There was a post on HN last week where the author provided a reasonable answer (or at least I think so) to this question.

The gist is that people who are willing to pay a subscription to disable ads are exactly the people that advertisers would like to target. When a service assembles a list of such people, the value of that list to the advertiser generally exceeds the sum of the individual payments provided by the subscribers.

You can increase the subscription fee, but then you'll have less subscribers and the ones who are left will be the most valuable to advertisers (i.e. they have the most disposable income). If you decrease the subscription fee, you will have more subscribers but not a lot more, because the primary obstacle for online subscription type services for most people is not the price but the idea of paying for something they are accustomed to getting for free.

So if you get 1000 people to pay $1.00 per month for your niche Swedish grunge music streaming service and double the price, a decent number of them will not like that and some of them will unsubscribe. But if you halve the price to $0.50 per month you won't see many new customers since many people aren't willing to pay even small amounts for music streaming. Advertisers, meanwhile, don't have these mental obstacles and just try to price things as objectively as possible. And they see a narrowly targeted list of people with a highly correlated list of interests and purchasing tendencies and value it appropriately.

This might explain why services like Hulu have gradually introduced more advertising into their paid subscription services. As long as two groups of people are are paying them (customers and advertisers) one will generally be willing to pay more. Economic forces on the internet seem to result in advertisers having more purchasing power here.

Do you have the link to the HN post you referred to?

Disclaimer: I'm making an educated guess here.

Most non-text media sites sell advertising directly, because most of the on-demand ad networks / platforms that provide a real-time auction are simply too slow. I've seen ad stacks that spend over 4 seconds determining which on-demand ad network will fulfill the view - usually accomplished via VAST chaining. VAST 'tags' are xml manifests for an advertisement, and can contain links to other VAST tags in lieu of supplying the ad unit data directly - and each hop on the VAST tag has the chance to drop a cookie on you.

Because ad networks suck for the user, selling direct advertisements is better. Because ad networks suck for CPM, selling direct advertisements is better.

Two primary metrics used to directly sell direct advertisement is viewership and consistency. Advertisers not only want to buy lots and lots of eyeballs, they want to be sure the entity selling those eyeballs can actually fulfill the desired number of eyeballs. The details of an ad buy can take quite some time to work out.

If the cost in man-hours to work a deal for $X eyeballs is more than CPM*($X/100), a company won't even try to sell the deal. Likewise, if the benefit of having $X eyeballs view your advertisement is less than the cost of your ad buy team's time plus the cost of the advertisement itself, you won't buy that placement. Since the ad buy team and the ad sales team's efforts are fairly constant with respect to the size of the deal, small deals are verboten - on both sides of the table.

And if you've got subscribers that don't see ads, that's leverage you don't have when you're trying to sell ads. And since your subscribers are probably more affluent, they're the audience your advertisers probably want to subscribe to anyways.

As soon as 'Ads' becomes the primary revenue driver (unless you're an advertising company), you've failed. IMHO.

This is spotify's approach.