Amdahl's Law, modernized, reveals a phase transition.
The classical version divides work into serial and parallel fractions. The modernized version (arXiv:2603.20654) divides resources between specialized accelerators and programmable compute, with efficiency differences and workload structures that determine how marginal hardware investment translates into value.
The key finding: optimal investment in specialized hardware doesn't decline gradually as workloads become more scalable. It drops to zero at a finite threshold. Below the critical scalable fraction, specialization wins — dedicated chips outperform general processors by enough to justify the investment. Above it, specialization becomes suboptimal regardless of how efficient the dedicated hardware is. The transition is discontinuous.
This explains a real puzzle in computing architecture. Domain-specific AI accelerators — TPUs, neural engines, custom ASICs — are measurably more efficient per watt and per dollar for their target workloads than GPUs. Yet GPUs dominate AI compute. The standard explanation is ecosystem lock-in, software inertia, market timing. The modernized Amdahl's Law offers a structural explanation: beyond the scalable-fraction threshold, the physics favors programmability over specialization, regardless of efficiency advantage.
The phase transition comes from an interaction between two properties: efficiency (how much more work the specialist does per unit resource) and scalability (how much additional value each marginal unit of compute provides). When scalability is high, programmable hardware can absorb arbitrary amounts of compute into the scalable workload. Specialized hardware can't — its fraction is fixed. At the threshold, the programmable pool's ability to grow overwhelms the specialist's ability to be efficient.
The deeper point: Amdahl's Law was always about the boundary between what scales and what doesn't. The modernized version shows that this boundary isn't just a limit on speedup. It's a phase transition in optimal design.