Rates verified 2026-07 · refreshed quarterly

Cloud GPUs & LLM APIs

Three ways to run a model you can’t run at home: rent a GPU by the hour, use a free API tier to prototype, or pay per token. Start with the one that matches your workload.

Cloud AI directory

Curated by Jakub Rusinowski · Entries verified 2026-07-08

This site is about running AI on your own hardware — but the cloud is the right tool surprisingly often: free API tiers for prototyping, rented GPUs for fine-tuning runs and try-before-you-buy, hosted open models when you want Llama-class output without owning a card. Here’s the honest map of what’s out there, and the math for when local wins.

Pricing notes are coarse by design (tiers change often); always confirm on the provider’s page. Most links are plain links; RunPod and Vast.ai are affiliate links.

Free LLM API tiers

Genuinely free API access — rate-limited, but real. Enough for prototyping, learning and low-volume tools before you spend anything on hardware or tokens.

NVIDIA NIM (build.nvidia.com)Free: Free hosted endpoints for Llama, Nemotron, Mistral & more with generous dev credits — sign-up only, no cardFree API credits for development; self-host NIM containers with NVIDIA AI Enterprise in productionTrying big open models (Llama 3.3 70B, Nemotron) over an OpenAI-compatible API before buying hardware Google AI StudioFree: Gemini Flash free tier — the most generous standing free allowance of the big threeFree-of-charge tier with per-minute/per-day rate limits; paid tier unlocks higher limitsPrototyping against a frontier-family model at zero cost GroqFree: Free developer tier with daily rate limits on open models (Llama, Qwen, Gemma)Pay-per-token on LPU hardware; famous for extreme tokens/secLatency-critical demos — hundreds of tokens/sec on 70B-class open models OpenRouterFree: Rotating set of :free model variants with daily request capsOne API key, 300+ models, pass-through pricing per modelComparing many models through one OpenAI-compatible endpoint before committing Mistral La PlateformeFree: Free experiment tier with rate limitsPer-token pricing on Mistral models; EU-hostedEU-hosted API access with a free on-ramp — relevant for GDPR-sensitive prototyping Cerebras InferenceFree: Free developer tier with daily token capsPer-token pricing on wafer-scale hardware; very high tokens/secFastest hosted open-model inference alongside Groq — good for agent loops GitHub ModelsFree: Free playground + API rate limits for GPT-4o, Llama, Phi, Mistral with any GitHub accountIncluded with a GitHub account; production use routes to Azure AIZero-signup-friction experimentation if you already live on GitHub Cloudflare Workers AIFree: Daily free allocation of neurons on every Cloudflare accountPer-neuron (unit) pricing beyond the free allocation; runs on Cloudflare edge GPUsSprinkling small-model inference (Llama 8B class) into edge apps without managing servers Hugging Face Inference ProvidersFree: Small monthly free credit allowance on every HF accountRoutes to partner providers at pass-through prices; monthly included credits on free & PRO plansCalling almost any open-weight model on the Hub without picking a vendor first

Cloud GPU rental

Rent a GPU by the hour instead of buying one. The right answer for fine-tuning runs, batch jobs and “try a 70B before committing to a $2,000 build”.

RunPodRTX 4090 ≈ $0.3–0.7/hr, A100/H100 by the hour; serverless per-second billing availableRenting a 24–80 GB GPU for hours-long jobs: fine-tuning, batch inference, trying 70B before buyingAffiliate link Vast.aiMarketplace of hosted GPUs — often the cheapest per hour (interruptible options)Cheapest possible GPU-hours when you can tolerate marketplace variabilityAffiliate link LambdaOn-demand A100/H100/B200 instances and clusters, per-hour billingSerious training/fine-tuning runs on datacenter GPUs with clean images Paperspace (DigitalOcean)Free: Free-tier Gradient notebooks with entry GPUs, subject to availabilityPer-hour GPU machines + Gradient notebooks; some free-GPU notebook availabilityNotebook-first experimentation with an easy upgrade path HyperstackAggressively priced A100/H100 per-hour instances in EU/NA regionsEU-region datacenter GPUs at low hourly rates TensorDockMarketplace pricing; consumer + datacenter GPUs from ~$0.2/hrBudget hourly GPUs with per-minute billing Google ColabFree: Free T4 GPU notebook sessions — the classic zero-cost on-rampFree notebooks with T4 GPUs (availability-based); Pro tiers add better GPUs and runtimesLearning, small fine-tunes (LoRA on 7–8B), and quick experiments without any setup

Hosted open-model APIs

Open-weight models (Llama, Qwen, Mistral, DeepSeek) served per token by specialists — the middle ground between frontier APIs and your own hardware.

Together AIFree: Small starter credit for new accountsPer-token pricing on 100+ open models (Llama 3.1 8B ≈ $0.18/M tokens)Production open-model serving with serious throughput and fine-tuning endpoints Fireworks AIFree: Starter credit for new accountsPer-token open-model serving; fast function-calling-optimized deploymentsLow-latency production inference on open models, incl. dedicated deployments DeepInfraAmong the lowest per-token prices for open models; OpenAI-compatibleCheapest hosted per-token rates on Llama/Qwen/Mistral-class models ReplicatePer-second (hardware-based) or per-token pricing; huge community model catalogRunning community models (image, audio, video, LLM) without packaging them yourself Novita AIFree: Small trial creditBudget per-token open-model APIs + GPU instancesLow-cost open-model APIs with an Asia-Pacific footprint

Frontier model APIs

The closed frontier models. Highest quality, zero setup, per-token billing — the benchmark every local setup gets compared against.

OpenAI APIGPT-4o $2.50/M input, $10/M output; GPT-4o Mini $0.15/$0.60Frontier quality and the broadest tooling ecosystem Anthropic APIClaude Sonnet $3/M input, $15/M output; Haiku $0.80/$4Best-in-class coding and long-context analysis Google Gemini APIFree: Free tier via Google AI Studio (see above)Gemini Flash $0.10/M input, $0.40/M output; Pro $1.25/$5Huge context windows and the lowest frontier-family prices

As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.

Local vs cloud: which should you use?

Editorial comparisons with a shared break-even calculator — honest about where each side wins. Cloud prices verified 2026-03-01.

Start hereLocal AI vs Cloud API Costs: The Real Break-Even Math (2026)Local AI vs cloud API costs compared with real numbers: hardware, electricity, and per-token prices. Break-even calculator shows when a GPU pays for itself.

Frequently asked questions

Where can I use AI models for free via API?

The most useful standing free tiers in 2026: NVIDIA NIM (build.nvidia.com) gives free hosted endpoints for Llama, Nemotron and Mistral-class models; Google AI Studio offers a generous Gemini Flash free tier; Groq and Cerebras run free rate-limited developer tiers on open models at extreme speed; OpenRouter exposes rotating free model variants; and GitHub Models works with any GitHub account.

What is the cheapest way to rent a GPU for AI?

Marketplace providers are cheapest: Vast.ai and TensorDock list consumer GPUs (RTX 3090/4090 class) from roughly $0.20–0.40/hour, RunPod adds per-second serverless billing, and Lambda/Hyperstack cover datacenter A100/H100 cards. For free experimentation, Google Colab still offers T4 notebook sessions at no cost.

Should I rent a cloud GPU or buy my own?

Rent when usage is occasional (fine-tuning runs, trying a 70B model before committing) — a few dollars per experiment. Buy when usage is sustained: at a few hours of daily inference, a $1,000–2,600 build typically undercuts rental within months. Our local-vs-cloud comparison pages include a break-even calculator.

What is the difference between hosted open-model APIs and frontier APIs?

Hosted open-model providers (Together, Fireworks, DeepInfra, Groq) serve open-weight models like Llama and Qwen per token — often 10–50× cheaper than frontier APIs, and you can migrate the same model to your own hardware later. Frontier APIs (OpenAI, Anthropic, Google) serve closed models with the highest quality ceiling.

By budget

By use case

By team size

By privacy & compliance

Prefer to own the hardware?

Tools