Air freight is also an odd comparison since it's usually time-sensitive and/or pricey ($100+ per pound).

I'm not sure I understand the point of your analogy with respect to wining/losing at a casino. If I follow it, you are effectively saying something akin to, when Google first became profitable, its achievements should have been completely dismissed because of a decade of prior losses. That's not how tech companies work. Most of the value creation is 5-10-20 years out. You bear enormous losses with the knowledge that returns compound and eventually render irrelevant prior years' losses.

Source: https://www.google.com/finance/quote/UBER:NYSE?hl=en

Couple of issues I can see: (1) most devices out there would probably be mobile, so no NVIDIA/CUDA for you; (2) even with binding to, say, Apple Silicon, you might still be memory limited, i.e. can you fit the entire net on a single mobile GPU; (3) network latency?

For Seam, we purchase, set up, and test many individual devices and systems in our lab in San Francisco. During the course of this work, we discover quite a few interesting things. When possible, we work directly with manufacturers on addressing the more concerning problems we find. We maintain an internal device database (partially available here https://www.seam.co/supported-devices-and-systems) where we keep track of our findings on devices we test & integrate. One area that I haven't seen addressed here is data-storage jurisdiction. imho, that might be one of the more concerning aspect.

happy to have a chat; my seam email is in my hn profile.

Fwiw My most reliable trick ended up piggie-backing off of a physical line going into Hong Kong from Shenzhen, and when roaming around China, using a vpn to get to that shenzhen gateway. As far as I can recall, that always worked. This led me to believe that most of the vpn traffic analysis (and blocking)was done at the edge of the GFW and not inside of it. Again, this could be outdated by now.