I don't want to disappoint you, but what you want to do is not really supported by many companies anymore. Maybe the cloud providers would be interested in someone close to the metal. A few companies still run their own data centers, but even for them, hardware is so cheap it's usually easier to just throw more hardware at the problem than optimizing what you've got.

The one area where you might want to focus is HPC. All the companies building huge GPT models need highly optimized hardware.