For some reason the article's explanation about "software ate everything" and jumping between each company isn't resonating intuitively with me, so I offer my own take on the situation:
For a long time, the hardware/data center offerings of Intel, AWS, etc. were close enough to what other companies needed to not make it worth their while to invest in inventing their own solutions to low % compute problems.
However, as:
-- compute loads and costs grew
-- types of compute became more specialized
-- design and build of one's own silicon became more accessible/differentiable
-- companies' needs and tolerances for paying a premium diverged enough from what Intel/AWS was offering,
it then became worthwhile for large companies (who can sustain such hardware development efforts) to design and build their own chips, either for cost reduction or functionality-enhancing reasons. Maybe they just saw the margins being achieved by sit-on-your-hands incumbents and decided, "we could do this too, and better".
This is a great summary. Sorry if it didn’t make sense. I think whats also interesting is it has become on one hand more accessible but on another less. Even with ARM cores the cost to get to 5nm is staggering and can really only be done by a few companies. It’s very obvious who has the scale.
Heck even semiconductor companies themselves are worried if they can make it to the 5nm node alone - the XLNX/AMD and especially the IPHI/MRVL merger can be viewed from this light.
What is most interesting is the timing. Software did eat everything, hardware used to keep up. Moore’s law died a quiet death and the solution is everyone makes custom chips for everything. Well then the people who have to most compute are going to make the most chips, but also at a time where the actual entry to do leading edge is the highest. It’s just a very cool time for them to have done it. I’m excited and will be following. Thanks for the feedback, the writing is hard sometimes.
For a long time, the hardware/data center offerings of Intel, AWS, etc. were close enough to what other companies needed to not make it worth their while to invest in inventing their own solutions to low % compute problems.
However, as:
-- compute loads and costs grew
-- types of compute became more specialized
-- design and build of one's own silicon became more accessible/differentiable
-- companies' needs and tolerances for paying a premium diverged enough from what Intel/AWS was offering,
it then became worthwhile for large companies (who can sustain such hardware development efforts) to design and build their own chips, either for cost reduction or functionality-enhancing reasons. Maybe they just saw the margins being achieved by sit-on-your-hands incumbents and decided, "we could do this too, and better".