The issue is also we don't know what to write. I'm being a bit extreme here, but not being able to do rational design limits things to mostly 1) build a lot of

The issue is also we don't know what to write. I'm being a bit extreme here, but not being able to do rational design limits things to mostly 1) build a lot of variants, using some educated guessing, 2) test those variants, 3) plug those results into some model and predict what to build next. So the value of each individual variant is relatively low, since you don't know if it will work or not. Plus a lot of other investments have to be made into the toolchain (developing assays for the thing you actually care about is not easy). In your plateaus, the possibility space explodes with each step function. We are already at the first rung of the ladder of building many 50bp variants. Generally, you'd probably want to be able to conduct an experimental cycle with 1000s of variants. At the upper end, you'd want to pay $5000-10000 for each cycle. There is demand for making protein-length sequences, but the prices the market is willing to pay is probably closer to $10/gene (fully customizable 2kbp) rather than the current price. Another thing is that not all sequences are created equally. Some sequences will just not synthesize well (maybe there's secondary structure), so to limit costs one ought to limit the number of retries per sequence.