How realistic is it for a company to build its own JVM?

It depends a lot on what they mean by their own JVM. A clean-room implementation is a very different project from forking OpenJDK and changing a few things. Even an internally packaged build with company certificates could get described loosely as an in-house runtime.

A fork still needs maintenance, including taking in upstream fixes and testing changes. But starting from an existing codebase is a much smaller lift than building a competitive runtime from scratch.

Writing a JVM that runs some bytecode is doable. Writing one that competes with OpenJDK on compatibility, performance, and features is a much bigger undertaking. A production-quality JIT compiler and optimizer alone are substantial projects.

A narrow internal runtime might be viable if it only needs to run a known set of simple programs and doesn't need full compatibility. That doesn't make it a general replacement for an established JVM, though.

There are real examples of companies developing JVM implementations, so it isn't inherently impossible. The harder question is whether the company has a specific need worth the ongoing engineering cost. For most uses, an existing runtime or a supported distribution seems like the more practical starting point.

Check the licensing premise too. OpenJDK is distributed under the GPL with the Classpath Exception, which generally means an application running on it does not have to be GPL just because of that. Modifying or distributing the JDK can raise different questions, so the exact use and distribution matter.