This is talking mostly about perceived "stars" in academia, who already do have tenured or tenure-track positions (or offers of them), but leave for industry in

This is talking mostly about perceived "stars" in academia, who already do have tenured or tenure-track positions (or offers of them), but leave for industry instead. So I think it's a different issue than the one you're talking about. For those people, one thing that's changed is that academia, in STEM at least, has grown into having a big management component. You're expected to run a significantly sized lab, and to continually bring in enough grants to pay for it. Professors at top universities spend a large amount of their time writing grants, hiring people, and networking with potential funding sources, and it's no longer really optional. There is a certain style of researcher who actually excels in this kind of position, managing a medium-sized research enterprise, and advancing a field by putting together good teams, providing high-level direction, and mentorship. It's like running a startup in some ways (obv. different in other ways). But not every researcher is good at or interested in that, and some would rather opt out and find a job that lets them focus more on their individual research. Another thing that's changed is pay. Academic salaries in CS have gone up significantly, to the point where $100k+ is common, and $150k+ for big names is also getting common, whereas professor jobs were traditionally more lower/mid-middle-class than upper-middle-class (more like "$75k+" than "$150k+"). But salaries at tech companies have gone up even faster. If a big name in deep learning gets offered $300k base plus a big stock package, the university's retention counteroffer isn't likely to come very close. Also, specifically in AI (which this article focuses on), data access and engineering/infrastructure support are other lures of industry. If you do machine translation, for example, moving to Google will solve many of your dataset problems. If you need a distributed cluster, there's expert IT staff working on it; if you do autonomous vehicles, there's expert engineering staff supporting your "demo platform". In academia you rarely can afford that kind of infrastructure or engineering support, so there's a lot more of grad students cobbling together DIY demos, and the postdoc hired mainly for his/her mathematics knowledge also moonlighting as the cluster admin.