If you can build small runs (<50base pairs) you can make small mutations to dna, or read particular sections of dna. If you can make dna larger than the average protein (~2000bp) you can invent new proteins from scratch rather than modify existing protein sequences. If you can make dna longer than a plasmid (~10,000bp) you can creatively invent a minimal viable replicable and deliverable unit (a plasmid (bacterial virus)). If you can do millions if base pairs you get to chromosomes and can invent Eucaryotic-transmissable storage.
But until you leap those plateaus, you can likely saturate the intervening market. So even if there’s massive pent up theoretical demand for the wholesale invention of genes, there’s no way to really demonstrate it in the current market.
This inability to estimate demand may make it a tricky spot to invest in.
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.
“Our system’s write-to-read latency is approximately 21 h. The majority of this time is taken by synthesis, viz., approximately 305 s per base, or 8.4 h to synthesize a 99-mer payload and 12 h to cleave and deprotect the oligonucleotides at room temperature. After synthesis, preparation takes an additional 30 min, and nanopore reading and online decoding take 6 min.”
Also the amount of data that can be written is very small in general for oligosynthesis. The most high throughout methods are microarrays by Agilent and others (not used in this paper). You can buy about 32 megabits for $6000 (http://www.customarrayinc.com/oligos_main.htm). The actual cost of synthesis is maybe $2000 as a guess.
So currently DNA storage would be good for small datasets that you would like to store for a long time. Physical density is useless most of the time if you can’t efficiently generate a lot of data to begin with.
We would need several orders of magnitude increase in the write capacity, which is slowly being worked on. Typically people would like to compare synthesis costs with Moore’s law, using Carlson curves ( http://www.synthesis.cc/synthesis/2016/03/on_dna_and_transis... ).
However there hasn’t been as much progress in synthesis as in sequencing. Why? My theory is that Theres not a very big market for DNA synthesis, so the big investments needed haven’t really been there. Maybe storage on dna could be that market, but it would need to show quick and easy wins (eg stepping stones of practicality like what early integrated circuits had).