The bottleneck has moved up the stack
AI code generation tools excel at producing working code. What they don't do well: design systems that perform under load, scale sustainably, or meet non-functional requirements like security and maintainability.
The pattern is consistent across enterprise deployments. AI-generated code introduces performance-critical mistakes—over-allocation, N+1 database queries, inefficient data structures. These aren't syntax errors. They're design failures that surface in production.
Anthropic demonstrated this last week with AI-resistant technical evaluations. They iterated on take-home tests specifically to identify skills Claude struggles with—memory optimization tricks, system-level reasoning, architectural trade-offs. The results matter for hiring: engineers who rely on AI for design decisions can't pass tests that require first-principles thinking.
The infrastructure isn't the problem
Much of the AI bottleneck discussion focuses on compute, memory, and energy constraints. Data centers need 40-50GW over three years. HBM memory limits on NVIDIA H100/H200 chips delay builds. Multi-year compute contracts lock in capacity.
These are real constraints—but they're infrastructure problems. The harder problem is architectural. Specialized models and quantization techniques already mitigate memory bottlenecks for short-context workloads. Power and permitting delays can be solved with capital.
What can't be solved with capital: teaching AI to make architectural trade-offs, anticipate edge cases, or design for operational sustainability.
What CTOs should watch
The "rising water effect" is widening the gap between engineers who use AI as a code assistant and those who design systems. Treat AI output as unreviewed junior code. The engineers who add value are the ones who can review it, spot the performance traps, and redesign the approach.
Enterprise AI is shifting toward domain-specific "Superagents" as general-purpose model costs rise. This reinforces the pattern: AI handles narrow, well-defined tasks. Humans handle system design, integration, and the messy reality of production environments.
Stanford's recent research predicts limited productivity gains outside programming. We're seeing why. Code generation is easier to automate than program design. The bottleneck moved—and it's not going away with better GPUs.