ガイド： GPU クラウド ユーザー向け
個人の AI 開発者に向けた、プロバイダー比較、デプロイ手順、率直なコスト分析。
How do I use Kimi K3 in LangChain? (ChatOpenAI + base_url, 2026)
Point LangChain's ChatOpenAI at https://cloudgpu.app/v1 with model kimi-k3 and $CLOUDGPU_API_KEY. Why Kimi K3 needs a large max_tokens, and what ChatOpenAI does with reasoning_content.
How to Use DeepSeek V4 Flash in Roo Code (OpenAI-Compatible Setup, 2026)
Point Roo Code at CloudGPU's OpenAI-compatible endpoint and use deepseek-v4-flash: base URL, API key and model ID, step by step.
How do I run Nemotron 3.5 Transcribe on a cloud GPU? | CloudGPU
Serve NVIDIA's Nemotron 3.5 Transcribe (34B total, 4.3B active) on a rented GPU with vLLM and NeMo Speech: the official hardware table, nvidia-smi and nvcc checks, the BF16 and NVFP4 commands, and a transcription request.
How AI labs use AI to train AI, and what a one-developer version looks like
Five ways AI labs have published about using models in training, each with its official source and date, plus a small-scale version for one developer.
How DeepSeek V4 API Billing Works: Per-Token Rates, Cached Input and One Balance
How DeepSeek V4 API billing works on cloudgpu.app: Flash and Pro are billed per token from one prepaid balance; current rates and cached input are on /api.
租 GPU 自建大模型还是按 token 调 API?盈亏平衡公式与实算(2026)
租 GPU 自建和按 token 调 API,成本什么时候持平?给出盈亏平衡公式、Qwen3 32B 在 32 GB 卡上的实算、利用率影响和可直接运行的脚本。
销毁云 GPU 实例前怎么保存数据:检查点、rsync、Hugging Face 上传清单(2026)
销毁租用的 GPU 实例会删除磁盘。用 rsync、scp、Hugging Face 上传或 rclone 把检查点、LoRA 适配器和生成结果拷出机器的方法与清单。
上下文长度要多少显存?KV 缓存计算公式与 Qwen3、Llama 3 实算表(2026)
KV 缓存随上下文线性增长:给出计算公式、Qwen3 / Llama 3 / Mistral 每 token 占用(取自 config.json),以及 24–80 GB 显卡各能放多长上下文。
Self-Host an LLM on a Rented GPU or Pay per Token? A Break-Even Calculator (2026)
When does renting a GPU beat a per-token API? The break-even formula, a Qwen3 32B example on a 32 GB card, utilisation and a script.
Save Your Work Before You Destroy a Cloud GPU: Checkpoints, rsync and Hub Uploads (2026)
Destroying a rented GPU deletes its disk. Copy checkpoints, LoRA adapters and outputs off the machine with rsync, scp, Hugging Face or rclone.
KV Cache VRAM Calculator: How Much Memory Does Context Length Need? (2026)
KV cache memory for any context length: the formula, per-token sizes for Qwen3, Llama 3 and Mistral from config.json, and what fits on 24–80 GB.
按分钟租一张 RTX 4090 48G 云 GPU（2026）
RTX 4090 48G 是什么、怎么判断你的模型是否需要 48 GB 显存、按分钟计费怎么算，以及怎样开好一台。价格见价格页。
Rent an RTX 4090 48G in the Cloud, Billed by the Minute (2026)
What the RTX 4090 48G is, how to work out whether your model needs 48 GB of VRAM, how per-minute billing works, and how to start one. Rates are on the pricing page.
How to Pay for the Claude API: USDT, Alipay, Google Pay (2026)
Top up one prepaid balance with Visa/Mastercard, Alipay, Google Pay, Apple Pay, USDT on TRC-20 or business bank transfer, then call Claude models over the API, billed per token.
How to Use Claude in Cursor, Cline and Continue With Your Own API Key (2026)
Set up Claude Sonnet 5.5 and Opus 5.5 in Cursor, Cline and Continue through one OpenAI-compatible endpoint: exact fields, config.yaml, and what to check when it fails.
Run Claude Code on a Custom API Endpoint (2026): Base URL, Key, Model Aliases
Point Claude Code at cloudgpu.app with a custom base URL and a cgw-sk key, map the Haiku slot, check it with one curl and /status, then save it in settings.json.
Claude API Proxy: Using Claude Through an OpenAI-Compatible Endpoint (2026)
What a Claude API proxy or gateway actually is, how calling Claude through an OpenAI-compatible endpoint works, and the honest trade-offs: what you gain, what you give up, and how to evaluate one before you trust it.
Claude API Access With Alipay, Card or USDT (2026)
Call Claude Opus and Sonnet through one OpenAI-compatible endpoint and pay with Alipay, Google Pay, Apple Pay, a bank card or USDT.
用支付宝、Google Pay、Apple Pay 租 GPU、买大模型 API：支持哪些付款方式、怎么充值（2026）
cloudgpu.app 的 GPU 和大模型 API 共用一个预付余额，可以用 Visa / Mastercard、支付宝、Google Pay、Apple Pay、USDT（TRC-20）或对公银行转账充值。这里写清楚每种方式怎么充。
Pay for Cloud GPUs and AI APIs with Alipay, Google Pay, Apple Pay, a Card or USDT (October 2026)
Every way to top up a cloudgpu.app balance: Visa and Mastercard, Alipay, Google Pay and Apple Pay in one card checkout, USDT on TRC-20, and bank transfer, and what each one is good for.
Fine-Tune With LLaMA-Factory on a Rented GPU, Billed by the Minute (2026)
How to run a LoRA or QLoRA fine-tune with LLaMA-Factory on a rented GPU: renting the card, installing LLaMA-Factory, training, and saving your adapter. Billing is per minute; rates are on the pricing page.
GLM-5 API, OpenAI-Compatible: Pay by Card or USDT (2026)
Call GLM-5.2, GLM-5.1 and GLM-5.3-Flash through an OpenAI-compatible endpoint, pay by card or USDT, and set it up in Cursor, Cline or the OpenAI SDK.
Kimi K2 and K3 API, OpenAI-Compatible: Pay by Card or USDT
Call Kimi K2 and K3, including kimi-k3 and kimi-k2.7-code, through an OpenAI-compatible endpoint and pay with a card or USDT. Model IDs are listed on /api.
按分钟租 GPU：4090 / 5090 / A100 怎么选、怎么付款（2026）
RTX 4090、RTX 5090、A100 按分钟计费，现价见 /pricing。支付宝、Google Pay、Apple Pay、银行卡（Visa / Mastercard）或 USDT（TRC-20）付款，不需要身份证。
跑 32B 模型选 4090 还是 5090？Qwen3 32B / DeepSeek-R1 32B / QwQ 在 24、32、48 GB 卡上的显存与每小时成本（2026）
32B 模型 Q4_K_M 权重 19 到 20 GB：24 GB 的 RTX 4090 放得下但只剩 8k 上下文；32 GB 的 RTX 5090（$0.59/h）能带 16k 到 32k，是推荐卡；48 GB 的 A6000 / 4090 48G 能上 Q8。附 Ollama 一条命令开机、每小时和每天的成本。价格 2026 年 9 月 15 日。
Run ComfyUI and FLUX.1 on a Rented RTX 4090, RTX 5090 or A100 by the Minute, With a Public HTTPS Endpoint (2026)
Deploy the ComfyUI template on an RTX 4090, RTX 5090 or A100 80G billed per minute, download FLUX.1 weights on the machine, export an API workflow and call /prompt over HTTPS with curl or Python.
How Much VRAM to Run Each Open Model? (2026 Table, 24 to 96 GB Cards)
A working table of VRAM needed to run Llama, Qwen, Gemma, Mistral, DeepSeek distills, gpt-oss and FLUX at FP16 and Q4, with KV-cache headroom, mapped to the GPU you should rent: RTX 3090, 4090, 5090, A6000, A100 or RTX PRO 6000.
按分钟租一张 4090/5090 跑 Ollama，拿到一个公网 HTTPS API 地址
在 cloudgpu.app 上按分钟租 RTX 4090 或 RTX 5090，Ollama 模板自带公网 OpenAI 兼容端点。附 curl、Python、Cursor 配置和背后的隧道原理，现价见 /pricing。
怎么用 DeepSeek V4、GLM、Kimi 的 API：USDT 或刷卡付款，OpenAI 兼容接口
用 USDT 或刷卡付款，拿到 DeepSeek V4 Flash/Pro、GLM-5、Kimi K3、MiniMax 的 API key。附 Cursor/Cline 配置方法和数据去向说明，现价见 /pricing。
Rent an RTX 5090 in the Cloud by the Minute (2026)
How to rent an RTX 5090 (32 GB) in the cloud with per-minute billing, how to pay without a US credit card, and what you get with the machine. Rates are on the pricing page.
Run Ollama on a Cloud GPU with a Public HTTPS API
Step by step: rent an RTX 4090 or 5090 by the minute, get Ollama running with a public OpenAI-compatible endpoint on cloudgpu.app, and call it from curl, Python or Cursor. Explains what happens under the hood.
DeepSeek V4 API With USDT or Card: Use the OpenAI SDK
Get a DeepSeek V4 Flash and Pro API key you can pay for with USDT or a card, and plug it into the OpenAI SDK, Cursor, Cline or OpenClaw. Rates are on /pricing.
How to Use DeepSeek in Cursor, Continue, and Cline, Paid With USDT or a Card
Step-by-step setup for using DeepSeek-V4 in Cursor, Continue.dev and Cline through an OpenAI-compatible endpoint, paid from a balance you top up with USDT or a card.