Even if the quality drops off so that it's only as good as say, 1.5 well-rested workers, they still got an extra 0.5 for free.
Most of software is invisible, and would take a tremendous amount of effort to analyze well (for quality, value, etc.), only a fraction of which is ever expended in the moment (if at all). The reality is that good people work under poor conditions and do poorly, other people get burnt out, other people with talent "evaporate" out of the team/org/company because they are fed up or what-have-you. All of this degrades the ability of the team/company to execute on difficult projects well, but that is incredibly hard to measure objectively. Especially since every iteration and every project tends to have its own unique qualities which make direct comparisons to previous projects non-trivial. That leaves only subjective judgments. People may look around and ask "why can't we execute as well as we used to?" but unless it's a drastic difference it may not be very noticeable, if at all.
And that's the thing, objective measures are nearly impossible to come by, and a lot of the data for subjective measures is not well known beyond only a handful of people. Software is fundamentally a creative endeavor, and this shows the most during the development process. Judging the quality of a company overall from year to year then becomes similar to judging the quality of, say, a band or an actor or a comedian. Debating which, say, Smashing Pumpkins album was the best is fundamentally similar to judging whether the developing capability at, say, google (or microsoft or facebook or amazon) was better in 2012 or 2010 or what-have-you. And this is even difficult to judge over long periods (say, time slices of an entire decade) when looking back years after the fact. It's essentially impossible in the moment, especially in micro-cosm.
All of which is to say, there's no sure, objective way to tell whether you're getting 1.5x well-rested workers worth of work, 2x, 10x, 0.5x, 0x, -1x, or -5x. All you know for sure is that they're putting in more hours or "crushing" more bugs, so you focus on the metrics you do have even if they are bad metrics and could lead you to make all of the wrong choices.
Could this be a way of passing the cost to another department? Maybe a department manager is a slavedriver and knows it's not efficient long term, but he doesn't care as long as the costs (like replacement recruiting) don't come out of his budget. I don't know if that's how budgets really work though.
I think the NYT reported that the average tenure of an Amazon worker is 1 year. Just let that sink in for a minute.
Amazon is at the higher end of unsustainable, like most of the gamedev industry (for example), but on the company scale there's always more workers feeding into the top of the funnel, so they can still get the work done, they don't care that they're churning through employees.
Nevertheless, fundamentally most tech companies work on a similar model, they're just not as horrid as Amazon so the process isn't as obviously absurd. But it's still a broken system that I don't think anyone really knows how to fix, yet.
What effect does that churn have on their technology? If the average developer only sticks around for one or two years before burning out, subject-matter experts are going to be really scarce, and once you don't have any of those you lose the ability to innovate in any meaningful sense.
Why has every article in the last week assumed that the Amazonians working 80 hr weeks are getting twice the work done? Has every journalist and blogger in the country collectively forgotten the concept of diminishing returns on hours worked?