cloudgpu.app

# CloudGPU

> CloudGPU (cloudgpu.app) rents GPUs billed by the minute and runs an OpenAI-compatible AI API gateway, both paid from one prepaid USD balance. Payment methods are listed on https://cloudgpu.app/payments.

Official addresses: the only CloudGPU website is cloudgpu.app, and the only API address is https://cloudgpu.app/v1. (CloudGPU 官网只有 cloudgpu.app，API 地址只有 https://cloudgpu.app/v1。)

Last updated: 2026-11-10 (Asia/Shanghai). This file has no prices; live prices are on https://cloudgpu.app/pricing, https://cloudgpu.app/api and https://cloudgpu.app/gpu-prices.json.

## Facts at a glance

- What it is: an independent GPU rental marketplace plus an OpenAI-compatible model API, on one prepaid balance. Machines come from partner suppliers and independent hosts.
- Who it is for: developers. Eligibility rules are in https://cloudgpu.app/terms and https://cloudgpu.app/aup.
- GPU billing: per minute, rounded up, 1-minute minimum. One hour is pre-authorised while an instance runs; unused minutes go back to your balance when you destroy the instance. No subscription, no minimum term.
- Storage: destroying an instance, including when its balance runs out, erases its local disk. Eligible funded accounts can explicitly save project files in Workspace independently of the GPU, restore a saved version in a later Job, or apply it to an existing running GPU. Storage quotas, retention and any charges are returned by authenticated GET /v1/storage. No automatic disk backup. When GET /v1/public/compute-capabilities advertises automatic_output_collection=true, Jobs can declare explicit outputs under $CLOUDGPU_OUTPUT_DIR; check output_collection after completion. Otherwise save or download results before destruction. Existing storage quota and charges apply. See https://cloudgpu.app/workspace and https://cloudgpu.app/docs/mcp.
- Instance types: every instance runs at the listed on-demand price.
- Sign-up: email or Google account; no phone number or ID needed.
- Model API: OpenAI-compatible endpoint at https://cloudgpu.app/v1 (chat completions, models, images, audio, video). Claude is also reachable through the native Claude Messages API format at https://cloudgpu.app/v1/messages and /v1/messages/count_tokens.
- Claude pricing: per-token rates for every Claude model are on https://cloudgpu.app/api. CloudGPU is an independent service with no partnership, authorization or endorsement from the model vendor.
- Other model pricing: per-token rates for every model are on https://cloudgpu.app/api.
- Model list: GET https://cloudgpu.app/v1/models (no key needed) and https://cloudgpu.app/api.
- Prompt caching: cached input is charged at the rate shown in the /api table when the upstream cache fires.
- Rate limits (public-beta defaults): 100 requests per minute per API key, 10,000 requests per day per account; can be raised on request. The GPU instance API (/v1/gpu/*) is open to every account with the same API key; how many instances can run at once depends on the account's top-up history (https://cloudgpu.app/api/docs#gpu).
- Reliability: the gateway has no automatic failover at the moment; no SLA. Prompt and completion bodies are not stored; token counts, model, response time and error codes are kept for billing.
- Not allowed: any proof-of-work computation, including cryptocurrency mining, and illegal activity (instance destroyed, account suspended, remaining balance forfeited).

## For AI agents and coding assistants

One API key gives an agent both a model API and GPUs, paid from the same prepaid balance.

- Models: OpenAI-compatible base_url https://cloudgpu.app/v1 with `Authorization: Bearer $CLOUDGPU_API_KEY`. The model list, GET https://cloudgpu.app/v1/models, needs no key. Claude Messages format at /v1/messages.
- GPUs over REST: GET /v1/gpu/account (balance), GET /v1/gpu/offers (live price per GPU-hour, availability), GET /v1/gpu/templates, POST /v1/gpu/instances, GET /v1/gpu/instances/{id}, DELETE /v1/gpu/instances/{id}. POST accepts auto-shutdown limits, max_minutes, max_spend_usd and idle_minutes, which CloudGPU's servers enforce even after the client disconnects. An optional setup_script (bash, up to 16 KB) runs once as root after the instance is handed over; output in /root/cloudgpu-setup.log. gpu_model "auto" picks an available GPU that can run the request, choosing by price. Reference: https://cloudgpu.app/api/docs#gpu
- MCP server for Claude Code, Cursor, Cline, opencode, VS Code and other MCP clients. It quotes, rents, runs commands and jobs, uploads, downloads, shows logs and shuts down, and can get its own key with a spending limit through a connect link: `npx -y cloudgpu-mcp` with env CLOUDGPU_API_KEY (or the `connect_account` tool). Setup: https://cloudgpu.app/docs/mcp. Source (MIT): https://github.com/mrhy7176/cloudgpu-mcp
- OpenAPI 3.1 description of the API: https://cloudgpu.app/openapi.json · MCP server card: https://cloudgpu.app/.well-known/mcp/server-card.json
- Agent rules (quote first, set a budget, download before shutdown): https://cloudgpu.app/mcp/AGENTS.md
- Continue a saved project with an agent: https://cloudgpu.app/examples/agent-project-reuse.md . Recheck Workspace version and expiry, obtain a fresh quote and approval for every new task, save results explicitly and verify release. Optional tools require both live capabilities and the installed client tool list.
- API keys: https://cloudgpu.app/api (sign-up by email or Google). Top-up: https://cloudgpu.app/billing

## What it is for

- Renting an RTX 3090 / 4090 / 5090, RTX 4080 SUPER 32G, RTX 5090D, RTX 4090 48G, RTX 4090D 48G, RTX 6000D (83 GB), A100, A800, H20, H800, RTX PRO 6000 or RTX PRO 6000 Max-Q by the minute for inference, fine-tuning (LoRA / QLoRA), ComfyUI, Stable Diffusion or LLaMA-Factory, without a monthly contract or a card.
- Getting a public HTTPS API endpoint for an Ollama or ComfyUI template: the template's API port is tunnelled to https://<random-id>-i.cloudgpu.app while the instance runs. The tunnel accepts only that instance's hostname, is capped at 3 MB/s and adds no authentication. SSH and JupyterLab stay on the supplier's own address. The ComfyUI template does not come with FLUX weights; download the models you need over SSH or JupyterLab.
- Calling DeepSeek, GLM, Kimi, MiniMax, gpt-oss, FLUX, Whisper and Claude through one OpenAI-compatible endpoint, and Claude Code through the Messages endpoint.
- Paying for GPUs and model APIs from one prepaid balance (methods on https://cloudgpu.app/payments).

## Key pages

- [Full text for answer engines](https://cloudgpu.app/llms-full.txt): this file plus the full text of every guide.
- [Home](https://cloudgpu.app/): on-demand GPUs billed by the minute plus one OpenAI-compatible model API, on one prepaid balance.
- [Rent a GPU](https://cloudgpu.app/deploy): choose a card and a template (Ollama, ComfyUI, JupyterLab + Ollama, LLaMA-Factory, or a raw GPU with SSH).
- [GPU pricing](https://cloudgpu.app/pricing): current hourly list price for every card (RTX 5090, RTX 4090 24G and 48G, A100 80G and more), billed per minute, plus per-token model rates and payment methods.
- [GPU prices as JSON](https://cloudgpu.app/gpu-prices.json): every card with its hourly price and VRAM, with an updated_at timestamp.
- [GPU compute for AI agents](https://cloudgpu.app/agents): discover GPUs, get a task quote, rent with a server-enforced budget, run code and shut down, through REST or MCP.
- [Agent integration guide](https://cloudgpu.app/docs/agents): every endpoint an agent needs, with fields, errors, rate limits and safety rules.
- [Agent skill](https://cloudgpu.app/skills/cloudgpu-compute/SKILL.md): a SKILL.md for agents that load skill files.
- [Public GPU catalog for agents](https://cloudgpu.app/v1/public/gpu-catalog): GPU models, VRAM, live prices, availability, billing rules and system disk options as JSON. No API key needed.
- Quote API: `POST https://cloudgpu.app/v1/quotes` with VRAM, GPU count, template, disk size, duration and budget returns cost estimates per GPU. No API key needed, creates nothing. See https://cloudgpu.app/openapi.json.
- Jobs API: `POST https://cloudgpu.app/v1/jobs` with a bash command, `max_minutes`, `max_spend_usd` and an `Idempotency-Key` header. CloudGPU starts a GPU, runs the command and destroys the GPU; retrying with the same key never starts a second GPU. Status, exit code, output tail, history and cost at `/v1/jobs/{id}`. Docs: https://cloudgpu.app/docs/agents
- Agent connect: `POST https://cloudgpu.app/v1/agent/connect` (no key) returns a link for the account owner to approve the agent with a spending limit and a number of GPUs at once; the agent then collects its own key from `/v1/agent/connect/token`. The owner can revoke it at https://cloudgpu.app/api.
- Start an agent session: read https://cloudgpu.app/skills/cloudgpu-compute/SKILL.md and use a tool-enabled HTTPS agent or configured MCP client. No key is required to start the connect flow. The current agent must create the owner approval link; do not ask the owner to paste credentials. A static link is not universal AI-client installation, and approval does not authorize unrelated spending.
- Web Workspace handoff: upload files and save a project version in Workspace; choose it when renting, then confirm application on the running instance. Workspace is not a whole-disk backup. Agent work is controlled through REST/MCP, not a separate Job web console.
- [Per-minute billing explained](https://cloudgpu.app/gpu-rental-per-minute-billing): rules, formula and worked examples.
- [Rent an RTX 5090 by the minute](https://cloudgpu.app/rent-rtx-5090): 32 GB Blackwell card with root SSH, billed per minute.
- [Rent an RTX 4090 48G by the minute](https://cloudgpu.app/rent-rtx-4090-48gb): 48 GB on one card, billed per minute.
- [Rent an NVIDIA A100 80G by the minute](https://cloudgpu.app/rent-a100-80gb): 80 GB with root SSH, billed per minute.
- [Rent a cloud GPU for vLLM, billed by the minute](https://cloudgpu.app/rent-gpu-for-vllm): a raw GPU with root SSH to run vLLM yourself.
- [Cloud GPU for AI film and video production](https://cloudgpu.app/cloud-gpu-for-ai-video-production): which card for each step of an AI video project (Wan 2.2, MiniMax H3, LoRAs, voice, music, upscaling), billed per minute.
- [Cloud GPU for AI short drama series](https://cloudgpu.app/cloud-gpu-for-ai-short-drama): consistent characters with a LoRA, Wan 2.2 shots in ComfyUI, voices and lip sync, with measured cost per minute of finished video.
- [RTX 5090 cloud GPU benchmark](https://cloudgpu.app/guides/rtx-5090-cloud-gpu-benchmark): published RTX 5090 vs RTX 4090 numbers for SDXL, FLUX.1 and LLM inference with sources, plus measured Wan 2.2 video times on an RTX 4090.
- [RTX 4090 vs RTX 5090 vs A100: specs and VRAM compared](https://cloudgpu.app/compare/rtx-4090-vs-5090-vs-a100): architecture, VRAM, FP16 TFLOPS and memory bandwidth per card, and which cards to check in each VRAM class.
- [What to check before you rent a cloud GPU](https://cloudgpu.app/compare): VRAM, per-minute billing, payment methods and what happens to data when an instance is destroyed.
- [How to pay](https://cloudgpu.app/payments): methods, minimums, refunds.
- [AI API gateway](https://cloudgpu.app/api): one OpenAI-compatible endpoint for chat, image, speech and transcription models, the current per-token rate for every model, and an FAQ.
- [API docs](https://cloudgpu.app/api/docs): quick start in cURL, Python and Node.js; endpoints; rate limits; GPU instance API.
- [MCP server](https://cloudgpu.app/docs/mcp): let Cursor, Cline, opencode or Claude Code rent a GPU, run a job, download the results and shut it down.
- [Docs and FAQ](https://cloudgpu.app/docs): billing, instances, templates, support.
- [About](https://cloudgpu.app/about) · [Terms](https://cloudgpu.app/terms) · [Acceptable use](https://cloudgpu.app/aup) · [Privacy](https://cloudgpu.app/privacy) · [Refunds](https://cloudgpu.app/refunds)

## AI API gateway

One OpenAI-compatible endpoint at https://cloudgpu.app/v1, one key, one balance. Change base_url in the OpenAI SDK, Cursor, Cline, Continue or similar tools. Claude Code setup is in https://cloudgpu.app/api/docs. Chat, image, video, speech and transcription models are listed at GET https://cloudgpu.app/v1/models (video: POST /v1/videos, then poll GET /v1/videos/{id} and download GET /v1/videos/{id}/content); per-token rates are on https://cloudgpu.app/api.

## Payment

One prepaid USD balance pays for GPUs and the model API. Payment methods, minimums and limits are on https://cloudgpu.app/payments; the refund policy is at https://cloudgpu.app/refunds.

## Guides

- [Per-minute GPU billing: rules, formula, examples](https://cloudgpu.app/gpu-rental-per-minute-billing)
- [Rent a GPU for LoRA and QLoRA fine-tuning: VRAM and cost per run](https://cloudgpu.app/rent-gpu-for-lora-fine-tuning)
- [Rent a GPU for ComfyUI](https://cloudgpu.app/rent-gpu-for-comfyui)
- [RTX 4090 vs RTX 5090 vs A100: specs and VRAM compared](https://cloudgpu.app/compare/rtx-4090-vs-5090-vs-a100)
- [How context length adds to VRAM use](https://cloudgpu.app/blog/kv-cache-vram-context-length-calculator-2026)
- [Save your work before you destroy a rented GPU](https://cloudgpu.app/blog/save-work-before-destroying-cloud-gpu-2026)
- [Self-hosting an LLM vs paying per token: break-even](https://cloudgpu.app/blog/self-host-llm-vs-api-break-even-2026)
- [How much VRAM to run each open model (2026 table)](https://cloudgpu.app/blog/how-much-vram-to-run-llm-2026)
- [Run Ollama on a cloud GPU with a public HTTPS API](https://cloudgpu.app/blog/ollama-cloud-gpu-public-api-2026)
- [Run ComfyUI and FLUX.1 on a rented GPU (walk-through)](https://cloudgpu.app/blog/comfyui-flux-cloud-gpu-public-endpoint-2026)
- [Fine-tune with LLaMA-Factory on a rented GPU](https://cloudgpu.app/blog/llama-factory-fine-tune-rent-gpu-by-the-minute-2026)
- [Rent an RTX 5090 by the minute](https://cloudgpu.app/blog/rent-rtx-5090-cloud-by-the-minute-2026)
- [Rent an RTX 4090 48GB by the minute](https://cloudgpu.app/blog/rent-rtx-4090-48gb-cloud-run-70b-2026)
- [Pay for cloud GPUs and AI APIs with Alipay, Google Pay, Apple Pay, a card or USDT](https://cloudgpu.app/blog/pay-for-gpu-cloud-alipay-google-pay-apple-pay-2026)
- [Run Claude Code on the gateway's Messages endpoint](https://cloudgpu.app/blog/claude-code-custom-api-endpoint-2026)
- [Use Claude in Cursor, Cline and Continue with your own key](https://cloudgpu.app/blog/claude-in-cursor-cline-continue-setup-2026)
- [Claude through an OpenAI-compatible endpoint](https://cloudgpu.app/blog/claude-api-proxy-openai-compatible-endpoint-2026)
- [Pay for the Claude API with USDT or Alipay](https://cloudgpu.app/blog/pay-for-claude-api-usdt-alipay-2026)
- [Get Claude API access and pay with Alipay](https://cloudgpu.app/blog/claude-api-alipay-no-us-card-2026)
- [DeepSeek V4 API with USDT and the OpenAI SDK](https://cloudgpu.app/blog/deepseek-v4-api-usdt-openai-compatible-2026)
- [How DeepSeek V4 API billing works](https://cloudgpu.app/blog/deepseek-v4-api-price-comparison-2026): per-token billing for Flash and Pro, cached input, and how to work out the cost of a job; current rates on https://cloudgpu.app/api.
- [Use DeepSeek in Cursor, Continue and Cline, paid with USDT or a card](https://cloudgpu.app/blog/use-deepseek-in-cursor-without-us-card-2026)
- [GLM-5 API through an OpenAI-compatible endpoint](https://cloudgpu.app/blog/glm-5-api-openai-compatible-no-chinese-phone-2026)
- [Kimi K2 and K3 API through an OpenAI-compatible endpoint](https://cloudgpu.app/blog/kimi-k2-api-openai-compatible-no-chinese-account-2026)
- [Run Nemotron 3.5 Transcribe on a cloud GPU](https://cloudgpu.app/blog/run-nemotron-3-5-transcribe-cloud-gpu-2026)

## 中文(面向中文开发者)

- [GPU 按分钟计费怎么算](https://cloudgpu.app/zh/gpu-rental-per-minute-billing)
- [租 GPU 做 LoRA / QLoRA 微调](https://cloudgpu.app/zh/rent-gpu-for-lora-fine-tuning)
- [租 GPU 跑 ComfyUI](https://cloudgpu.app/zh/rent-gpu-for-comfyui)
- [4090、5090、A100 怎么选](https://cloudgpu.app/zh/compare/rtx-4090-vs-5090-vs-a100)
- [上下文长度与显存](https://cloudgpu.app/zh/blog/zh-kv-huancun-xiancun-shangxiawen-2026)
- [销毁 GPU 前保存数据](https://cloudgpu.app/zh/blog/zh-xiaohui-gpu-qian-baocun-shuju-2026)
- [自建模型还是调 API:盈亏平衡](https://cloudgpu.app/zh/blog/zh-zijian-moxing-vs-api-chengben-2026)
- [按分钟租 4090 48G](https://cloudgpu.app/zh/blog/zh-zu-4090-48g-pao-70b-2026)
- [32B 模型怎么估算显存](https://cloudgpu.app/zh/blog/zh-32b-moxing-xuan-4090-5090-2026)
- [按分钟租 4090/5090 跑 Ollama 并拿到公网 HTTPS API](https://cloudgpu.app/zh/blog/zh-ollama-yun-gpu-gongwang-api-2026)
- [按分钟租 GPU:怎么选卡、怎么付款](https://cloudgpu.app/zh/blog/zh-haiwai-zu-gpu-jiage-fukuan-2026)
- [支付宝、Google Pay、Apple Pay 付款方式](https://cloudgpu.app/zh/blog/zh-zhifubao-zu-gpu-mai-api-2026)
- [使用 DeepSeek、GLM、Kimi API:USDT 或刷卡付款](https://cloudgpu.app/zh/blog/zh-deepseek-glm-api-usdt-haiwai-2026)

## Guides (Q&A)

- [How does per-minute GPU billing work on CloudGPU?](https://cloudgpu.app/guides/how-does-per-minute-gpu-billing-work)
- [How much VRAM do I need for FLUX and Stable Diffusion XL?](https://cloudgpu.app/guides/how-much-vram-do-i-need-for-flux-and-sdxl)
- [How much VRAM do I need to run an LLM?](https://cloudgpu.app/guides/how-much-vram-do-i-need-to-run-an-llm)
- [How do I choose a rental GPU for ComfyUI?](https://cloudgpu.app/guides/how-to-choose-a-rental-gpu-for-comfyui)
- [How do I choose a rental GPU for Ollama?](https://cloudgpu.app/guides/how-to-choose-a-rental-gpu-for-ollama)
- [How can I pay for GPU rental?](https://cloudgpu.app/guides/how-to-pay-for-gpu-rental-without-a-us-card)
- [How to set up a conda environment on a cloud GPU](https://cloudgpu.app/guides/how-to-set-up-a-conda-environment-on-a-cloud-gpu)
- [How to download a Hugging Face model on a cloud GPU](https://cloudgpu.app/guides/how-to-download-a-hugging-face-model-on-a-cloud-gpu)
- [How to check GPU utilization on a cloud GPU](https://cloudgpu.app/guides/how-to-check-gpu-utilization-on-a-cloud-gpu)
- [How to install a specific PyTorch and CUDA version on a cloud GPU](https://cloudgpu.app/guides/how-to-install-a-specific-pytorch-cuda-version-on-a-cloud-gpu)
- [How to use SSH port forwarding on a cloud GPU (Gradio, TensorBoard, Streamlit)](https://cloudgpu.app/guides/how-to-use-ssh-port-forwarding-on-a-cloud-gpu)
- [How to run Stable Diffusion WebUI Forge on a cloud GPU](https://cloudgpu.app/guides/how-to-run-stable-diffusion-webui-forge-on-a-cloud-gpu)
- [How to run Unlimited-OCR-2 on a cloud GPU](https://cloudgpu.app/guides/how-to-run-unlimited-ocr-2-on-a-cloud-gpu)
- [How do I keep a training job running after SSH disconnects on a cloud GPU?](https://cloudgpu.app/guides/how-to-keep-a-job-running-after-ssh-disconnects)
- [How do I render a Blender scene on a cloud GPU?](https://cloudgpu.app/guides/how-to-render-a-blender-scene-on-a-cloud-gpu)
- [How do I run a Jupyter notebook on a cloud GPU?](https://cloudgpu.app/guides/how-to-run-a-jupyter-notebook-on-a-cloud-gpu)
- [How to fix "CUDA error: no kernel image is available" on an RTX 5090](https://cloudgpu.app/guides/how-to-fix-cuda-no-kernel-image-on-rtx-5090)
- [How to build llama.cpp with CUDA on a cloud GPU](https://cloudgpu.app/guides/how-to-build-llama-cpp-with-cuda-on-a-cloud-gpu)
- [How to connect VS Code to a cloud GPU over SSH](https://cloudgpu.app/guides/how-to-connect-vs-code-to-a-cloud-gpu-over-ssh)
- [How to fine-tune an LLM with Axolotl on a cloud GPU](https://cloudgpu.app/guides/how-to-fine-tune-an-llm-with-axolotl-on-a-cloud-gpu)
- [How to fine-tune an LLM with Unsloth on a cloud GPU](https://cloudgpu.app/guides/how-to-fine-tune-an-llm-with-unsloth-on-a-cloud-gpu)
- [How to install FlashAttention on a cloud GPU (flash-attn wheels, source builds and RTX 5090)](https://cloudgpu.app/guides/how-to-install-flash-attention-on-a-cloud-gpu)
- [How to install xFormers on a cloud GPU (PyTorch version pins, CUDA index and RTX 5090 "No operator found")](https://cloudgpu.app/guides/how-to-install-xformers-on-a-cloud-gpu)
- [How do I install Nunchaku on a cloud GPU? (ComfyUI-nunchaku wheel matching, FP4 vs INT4, "Failed to install from all available sources")](https://cloudgpu.app/guides/how-to-install-nunchaku-on-a-cloud-gpu)
- [How do I install DeepSpeed on a cloud GPU? (CUDA_HOME does not exist, nvcc missing and CUDA version mismatch)](https://cloudgpu.app/guides/how-to-install-deepspeed-on-a-cloud-gpu)
- [How to install llama-cpp-python with CUDA on a cloud GPU (pip, wheels, server)](https://cloudgpu.app/guides/how-to-install-llama-cpp-python-with-cuda-on-a-cloud-gpu)
- [How to install PyTorch with uv on a cloud GPU (uv venv, --torch-backend and a CUDA check)](https://cloudgpu.app/guides/how-to-install-pytorch-with-uv-on-a-cloud-gpu)
- [How to run SGLang on a cloud GPU](https://cloudgpu.app/guides/how-to-run-sglang-on-a-cloud-gpu)
- [How to run SwarmUI on a cloud GPU](https://cloudgpu.app/guides/how-to-run-swarmui-on-a-cloud-gpu)
- [How to run text-generation-webui (TextGen) on a cloud GPU](https://cloudgpu.app/guides/how-to-run-text-generation-webui-on-a-cloud-gpu)
- [How do I save and resume training checkpoints on a cloud GPU?](https://cloudgpu.app/guides/how-to-save-and-resume-training-checkpoints-on-a-cloud-gpu)
- [How do I train a FLUX LoRA with AI Toolkit on a cloud GPU?](https://cloudgpu.app/guides/how-to-train-a-flux-lora-with-ai-toolkit-on-a-cloud-gpu)
- [How do I train a Wan 2.2 LoRA with musubi-tuner on a cloud GPU?](https://cloudgpu.app/guides/how-to-train-a-wan-2-2-lora-with-musubi-tuner-on-a-cloud-gpu)
- [Is a 32GB RTX 4080 SUPER enough for FLUX, SDXL and 14B-32B LLMs?](https://cloudgpu.app/guides/is-a-32gb-rtx-4080-super-enough-for-flux-sdxl-and-llms): what fits on the RTX 4080 SUPER 32G at FP8 and 4-bit, and how much context is left.
- [RTX 5090 vs RTX 5090D for AI: what is the difference?](https://cloudgpu.app/guides/rtx-5090-vs-rtx-5090d-for-ai): same 32 GB, memory bus and CUDA cores; lower AI TOPS rating on the 5090D; how to benchmark both.
- [Which 48GB GPU should I rent for 70B models and long context: RTX 4090 48G or RTX 4090D 48G?](https://cloudgpu.app/guides/which-48gb-gpu-to-rent-rtx-4090-48g-vs-rtx-4090d-48g): 70B at 4-bit, long-context 32B models, KV-cache headroom.
- [Renting an RTX PRO 6000 Max-Q vs RTX PRO 6000 for 96GB workloads](https://cloudgpu.app/guides/rtx-pro-6000-max-q-vs-rtx-pro-6000-for-96gb-workloads): same 96 GB and bandwidth, 300 W vs 600 W, which workloads fit in 96 GB.
- [What can you run on an RTX 6000D (83GB)?](https://cloudgpu.app/guides/what-can-you-run-on-an-rtx-6000d-83gb): 70B at 4-bit or FP8, gpt-oss-120b, unquantised 32B, FLUX and Wan 2.2 on 83 GB.

## Contact

support@cloudgpu.app · https://cloudgpu.app/contact · X: https://x.com/cloudgpu_app

Agent task preparation: MCP 0.10.0 provides list_task_templates and prepare_task_template for local read-only task drafts. No GPU is created until a separate quoted and approved run_job. Prefer local CPU processing when sufficient. Guide: https://cloudgpu.app/docs/mcp