> ## Content Index
> Fetch the complete content index at: https://www.techstories.co/llms.txt
> Use this file to discover other available public pages before exploring further.

# Your GPU can save you millions of cloud tokens a day
- URL: https://www.techstories.co/your-gpu-can-save-you-millions-of-cloud-tokens-a-day/
- Published: 2026-11-20T09:41:47.000Z
- Updated: 2026-11-22T13:08:14.000Z
- Description: The token maths behind running agents on a consumer card.
- Author: Paul Mah
- Tags: AI, Gadgets

As Anthropic goes public, AI agents are going mainstream. But did you know you can run them on your own PC? Here's why you should, and how.

### Cloud, local, or both?

Agents can be addictive. Get them to diagnose a problem on your PC, tally up your receipts, or write a difficult email, and they do it so well that you keep coming back for more.

There are cloud agents such as [OpenAI Dots](https://openai.com/index/introducing-dots/?ref=techstories.co) and [Meta Muse](https://ai.meta.com/muse/?ref=techstories.co), and local ones such as OpenClaw or Hermes Agent that run on your own devices. So which is better, cloud or local? I personally think the future will be hybrid.

Why go local at all, though? There are two main reasons. The first is cost. Being able to handle simple tasks on your PC means you're not burning expensive cloud tokens on them, which matters more and more as cloud token quotas are gradually reduced. The second is control. For hobbyists managing their home automation or smart devices with AI, it makes far more sense to do so locally than through a cloud AI, whether you look at it in terms of cost, reliance on connectivity, or privacy.

### The token maths

So how much does an agent actually cost to run? This is typically calculated in tokens, and the figures vary widely. But as a rough guide, checking for unread email takes somewhere between 1,000 and 3,000 tokens, a web search runs to 10,000 to 20,000, and a daily inbox briefing comes in at 10,000 to 30,000.

Those numbers add up over time, and can potentially reach millions of tokens a day. Thankfully, consumer GPUs such as the Nvidia 5090 can easily generate 60 to 70 tokens per second (tps), rising to 200 to 300 tps with concurrent streams. Put that to work on the everyday jobs, and the cloud token savings are considerable.

![](https://www.techstories.co/content/images/2026/11/source-ed528e484cbbd0c6.jpg)

![](https://www.techstories.co/content/images/2026/11/source-fc91399ccd260a5e.jpg)

![](https://www.techstories.co/content/images/2026/11/source-804305349595a089.jpg)

### Three ways to go local

Looking to use local agents? There are three options, depending on how much power you need.

The first, and simplest, is agentic flexibility on your PC. You run the AI agent directly on the machine. With a modern consumer GPU, you generate tokens locally to power agents such as OpenClaw and Hermes Agent. The advantage? No bill shock from API billing, and no need to eat into the precious tokens you're saving for that coding project.

The second is for the power user. I was initially sceptical, but when I tested the [Asus GX10](https://www.asus.com/sg/networking-iot-servers/desktop-ai-supercomputer/ultra-small-ai-supercomputers/asus-ascent-gx10/?ref=techstories.co), I was quickly blown away by its ability to run top AI models from such a small but capable box. Three or more GX10 units can even be wired together over their 200Gbps ConnectX-7 links to run larger models or to output tokens faster.

The third is departmental-level power. If you need even more, consider the [Asus ET900N G3](https://www.asus.com/displays-desktops/workstations/performance/expertcenter-pro-et900n-g3/?ref=techstories.co). It comes with 72 processor cores, a Blackwell GPU, and enough memory to run DeepSeek 4.1 and GLM 5.3 locally. I'm currently testing the ET900N myself, and I'll report back in a couple of weeks or so.

In the meantime, it's worth asking whether the PC already sitting on your desk could be doing some of your agents' work for you.