| One caveat to alerts (and automatically acting on alerts) is that there are delays[0] between costs being incurred and alerted. I can't find a Google source for what the delay is, but a source online say it could be "24 hours [to] a few days."[1] This has been a major reason why I reach for OpenAI models before Gemini, but also why I'd rather use services like RunPod for training jobs. For a small boostrapped company like mine, it feels terrifyingly easy to rack up a company-ending AI bill. The cloud companies try to limit these accidents through cranking your quotas down to nothing, but this also means that my small company can't just whip up 8xH100 easily without major ceremony, and I have routinely been rejected the GPUs quotas I needed for projects. Accidentally leaving that kind of node on for the 24 hours that it might take to get an alert would rack up a $2,000+ bill, compared to $500 on RunPod, which will also stop the instance when you run out of money. I've loved working with major cloud providers at growing VC-funded startups that have credits, TAMs and bigger budgets for errors. But hyperscalers are fairly difficult for a pre-scale bootstrapped business, and arguably not designed or optimized for it. [0] https://docs.cloud.google.com/billing/docs/how-to/disable-bi...
[1] https://support.terra.bio/hc/en-us/articles/360057589931-How... |