And it really did not scale. If you got too many computers on one Ethernet network together, things would start to really bog down.
We would address this by splitting an ethernet network into two. When it got too slow, we'd take our network of (say) 1.2.3.0 - 1.2.3.255 and split it into two networks, 1.2.3.0 - 1.2.3.127 and 1.2.3.128 - 1.2.3.255. Each of these ethernet networks would get its own separate port in the IP router. (Which could require adding cards to the router. And put extra processing load on it.)
DHCP did not exist yet, so we had to manually go to each computer and change its broadcast address, subnet mask, and default route.
This was disruptive, so we tried to keep all addresses in the lower portion of the address range so that they would continue to function well enough during such a transition. For example, a host at 1.2.3.10 with subnet mask 255.255.255.0, router address 1.2.3.1, and broadcast address 1.2.3.255 would still function when we changed its ethernet to be 1.2.3.0 - 1.2.3.127 instead of 1.2.3.0 - 1.2.3.255. Its broadcast address and netmask were wrong, but it wasn't completely offline. (By contrast, a host with address 1.2.3.200 would be out of range for the new network parameters and would have problems immediately.)
Then Grand Junction came up with the idea of ethernet switches to isolate traffic, and we didn't get one because they were a bajillion dollars. I don't know the exact prices, but I would guesstimate something like $5000 to $10000 for a switch with around 10 ports.
But eventually they came down in price. With an ethernet switch, everything was amazingly easy in comparison. It basically did the same thing I'm talking about (splitting an ethernet into pieces so the devices don't try to talk over each other), except instead of it being a weeks-long manual process, it happened automatically in real time.
Once switches became standard, ethernet was pretty scalable. Broadcast traffic still isn't scalable if it's used too much, but otherwise it basically works fine.
(Also, ethernet bridges were a thing at the time, and they are basically 2-port switches. But arguably a switch with a bunch of ports was a different way of thinking about how to do an Ethernet network.)
I think the idea was that anything more than this was likely to lead to long enough round-trip times between devices at the far ends of the Ethernet that the CSMA/CD algorithm couldn't be guaranteed to work.
Respect!