At 9 a.m., I checked my power bill first
Last Wednesday at 9 a.m., I was revising an automated quotation proposal for a client in Hangzhou. I opened the server bill and saw a jump from the previous month. My first thought was, “Have I been using this too heavily?” Then I realized the data centers powering AI services are competing with cities and factories for electricity. People like us, running side projects or small teams, usually see only “how much does each generation cost,” not the much larger energy bill underneath.
I’ve been stuck here too. I once thought avoiding the most expensive plan would save money, then ran a large batch of repetitive tasks and burned through the budget anyway. This is not overthinking. Costs really do travel from the data center to our tool prices, electricity bills, and community expenses.
This is not a new tool; it changes how we choose tools
Simply put, data centers are centralized places where AI and web services run, and they need a lot of electricity. The Cornell University report was not about one particular piece of software. It focused on the external costs created by this expansion: some lower-income residents in rural communities may face more direct price pressure.
This matters to us because it will affect which services can remain cheap over the long term, which features shift to usage-based pricing, and which small companies prioritize lighter-weight solutions.
I know a freelance designer named Chen Lan. She writes menu copy for four restaurant clients each month. She used to put every revision into one expensive subscription and always went over budget by month-end. Then she separated first drafts from major revisions: she used cheaper pay-as-you-go tools for first drafts and called a more expensive model only after the client confirmed the direction. In June, she cut the cost per task from RMB 31 to RMB 18. She saved money not by fighting the system, but by running fewer unnecessary tasks.
If I tried this today, the budget and first step are specific
The cost of testing this approach is concrete: prepare less than RMB 50 and try it for one hour. The technical barrier is low—registering an account and opening the usage page is enough. The first step is not switching tools. Log in to the AI platform you already use, find “usage” or “billing,” and note your three most expensive task types from last week.
Then test them in this order: turn off unnecessary background automations; move lower-difficulty tasks such as summaries and classification to cheaper models; switch to a stronger model only for important client work. After each batch, record the actual cost and number of revisions. I once treated “the stronger the model, the safer” as a rule, then paid for more than 20 generations for a single formatting change. I got that step wrong.
Not everyone needs to do this. If your business barely uses AI, or you value stability more than savings, skipping this for now is fine. What matters is that when traffic is stable, we are not using power-hungry capabilities for every small job.
Match the approach to your stage—change one thing at a time
If I were just starting out, I would first set aside a RMB 50 trial credit and run it only on the one or two most repetitive tasks in my actual business, without rushing to migrate everything.
If I already had one or two clients, I would pull the last month’s task records, separate “high-value judgment” from “ordinary generation,” and review usage and rework costs once a week.
If I were scaling up, I would calculate the actual consumption per client before discussing a team plan or automation. Otherwise, as the headcount grows, inefficient tasks get replicated many times over.
AI will keep entering the tools we use, but our goal is not to chase every change. Understanding where every yuan goes is already more stable than blindly chasing the newest model.