Here is the material-science heart of the ternary thesis. Multiplication is not expensive because arithmetic is hard — it is expensive because an n-bit array multiplier is a two-dimensional grid: ~n^2 AND-plus-adder partial-product cells laid down in silicon. That quadratic block is the single densest thing in a MAC unit, and it is precisely what forces you onto a leading-edge node — you need the smallest transistors money can buy to pack n^2 cells into a tolerable area and power budget. Now make the weights ternary. A weight of +1 says 'add the activation', -1 says 'subtract it', 0 says 'do nothing'. There is no partial-product grid at all — just one n-bit add/subtract lane that scales as ~n. For n=8 that is 64 gates versus 8: an 8x collapse you compute live below, and the gap only widens as n grows because you are trading a quadratic for a linear. Delete the quadratic block and the area/power pressure that demanded 3 nm evaporates. A mature, sovereign, high-yield node — 22, 45, 130, even 180 nm, with cheap mask sets and MPW shuttle runs anyone can buy into — is now enough to do real edge inference. Be honest about the frontier's edge: exotic ternary-logic DEVICES (CNT-FET with adjustable Vth, tunnelling TCMOS, semi-floating-gate) exist but are early, so the near-term win is DIGITAL ternary weights plus in-memory compute on ordinary CMOS, not a new transistor. Recall that CMOS already spends 2 devices per bit — pull-up and pull-down — buying noise margin; ternary is not fighting that reality, it is removing the block that made the reality unaffordable. The barrier was never physics. It was the multiply. Take the multiply away and the fab stops being a gate.