"but I think that the commodity level of most hardware still makes it super unlikely."
"If a company is willing to use Office 365"
Oh I'm with you on this. Steve Walker's Computer Security Initiative and the Orange Book gave us lots of highly-secure stuff for defense, etc. They got rid of that for cheap, fast, fully-featured COTS. The same happens in business, aerospace, etc. The exceptions are usually pre-made appliances or, in aerospace, better components in the DO-178B Level A stuff. Much of it is shoddy. A good chunk of the defense fabs' business is probably replacing legacy parts in old equipment at prices guaranteed through corrupt contracts. They don't give a shit about security in general: just money. ;)
"But when it comes right down to it, there's billions of weak points at much higher levels and much more easily compromised than a chipset or compiler."
There's lots of weak points. Stopping code injection from all known vectors with simple, proven techniques at CPU and language levels eliminates the whole malware problem if apps are whitelisted, built from source, and include no executable scripting/JIT. There's ways to conveniently enforce POLA within a system (eg CapDesk), do secure (even automated) configurations of networks (Boeing's Survivability Grammers), and so on. There's components for most use-cases just waiting to be productized, integrated and sold to larger audience. Given this is 90+% of attacks, it's certainly worth pushing to establish a stronger baseline for companies that want less loss of secrets, availability, data, etc.
There's only so much the traditional methods like coaching and monitoring can do if one can simply open a folder (not a file!) that immediately results in full-control of machine by malware since a thumbnail rendered. Endless crap like that exploiting underlying foundation of quicksand. The HW and SW of endpoints, at least at lowest layers, need to get in check for that other stuff to be meaningful. I'm also in favor of an integrated networking stack that makes different applications, even if using TCP or HTTP, look visibly different at the packet level so NIDS spots weird patterns more easily. Like the MLS extension but not MLS policy itself. Do it at application layers with stuff like Ethos's eTypes or security-enhanced ZeroMQ where developers don't worry about plumbing much.
" It's a good academic exercise though."
It's also an industry bringing in tens of millions of dollars at least. That's with costs that are too high, lack of key software support, and little to no advertising. I imagine it could be larger than tens of millions with such obstacles reduced or eliminated.