Chains that welcome purpose-built chips end up with their hashrate concentrated in whoever can afford to have chips made. Chains that fight them spend their lives losing that fight, because somebody eventually builds the chip anyway.
We run both. Purpose-built hardware gets a lane designed for it. Graphics cards get a lane where they are the intended machine rather than the tolerated one. Each sets its own difficulty, and a limiter keeps either from taking the network.
How it works
The first three are how you run two lanes without one eating the other. The fourth is the reason there is a second lane at all.
Lane A is a conventional puzzle that purpose-built chips are very good at. Lane B is built so a graphics card is the machine it was designed around. They run at the same time, on the same chain, and each produces half of every month's reward.
The lanes do not share a difficulty setting. A wave of new chips arriving on Lane A makes Lane A harder and leaves Lane B exactly where it was. They are separate markets for separate hardware, and neither one's boom prices the other out.
Separate difficulty is not enough on its own. If one lane becomes far more profitable, machines migrate and the other lane thins out. So past a threshold, that lane's next adjustment gets an extra push upward. You can watch it engage in the animation above.
Lane B is memory hard, which is the usual way to keep graphics cards competitive. What is unusual is the shape: the arithmetic is the arithmetic of running a model. So pointing a card at this chain is not a choice between mining and doing useful work. It is the same card doing both, on price.
Why not just pick one
Every chain eventually has this argument. Welcome purpose-built chips and you get efficiency, durability, and operators whose money is committed to your network specifically. You also get a hashrate that concentrates in whoever can afford a fabrication run, which is a very short list.
Fight them instead and you are committing to a permanent arms race you will lose, because resistance to specialised hardware is a temporary property of a puzzle rather than a permanent one. Somebody builds the chip in the end, and then you have a chain designed around a promise it can no longer keep.
Running two lanes does not resolve the argument so much as decline to have it. Both kinds of hardware get somewhere they are the intended machine. Neither has to displace the other to earn, because the reward is split before either of them starts.
every retarget, for each lane:
share = that lane's blocks over the recent window
÷ all blocks in that window
adjust difficulty toward an even split
if share is over 65%:
push that lane's difficulty up further
the corrective is zero exactly at the threshold,
so it eases in rather than snapping onThe animation above is running that rule. Watch one lane climb past sixty-five per cent, the corrective engage on its difficulty bar, and the share come back down. The percentage beside each lane is counted from the ticker of recent blocks below it, so you can check it by eye.
The threshold and the window are the current design, not a settled number. Set them too tight and independent difficulty stops meaning anything; too loose and the limiter never engages when it should. Both come out of testnet simulation.