the model at once.

How to deploy a private LLM (Llama 3.1, Qwen) on a cloud GPU with Ollama or vLLM: VRAM sizing, setup and lockdown

RunPod, Vultr or DigitalOcean for cloud GPUs?

Specialist GPU clouds such as RunPod usually list lower hourly prices and more GPU models. General clouds such as Vultr and DigitalOcean cost more per hour but include networking, block storage and the rest of a full platform. Compare the hourly price for the exact GPU and VRAM, what storage costs while the machine is stopped, and egress fees.

RunPod vs Vultr for running AI models (2026): hourly GPU cost, hardware availability and infrastructure compared

Is self-hosting an LLM cheaper than paying for an API?

Only when the GPU stays busy. An hourly GPU costs the same idle or loaded, while an API charges per token, so low or bursty traffic is usually cheaper through an API. Self-hosting wins on steady high volume, on data that must stay on your own servers, or when you need a fine-tuned model.

Cloud GPU pricing compared: RunPod, Vultr, DigitalOcean, Lambda and AWS, and how to turn an hourly rate into cost per million tokens

All ai & cloud gpus guides

AI & cloud GPUs8 min read

Hosting private AI models for e-commerce clients: tenant isolation, licenses, cost allocation and SLAs

A private model service is managed infrastructure with obligations, not a product you set up once. Check every license first: the model's (Llama's community license and acceptable-use policy, Apache 2.0 for many Qwen and Mistral models) and the tools' (n8n's Sustainable Use License doesn't allow selling hosted n8n). Isolate clients with separate API keys, separate logs and, for sensitive data, separate instances; per-client LoRA adapters on one base model are cheaper but share the GPU. Log tokens and latency per key, split each month's GPU bill by usage, and compare every client's revenue and p95 latency with its plan. Price from that measured cost plus your operating time, not from what the market seems to bear.

AI & cloud GPUs8 min read

Cloud GPU pricing compared: RunPod, Vultr, DigitalOcean, Lambda and AWS, and how to turn an hourly rate into cost per million tokens

On 23 October 2026 an L40S (48 GB) cost $0.79 to $1.09 an hour on RunPod, $1.57 on DigitalOcean, $1.671 on Vultr and $1.861 on demand on AWS (g6e.xlarge); an H100 SXM cost $2.69 to $3.99 on RunPod, $2.99 per GPU in Vultr's 8-GPU server, $3.99 to $4.29 on Lambda and $6.88 per GPU on demand on AWS (p5.48xlarge). The hourly rate is only half the comparison: billing granularity, storage while stopped, minimum terms and above all the tokens per second you actually get decide the price per job. Rent the candidate GPU for an hour, run the benchmark script in this guide at your real concurrency, and divide.

AI & cloud GPUs10 min read

Fine-tuning an open model with LoRA or QLoRA on your support tickets: data prep, GPU memory, spot checkpoints and evaluation

Fine-tune for tone, format and recurring answers; keep facts that change (prices, policies, stock) in retrieval. Export resolved tickets, mask emails, phones, card and order numbers, drop duplicates and one-line replies, and split train and eval by ticket. LoRA on all seven projections of Llama 3.1 8B at rank 16 trains 41.9 million parameters; with QLoRA the 8B base sits in 4-bit, so a 24 GB GPU is usually enough at 2k-token examples, while 70B needs an 80 GB card. Checkpoint every few minutes so a reclaimed spot instance resumes, compute only the loss on the assistant's tokens, and compare eval loss and blind-graded answers before and after.

AI & cloud GPUs8 min read

Running CrewAI or LangGraph agents on a Linux VPS: budgets, approval gates, idempotent actions and process supervision

Treat an unattended agent as a program that spends money and can write to your systems. Give every run hard limits on steps, tokens, dollars and wall-clock time, stop it when it calls the same tool with the same arguments over and over, and make every tool that changes something (prices, listings, emails, refunds) wait for a person to approve that exact call. Record approved actions so a restarted run can't repeat them, log every model and tool call to a file, and run the agent from a systemd timer with a dedicated user, memory and CPU limits, secrets in a root-only file and a heartbeat to an external monitor.

AI & cloud GPUs9 min read

Product lifestyle photos with FLUX or SDXL in ComfyUI on a cloud GPU: VRAM, licenses, batch automation and keeping the product real

Run ComfyUI on a rented 24 to 48 GB GPU, bound to localhost and reached through an SSH tunnel, because it has no login of its own. Generate only the background: cut the product out with a transparent background, let the model repaint the transparent area, then paste the original product pixels back on top so the item buyers receive is exactly the item they saw. Check the model license first (FLUX.1 [schnell] is Apache 2.0, FLUX.1 [dev] weights are under a non-commercial license), install as few custom nodes as possible, and drive batches through ComfyUI's API with a script that records the source photo, prompt, seed and workflow hash for every image.

AI & cloud GPUs8 min read

Renting vs buying a GPU for AI work: a cost model for on-demand, spot, reserved and owned hardware

Price every option per useful GPU-hour at your real monthly usage. An owned RTX 4090 workstation mostly costs depreciation, so it's expensive at 40 hours a month and cheap at 400. On-demand cloud scales linearly with use; spot is cheaper per hour but loses about half a checkpoint interval of work plus the restart time at every interruption; reserved capacity costs the same whether you use it or not, so it only pays at high, steady utilization. With the October 2026 prices and the example assumptions in this guide, the workstation beat RunPod's Secure Cloud RTX 4090 from about 76 hours a month but needed 234 hours a month to beat its Community Cloud price. Run the model with your own numbers before buying anything.

AI & cloud GPUs7 min read

Batch-processing product reviews and listings with an open model on a rented GPU: throughput, cost per 1,000 rows and resumable runs

Serve an 8B-class model with vLLM on one rented GPU, send it many requests in parallel (vLLM batches them on the GPU), ask for JSON with fixed keys, and write each result to a JSONL file the moment it's done so a stopped or reclaimed instance resumes where it left off. Validate every answer and retry bad JSON once. Your cost per 1,000 rows is the hourly price divided by rows per hour, so measure rows per hour on 200 real rows first and compare it with an API's per-token price at your volume before committing.