Easy solution - use hyperscalers with super expensive API charge only when latency really matters. Otherwise build your own DC. Easy to expect customers don't care latency that much over money.