My intuition is that network transfer uses almost no energy, on a marginal basis, because the network links are always on anyway. When you pay for bandwidth, you're paying the amortized costs of building and maintaining the infrastructure.

Rendering a SPA on a client device consumes CPU time, which definitely uses energy proportional to the number of client devices.

Assume the server uses less clock cycles to render than the client--e.g., because of optimized software, hot caches, or tuned hardware.

Therefore rendering on the server should always be more energy efficient than on the client.