How much does it cost the publisher to publish an academic article?

Ubiquity Press breaks down their £300 ($500) APC as follows:

• 38% indirect costs for things not related to the publishing of a single paper but which are needed for the business (£114 or $190)
• 34% covers editorial and production aspects, which appears to be the costs associated with producing the paper, managing submissions, responding to authors, preparing proofs, typesetting, XML etc. (£102 or $170)
• 16% is a waiver premium charged so they can offer 0 or low APCs to people who genuinely cannot pay (£48 or $80)
• 8% is used to pay for indexing, archival (in case they go bust), DOI etc (£24 or $40)
• 4% goes towards costs of billing you and taking payment (£12 or $20)

Depending on what you consider to be the actual publishing costs (here probably the 34% editorial & production costs + 8% Indexing & Archiving) you would be looking at ~ £126 or $210.

Ubiquity don't break their indirect costs down into server/platform costs; this all goes into the 38% indirect cost column.

I know this doesn't refer strictly to the final version of a paper, but the arXiv pre-print server provides a useful bit of information to contribute to this discussion. According to its website, it receives around 76,000 publications per year. Its operating costs are on the order of $826,000 per year.

You do the maths, and it comes to just over $10/article. This is without any of the bells and whistles that come with traditional publishing, but provides a nice baseline estimate of what it takes to publish a research article online.

It depends to an extent on how technically-savvy the author community is, and thus what services they need or do not need to be done for them.

For computer science journals, the cost of production is extremely low because authors can typically be expected to do their own typesetting.

An efficient, peer-reviewed, top quality journal can thus be run at a cost of just $6.50 per paper. There is an excellent, detailed breakdown of this figure given by Stuart Shieber about the Journal of Machine Learning Research (JMLR) here.