Make money doing the work you believe in
Sharing my reflections from Jonathan Ross’s (now at NVIDIA, founder of Groq, architect behind Google’s first TPU) comments on memory at his recent interview at Sohn.
There is a tension in memory right now.
Jevons Paradox keeps memory alive. If it is cheaper, there will be more demand. So demand will always expand to fill the available supply.
But if memory gets too expensive for too long, people don’t just buy less, they will engineer around it entirely (via software and/or hardware).
Jevons Paradox only wins if supply grows fast enough to stay ahead of the pain. Engineers keep building memory-hungry things because they can.
But if scarcity persists, the best engineers stop treating memory as a resource and instead treat it as a problem to eliminate, focusing on the constraint. Scarcity forces efficiency.
By protecting record-high profit margins by constraining supply and not building capacity fast enough, the memory players accelerate the engineering movement that makes their product less necessary tomorrow. They win the pricing battle, but lose the demand war. This is where they differ from TSMC.
Jevons Paradox only works if we feed it. Starve the market long enough and Jevons flips. If demand doesn’t expand to fill supply, it could start contracting to escape it. That’s why the memory players have to build more fabs faster, don’t give the smartest engineers in the world the chance to make you obsolete.
Example: DeepSeek-V4-Pro has a 90% reduction in KV cache memory burden compared to DeepSeek-V3.2.
