Power is the real bottleneck.

The golden era: Dennard scaling

From 1974 to about 2006, transistors shrank on a predictable cadence, and something magical happened: as they got smaller, they also got faster and used proportionally less power. This is Dennard scaling. Every generation, chip designers got a free lunch: more transistors, higher clocks, same power budget.

Why it stopped

Below about 65nm, transistors started leaking. Current bled through the gate even when it was "off." You couldn't lower voltage anymore without leakage overwhelming the savings. Power per transistor stopped falling.

So even though transistors kept shrinking, you could no longer power all of them at once. Enter the dark silicon problem: a chip might have 10 billion transistors, but at any given moment, only a fraction can be active without melting the chip.

What replaced the free lunch

Since 2006, performance gains have come from:

• More cores — parallelism instead of higher clocks
• Specialized accelerators — GPUs, NPUs, TPUs that do specific work more efficiently than a general CPU
• Better microarchitecture — deeper pipelines, wider issue, better branch prediction
• 3D stacking — HBM memory, chiplet designs
• Aggressive power management — turn cores off when idle, run at lower voltage when possible

The thermal ceiling

Every watt a chip consumes must be removed as heat. A modern flagship CPU dissipates 250W in ~600 mm² of silicon. That's a power density higher than a stovetop burner. Cooling is now a first-class chip design constraint, not an afterthought.

This is why phone chips throttle. Why data centers are built next to rivers. Why NVIDIA's H100 costs $30K partly for the packaging and cooling. Why liquid cooling is coming to consumer PCs. It all traces back to Dennard.

What's next

• Backside power delivery — flip power routes to the bottom of the wafer, freeing top layers for signal routing
• 2D materials (graphene, MoS₂) — potential atomic-scale channels
• Photonics — replace some copper interconnects with light
• Analog / neuromorphic compute — trade precision for orders of magnitude less power

None of these are silver bullets. All of them are attempts to work around the same fundamental constraint: you can only move so many electrons through so much silicon before it melts.