The hardest part of starting a company may be miscalculating costs, not finding customers
Last week in Hangzhou, I met a friend building an AI customer-service product. Halfway through his budget spreadsheet, he was preparing to buy a batch of GPUs because “the prices might keep rising.” Then he asked me: if he ordered now, would he be stockpiling resources or strangling his cash flow? I’ve fallen into the same trap—after seeing everyone else rush, I forget whether my own orders are even sufficient.
According to a TechPowerUp report, Nvidia is alleged to be artificially creating demand and scarcity by funding AI startups that buy its GPUs. The report uses “allegedly,” so the claim still needs to be treated cautiously. The issue is now being discussed among investors and AI founders. For us, the useful question isn’t whether we can guess the vendor’s strategy. It’s how to avoid letting industry panic become our purchasing rationale.
The number that matters: who pays when the machine sits idle
Yan Lin, who handles AI document processing in Beijing, didn’t chase GPUs last month. Instead, she broke down her actual daily usage into peak and low periods, then compared rented compute with long-term purchasing. She found that her team had plenty of customers during the day, but the GPUs often sat idle at night. Optimizing the job queue and renting on demand was more stable than buying capacity upfront. When we’re running side-project tools or small-team services, what we usually lack isn’t “compute freedom,” but stable customers.
Today’s replication cost: build a 30-minute decision sheet
Money: 0 yuan, using a free spreadsheet. Time: 30 minutes. Technical barrier: no coding required—just customer count, cost per processing run, and equipment utilization rate. First step: open a spreadsheet on your computer, create four columns labeled “usage, unit price, utilization rate, monthly spend,” and fill in the real numbers from the last seven days. Don’t substitute industry rumors for your own data. If you do need to buy, add maintenance, electricity, and advance-payment costs.
Decide from your stage, don’t let shortages lead the way
If I had just started and didn’t have stable customers yet, I’d rent by the job or use free quotas first and keep the budget for customer acquisition. It’s fine not to try yet; once demand appears, I can calculate again. If I had one or two stable customers, I’d first check whether the revenue each customer contributes covers the equipment cost before reserving a purchase. If I were scaling and usage had stayed stable for several consecutive weeks, I’d compare long-term purchasing, renting, and cloud options while keeping cash available in case prices rise. Vendor strategy is worth tracking, but panic shouldn’t place the order.