Re-posting this because otherwise I think it'd get lost under the OP, which some people might, not unreasonably, instantly skip by subconscious reflex. (I didn't even endorse it to the point that you would pay attention to it just because someone who almost never posts AI content thinks it's worth highlighting.)
Messing with Suno for the first time, I noticed that, even in the free tier, Suno was generating songs in less time than it takes to listen to one. We've already had discussion of music generation that can massively flood the industry. Indeed, we were already past the point of "more music worth listening to is written than a human could ever hear, even if they did nothing else" while music was still only written by humans. It is a matter of taste as to what if any AI music counts as "worth listening to" and in what sense, but most music is not heard with a high level of critical artistic analysis to begin with, like what you listen to while driving just because it's on the radio and good enough, or low energy synthwave while studying.
At least, AI music has reached the level where you can at least believe a human wrote it with workmanlike skill and thought it was good enough to publish if you're not actively searching for tells. (Well, there is a certain sense in the lyrics that I feel like I can hear sometimes that I can't describe, that tells me, "these are AI". But who knows how long that will last, or if humans writers become more like that from AI influences in their media environment.)
But the prior volume, AI or human, still assumes you're listening to the same music as others: one song is played many times. But we're now at the level where a sufficiently funded Suno account could produce an endless stream of songs that are generated and then played once.
But how much funding is "sufficiently funded"? Imagine a 24/7 station that generates songs that are played once and then deleted. Songs cost 5 credits (well, 10, because Suno always generates 2 songs on each attempt, but we'll imagine you can negotiate this even as we're assuming full retail price otherwise)…
(4,000 credits / 30 USD) * (1 song / 5 credits) = 80 songs / 3 USD
Max song length is 8 minutes but 3.5 looks closer to average (tho symphonic metal is somewhat longer on average) by eyeballing it, let's call it 3 minutes.
(80 songs / 3 USD) * (3 min / song) = 80 min / USD
(1 USD / 80 min) * (525,949.2 min / 1 Gregorian year) ≈ 6,574 USD/year
These are full retail and considering only extra credits, the subscription credits would reduce it somewhat. Anyone that this volume might be able to negotiate lower, but credit costs are also probably more closely linked to raw compute costs since you can't subsidize or loss lead when a commercial customer picks their volume. On the other hand, constant fixed usage on multi-year contracts is ideal for capacity planning and amortization of TPU capital.
So, about 6.5 k$/year is the current cost of a continuous stream of bespoke music at the current state of the art. This is as expensive and as low quality as it will ever be.