A slow algorithm could provide just as much security at 100 hashes/second as a fast algorithm at 1 billion hashes/second. Hashes/second does not matter, what matters is how difficult it is to achieve 51% of the network hashrate.
How difficult == how much power or hardware do you need. Therefore, for achieving the same level of security, you need to invest the same amount of real-world capital in form of electricity and hardware. Security is directly proportional to 'waste'.
On the other hand with CPU mining, CPU's are complex and difficult to feed (you often won't find more than one on a motherboard) and the capital (CAPEX) versus power cost (OPEX) ratio is very high. You simply can't easily scale up to 100 computers (the equivalent of 1 asic hashboard) very easily. Especially when these large miners can't dominate the market, everyone with a laptop from angola to zimbabwe will be competing with them.
I understand that ratio, that makes sense. However, your numbers are a bit off. If we assume a 3kW miner that costs roughly 3000$ (it was over 10,000$ at the top).For $10k, you can buy 10 rigs consisting of a motherboard, 12900KF and a PSU. The chip pulls 250W, probably 300W at the wall; achieving a similar ratio as with the ASIC.
With current ASIC prices (same machine going for $3k), the ratio may be 3x better on CPU mining, but it's not orders of magnitude better.
That's assuming nobody will build mining motherboards with 4 chips each or something like that. You know... like nobody expected SHA256 ASICs at the start.
