I really wanted to tell her -- what she needed was compute, not hardware. Her best bet was to rent GPU which is easily available from many services. If she was paid to run inference, cost is not a concern. If she did it for her own work, she wouldn't need to shell out a ton of money initially and then be left with hardware that she no longer needs when the work is done. In any case, she wouldn't need to worry about purchasing hardware or battery life. ssh to a remote machine is simply way superior than running inference on a laptop she carries around by any means except in a few rare occasions (e.g. no Internet access).
Same thing for a company. Why would any company who has a business to run get themselves burned in acquiring GPUs, setting up racks, load balancing etc when Bedrock/Azure Foundry etc is right there with tons of both open and closed models and ZDR? And then see hardware depreciate in just a few years?