Placement Is the Decision Routing Can’t Undo
Every PCB engineer has used an autorouter, or at least argued about one. Almost none would hand a real board to an autoplacer. Routing got automated decades ago. Placement stayed in human hands.
That neglect matters more than it looks, because placement is the one decision routing cannot undo. Put a power management IC two millimeters too close to a sensitive analog front end and no amount of routing cleverness undoes the noise coupling. The geometry is already wrong. A routing mistake stays local and fixable. A placement mistake propagates through everything downstream, surfacing later as congestion, as a timing problem, as the late component shuffle that blows the schedule. Placement constrains routing. Routing cannot fix placement.
So why is placement still done almost entirely by hand? Not because it is unimportant, and not because the math is uniquely hard. The bottleneck in PCB design was never only the algorithm. It is capturing what the engineer already knows.
The Real Bottleneck
Take routing, the half that got most of the attention. We built a routing engine that handles real production boards, and it took years to get there. But for all that, the thing standing between engineers and the result was not only the algorithm. It was everything that had to happen first: defining net classes, marking differential pairs, declaring what must stay clear of what, telling the tool what the board is actually for. The engine was ready. The intent was trapped in the engineer’s head and in a schematic the tool could not read.
That is the pattern under most PCB automation disappointment. Not that the engine is easy, it is not, but that even a great one is inert until someone captures the intent it runs on. Get that capture wrong and the best engine in the world looks broken. Get it right and it looks like magic.
Why Placement Resists Automation
It comes down to the same thing. Placement’s intent is almost never written down. Routing runs on explicit, measurable rules: trace widths, clearances, layer assignments. An algorithm can check a route against them and know whether it passed. Placement runs on intent that lives nowhere the tool can see it. Which parts belong near each other, which must stay apart, what the engineer is trying to protect. Those relationships are implicit, often in tension, and they come from understanding why the board is built the way it is. The typical autoplacer does not fail because the math defeats it. It fails because it is placing without ever knowing what the engineer meant.
What Changes Now
Two things make this tractable in a way it was not before. First, models that learn spatial patterns from real boards: train on enough designs and a system starts to recognize that decoupling capacitors belong near power pins, not because someone wrote the rule, but because every good board does it. Second, and bigger, language models that can finally read the parts of a design that never reached the layout tool: the schematic structure, the datasheet notes, the intent sitting in the engineer’s head.
That second piece is the one that was always missing. It is the same unlock in placement as in routing: capture what the engineer knows, in the form they already express it, and give the algorithm the context it never had. Building the optimizer was a frontier of its own. Reading intent is the one still open.
The Tool Carries the Load
Capturing intent is not the same as taking over, and the difference matters. The intent being captured is the engineer’s. Placement is the part of the job most engineers want to own. It is where judgment lives, where experience shows, where the decisions that define a board get made. A good tool does not replace that. It carries the load that does not need a human: reading constraints out of the schematic, surfacing conflicts early, turning intent into a strong starting point instead of a blank board the engineer pushes around by hand.
The engineers who consistently ship working boards on tight schedules already work this way. They spend their judgment on the decisions that matter and want the rest handled. A tool worth using respects that line.
This is how we see it at DeepPCB. The engine took years to build, and it works. It was never what held engineers back. What decides whether any of this is usable is capturing what the engineer knows and handing it to the engine. Routing is where we saw it first. Placement is the same problem, one step upstream. The industry spent sixty years building better algorithms. The board was always waiting on something simpler: someone to listen to the engineer.