A dimension can contain a SQL expression over one source row. Model profit as revenue minus cost, then make it available to other fields through a LookML reference. Subtracting a cost total from each individual row would mix grains. Keep the dimension numeric. The editor accepts references to source columns and to fields, and rejects unsupported SQL instead of guessing its result.
Use the field picker and the generated result together. The warehouse contains repeated customer keys, orders with several items and missing values. Those rows make a plausible-looking total insufficient evidence. Run the requested definition, inspect its grain, and use the feedback to locate the mismatch. Every dataset here is fictional.
Add a reusable row-level profit dimension.