I did some math, and running my Nvidia GB10 workstation (Asus GX10) costs me maximum: 12~13 USD / month or 150~160 USD / year It is a little bit above half the price of ChatGPT plus subscription. For that, I get to run models that can fit in 128 GB of memory How I calculated: You can see how much power your apartment uses in Singapore in half-hourly resolution. We turned off all devices and A/C while we sleep, and got only the fridge and the GB10 remaining From that, we see it uses around 80-100 Watt while I was running an inference workload overnight. So this is like an upper bound I take it as 90 Watt. Electricity here costs 0.25 SGD / kWh 0.09 * 0.25 * 24 * 30 * (SGD/USD conversion rate) = 12~13 USD / month = 150~160 USD / year Local models are getting very good now, small ones roughly around GPT 5.x-mini level. This workstation makes all sorts of workloads possible for me that would otherwise cost a ton on the API It is also my always on workstation that works overnight. I use Codex for my work, and my workstation is always running agents. It never sleeps. I never have to worry about keeping my laptop lid open. I connect and monitor the agents anytime on my phone using mosh and herdr We have crossed a threshold. Running local models is cheaper than a big token sub for quite a few workloads already. If you are running a business, that makes a difference The localening is here