They just realized very early which way to go. If others (AMD, Intel) saw writing on the wall earlier, we would have had competition. This way, it will take years for them to catch up.
Wrong. AMD had HSA based SoC back ten years ago. The problem with this approach is the Memory is Fixed width, a.k.a not expandable or scalable. The M1 has a subset of processes and applications spaces it shows promise in single core performance.
In reality, every applications running today is multi-core/multithreaded and large applications that require large data sets and are memory and computational intensive this CPU will not be used on. It would crash.
We have a fixed 16GB shared memory space. That's it. I routinely run 30GB rendering samples. This CPU would crash in Blender.
AAA games now requiring 12GB in dedicated VRAM for 4K Godsplay this CPU would crash and lock up the entire system.
These systems are disposable consumable, low power, low processing little efficient boxes for basic Web Browsing, Office Work. Heavy intensive, high core/high thread, memory intensive applications that thrive the more resources you throw at them--Computational Fluid Dynamics, Finite Element Analysis, large data sets that need to be stored in memory none of these and more will work with the M1.
If you're betting this solution has some linear scalable 1:1 capability of adding more cores, more shared memory as if the memory is limited on the SoC by your imagination then you'll be sorely disappointed.
Back in 2019 when Apple Demoed the Mac Pro onstage with 1024 channel strips running a full 100 unit orchestra w/o taxing a 28 core Xeon with 56 threads, an Afterburner and Dual Radeon Pro Duos with 128GB of HBM memory they could max Logic Pro [X at the time 10.5] without barely pushing all 28 cores, but it utilized all 28 cores.
That same score wouldn't even load in any of today's or for years to come M series SoCs. They'll have to offer a completely different Workstation class set of chips for that to ever become a reality.
By the way, AMD's Zen 4 is including their own Neural Engine FPGAs drawing upon the Xilinx merger, their own Tensor Cores, RDNA 3.0 based 5nm fab SoC that won't be limited to 16/32/64/128GB of HBM2e memory [that they can leverage even now seeing as they were first to market to use such]. Zen has a present limit of 2TB and the Zen 4 on DDR5/LPDDR5 will expand that memory footprint for the supercomputing markets, big data center markets.
People seem to delude themselves that this SoC is the future when it's a specialized solution.