Most stuff eventually turns into cost so I'll ignore "costly" effects like heat/power and performance per dollar and focus on max performance scale for workloads and other unique differences not achievable via just throwing more money at the alternative.
Per socket performance scaling is higher for equivalent tier sockets. At hyperscale that goes back into the price benefit (buy and maintain less physical data center) but for an individual server workload or individual user that also turns into a performance benefit, particularly for non NUMA aware workloads on the server side and just plain availability of such core counts for performance on the desktop or workstation side.
PCIe wise you get about twice the lanes (128 total on AMD) of even a 40 core 8380 in a the base 8 core model of Epyc or a Threadripper workstation CPU.
A place Intel still wins is total NUMA scaling. For a NUMA aware app like SAP HANA Intel can scale to 8 sockets while AMD currently tops out at 2 so you can reach about 2x as many total threads that way.
I had a sister comment which wasn't as thorough as yours so I deleted it. It's worth adding though that for mobile applications, power consumption isn't just a cost factor, since better efficiency means you can have tighter packaging, get more battery life, not roast your lap as much, have quieter cooling, etc.
For non-NUMA aware workloads with high inter-core coordination (for example, a write heavy database workload) Intel will still perform much better because the cross-chiplet latency of EPYC chips is very high. Going through the IO die and to another chip is about as expensive as going to main memory.
Hyperscalers are running web servers which is a different story. But if you're running web servers you might be better off with Graviton in perf/$.
Per socket performance scaling is higher for equivalent tier sockets. At hyperscale that goes back into the price benefit (buy and maintain less physical data center) but for an individual server workload or individual user that also turns into a performance benefit, particularly for non NUMA aware workloads on the server side and just plain availability of such core counts for performance on the desktop or workstation side.
PCIe wise you get about twice the lanes (128 total on AMD) of even a 40 core 8380 in a the base 8 core model of Epyc or a Threadripper workstation CPU.
A place Intel still wins is total NUMA scaling. For a NUMA aware app like SAP HANA Intel can scale to 8 sockets while AMD currently tops out at 2 so you can reach about 2x as many total threads that way.