I think its just the beginning. They are testing the market for their own silicon and reception from market. They are also going to need a userbase for Graviton products as well so portability in the beginning is important. This is true in any product. In the early days, increase market and mind share. So you need to make it as easy to adopt as possible. Then, add differentiation, so you product stand out from the others to continue attract users. Then slowly add some barriers for people to leave your product for a competitor's product. You want to increase moat and protection against competitors.
I don't think Amazon will throw in the resources to build a chip just to offer cheaper prices than competitors alone. Pretend you are Amazon. Remember, the cost of building a chip is similar between you and another guy. And if you are getting less return on investment compared the other guy why bother? It has to be that getting into chips will give you more return on investment than the other guy. But return on investment comes from many forms. For example, using graviton to run VMs is cheaper and leading your customer to use AWS more and less likely to leave AWS is another form of return. In this strategy, you need to build up the barrier for customers to leave. You could do it by making the stuff customer use with their VMs have lock-in effect, for example databases or storage that only runs or works better on AWS. Or you can directly make sure things customers run on Graviton cannot be run on anywhere else.
By the way, a better hardware performance is hard to maintain. If everyone uses the same instruction sets and the same manufacturing process, the only way you can be faster than the other guy by better circuit design. But there is only so much you can squeeze out of a circuit design and its really hard to maintain a sizable lead for many many years. Look at how AMD designed a better circuit and Intel's competitive position is gone. Changing the way software interact with the hardware is next easiest thing to do, and especially easy if you have a lot of leverage over software, OS, APIs, programming languages and developers. Basically what companies like Amazon, Microsoft, Google and Apple all excel at. See the addition of matrix operations for ML significantly increased performance. And how Apple can use their control over iOS APIs to build CoreML, get developers to use it, and have add special instructions and function blocks to their hardware to accelerate core ml. It will be interesting to see how this develops, but it would be only natural for these companies to explore this direction.