I feel like I need to point out a few things about Amazon corporate culture, which are really relevant to the discussion here.
The core of Amazon as a company is that they vertically integrate everything by implementing competitors to all their dependencies, use them internally, and make them available as a paid service to outsiders. They are extremely predictable at this -- if they spend a lot of money on something outside the company, then that will be a new business unit.
The reason this cultural part is so important is that not all decisions are made to be cash-optimal. Even if they chose to not do it now, no-one at Amazon would ever get fired for daring to suggest that they should make their own CPUs. The decision wasn't really ever an if, but a when. What other dependencies should Amazon replace before tackling the CPU problem? It seems like they have finally worked down the list and it's CPU time.
I don't see how Amazon can internally justify an A72-based ARM server CPU in 2019 (or beyond) when price/performance and performance/watt will be so much better on products from other companies.
There's no sense whatsoever to bet on a A72-based server CPU in 2019, and that's likely not what they are doing. Amazon is one of the most methodical and least quarterly-results-bound large companies in the world. They understand that large undertakings take a long time, and should be done in stages. The job of the current CPU is likely to help mature the software and the platforms in advance of something more substantial in the future.
Hell if they want to go ARM, they would be better off getting in bed with Marvell/Cavium than trying to roll their own (Huawei is not really an option despite having a technically excellent 64c ARM server CPU).
Amazon doesn't really do partnerships like that. Jeff Bezos is approximately the world's greatest control freak. Buying up an external company to set up as his newest business unit would be entirely within his m.o., but the key part is that he will never make a deal that leaves even a sliver of real agency to his "partners". The rule is that what Bezos decrees is law, and if that's not okay with you they'd rather not deal with you and build up a competitor to you instead. Marvell is a bit on the large side for an Amazon acquisition, but it would not be unprecedented. Maybe if the internal CPU project is a dismal failure?
That's basically what Huawei did (with help from their HiSilicon division which had experience developing mobile ARM CPUs). I still get the sneaking suspicion that Amazon may throw in the towel, though. History is filled with the wreckage of failed CPU design efforts. I would say that "time will tell", but in Amazon's case, how would we ever know? They probably wouldn't release specs on anything they designed strictly for internal use, and ODM sales are almost entirely "under the table". We won't know if they continue buying hardware from outside vendors, either.
But do you know what history is not full of? Abandoned Amazon product lines. Just as no-one would get fired for suggesting building a new CPU to Bezos, everyone would know exactly what the two-word reply from Bezos would be if you went up to him and said that "This was too hard, we should give up." Failure at Amazon means they try again. And again. They have the resources, and so far they've managed to brute-force even the hard problems.
With this, I don't mean to describe Amazon as a flawless, unbeatable juggernaut. They have plenty of their own problems, and it's entirely possible that they will have their CPU project crash and burn so many times it becomes a running joke. It's just that their culture and flaws are sort of the inverse of Intel in the past decade. They don't start a thousand projects only to shutter them down after the first iteration turns out to be a failure. When they fail, Bezos fires some people, reorganizes things (often seemingly as if he did it entirely randomly...), puts new people in charge and they try again. For now, that approach has eventually yielded results.
Facebook
Twitter