Ollama Cloud Pricing 2026: Plans vs Renting a GPU

Ollama Cloud pricing changed on 31 Aug 2026: the old $20 plan with 5-hour and weekly caps became per-token billing with a monthly credit pool. That makes it much easier to compare against the other way to run Ollama in the cloud, which is renting a GPU and running ollama serve yourself. This post lays out the current plans, what $20 actually buys, and the break-even math against a rented H100 GPU rental and an RTX 5090.

TL;DR: Ollama Cloud Pricing vs Renting a GPU

Ollama Cloud Pro costs $20/month for $60 of token usage; a rented Spheron GPU is a flat hourly rate.

Ollama plans as of 20 Oct 2026; Spheron rate as of 20 Oct 2026. Ollama Cloud wins on light use, a rented GPU running vLLM at sustained volume.

How Ollama Cloud Pricing Works

Ollama Cloud runs models on datacenter GPUs that Ollama provides. You call them through the same ollama CLI and API you use locally, so a model tagged for cloud runs remotely while everything else stays on your machine.

The Ollama pricing page and the 31 Aug 2026 pricing announcement set out how billing works now:

• Usage is metered in tokens, not GPU time. Each model has its own input, cached input and output rate per million tokens.
• Paid plans include more usage than they cost. Pro is $20 and includes $60 of usage; Max is $100 and includes $300. The included pool resets on your billing day and does not roll over.
• Overage is the same rate. When the pool runs out, you keep going at the listed per-token rates using purchased credits. There is no separate overage fee.
• No more session or weekly caps. The 5-hour and weekly limits on the older plans are gone on the new ones. Legacy subscribers keep the old terms until they change plan.
• Concurrency is the real ceiling. Free allows 1 request at a time, Pro 3, Max and Team 10. Extra requests queue, and are rejected once the queue fills.
• Off-peak discounts exist on some models. DeepSeek models bill at half rate outside 12:00-18:00 UTC on weekdays and all weekend.

Ollama Cloud per-token rates

A sample of the listed rates, per million tokens, as of 20 Oct 2026:

These are in the same range as hosted open-model APIs. For context on how other providers price the same models, the Fireworks AI pricing breakdown covers a direct competitor.

What $20 a month buys

Take gpt-oss:120b with a typical chat or agent mix of three input tokens for every output token and no cache hits. Each million output tokens then costs $0.60 plus $0.45 of input, $1.05 in total. The Pro plan's $60 pool covers about 57 million output tokens and 171 million input tokens a month.

The same $20 rents an H100 on a marketplace for well under a day. For one developer chatting with a model and running a coding agent, Ollama Cloud Pro is hard to beat.

Running Ollama on a Rented GPU Instead

The alternative is to rent a GPU and run Ollama yourself. It is the same software you run on your laptop, so the setup takes a few minutes:

curl -fsSL https://ollama.com/install.sh | sh
sudo systemctl edit ollama   # under [Service], add: Environment="OLLAMA_NUM_PARALLEL=4"
sudo systemctl restart ollama
ollama pull gpt-oss:120b

Keep the API bound to localhost and reach it through an SSH tunnel, since Ollama has no authentication of its own:

ssh -L 11434:localhost:11434 user@your-gpu-server

Your apps then call http://localhost:11434, including the OpenAI-compatible /v1/chat/completions route. If you have not set up Ollama on a GPU before, our guide to running LLMs with Ollama covers GPU detection, VRAM sizing and quantization formats.

What changes compared to Ollama Cloud:

• You pay by the hour, not the token. A busy hour and an idle hour cost the same.
• Any model works. Every GGUF on Hugging Face, your own fine-tunes, and older models Ollama Cloud does not host.
• Concurrency is up to you. OLLAMA_NUM_PARALLEL sets it, bounded by VRAM for the KV cache.
• Prompts stay on your server. Useful when client data cannot go to a third-party API.

Break-Even Math: Ollama Cloud vs a Rented GPU

The comparison comes down to one number: how many output tokens per hour your rented GPU actually produces. Divide the hourly rate by that and you get cost per million tokens, which you can set directly against Ollama Cloud's rates.

That throughput depends heavily on the serving engine. Ollama is built for one user at a time and batches concurrent requests far less efficiently than a production server. vLLM uses continuous batching and can push several times more tokens through the same GPU when many requests arrive at once. Our Ollama vs vLLM comparison explains why.

Worked example, with rates frozen on purpose so the ratios stay consistent: H100 PCIe at $2.64/hr (the live Spheron on-demand rate on 20 Oct 2026) and RTX 5090 at $0.84/hr (the last recorded Spheron on-demand rate, 17 Sep 2026). Throughput figures are planning assumptions for aggregate output across concurrent requests, not benchmarks; measure your own before deciding.

The Ollama Cloud column includes three input tokens per output token at list rates, with no cache discount. These rates are a snapshot; for today's Spheron rates, see live GPU pricing. The current H100 rate is $2.75/hr and the RTX 5090 rate is $0.78/hr.

Two things stand out. First, running Ollama itself on a rented GPU does not beat Ollama Cloud on cost for these models, even with the GPU busy every hour. A hosted service spreads each GPU across many customers' requests; your single Ollama server only has your traffic, and Ollama batches it poorly. Second, switching the rented GPU to vLLM flips the result. An H100 serving gpt-oss:120b through vLLM undercuts Ollama Cloud once it averages about 28% of its capacity, which over a month at the frozen rate is roughly 1.8 billion output tokens. Below that volume, Ollama Cloud is cheaper.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 20 Oct 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.

Which One Should You Use?

Pick Ollama Cloud when:

• You are one developer or a small team and usage is a few hundred million tokens a month or less.
• The models you want (gpt-oss, Gemma, DeepSeek, Kimi, GLM) are on its list.
• Three or ten concurrent requests is enough.
• You would rather not manage a server at all.

Pick a rented GPU when:

• Volume is high and steady enough to keep a GPU above roughly a quarter of its capacity, and you are willing to serve through vLLM or SGLang instead of Ollama.
• You need a model Ollama Cloud does not host, such as a fine-tune or a niche GGUF.
• You need more than 10 concurrent requests, or tighter latency control than a shared queue gives.
• Prompts or outputs must not leave infrastructure you control.

A common pattern is to start on Ollama Cloud, then move the heavy, predictable traffic to a rented GPU once the monthly bill shows it. Keep the per-token bill for the long tail. If your traffic is bursty rather than steady, the serverless vs on-demand vs reserved guide covers the billing models in between, and the GPU cost optimization playbook has the levers for cutting a self-hosted bill further. For smaller models, a single Spheron RTX 5090 holds gpt-oss:20b with plenty of KV cache headroom.

Outgrowing a per-token bill? Run Ollama or vLLM on a dedicated GPU with per-minute billing and no token metering.

H100 on Spheron → | RTX 5090 capacity → | View all pricing →

Get started on Spheron →