Physical isolation, control of whole environment, and more security/reliability/predictability when these are leveraged properly. You also have less legal risk if the FBI decides to kick in that colo's doors as they have in the past and seize shared servers just because one client was on it. I doubt the quantitative risk of this is high but I like risks not existing where possible.

And you're never investing in IT if you are doing IaaS or PaaS: you're just paying a third party for a temporary good. Colocation is investing in IT because you own the equipment and what's in it. The other parts are their problem with them having a financial incentive (eg competitiveness) to improve them over time.

Cloud can shut down tomorrow and what do you get out of that? If it's my PC with my data, I can get it back from them somehow and people trying to stop that might face criminal charges. A VPS service shutting down means my system probably ceases to exist unless they have specific provisions for the situation that work during bankruptcy. You might know of them since you clearly study the cloud offerings more. I don't so I stay in a low-risk situation.

So you don't want to use AWS because the risk Amazon will shut it down tomorrow?

I like how you ignored the most important stuff (first sentence) and a peripheral risk with precedent (second sentence) to focus on the lowest concern I had. Plus, that concern applies to more than Amazon: many vendors out there. Nice troll tactic, though.

You raised the point. I just called attention to it.

Also, ad hominem.

To just that. What of the main points of dedicated vs VPS?

Physical separation from other users defeats virtualization attack surface and many, covert channels immediately. Control down to hardware layer lets you do neat improvements to performance (eg custom drivers), maintenance (eg transactional kernels), and security (eg CheriBSD on FPGA-based CHERI processor). I also pointed out you can get more predictable with the implication of using a RTOS or other deterministic software + hardware combos. Heck, might even send in AS/400's, NonStop's, and/or OpenVMS clusters on cheap Alpha/Itanium servers to get uptime you haven't seen in cloud world: 17+ years for one OpenVMS cluster.

Those are the strongest points that cloud can't touch at all so far. The cloud-style research on making something comparable with strong correctness arguments is in infancy and almost all academic R&D. Which implies things about what they're using now... Of course, the outages and papers at DEFCON etc already told us that, didn't they?