Is AI a Bubble? I Didn't Think So Until I Heard of SDD (hyperdev.matsuoka.com)

Specification-Driven Development (SDD) has suddenly crystallized as both a practical response to AI “vibe coding” and a convenient narrative propping up runaway valuations. Startups and incumbents launched spec-first tools (GitHub’s Spec Kit, Amazon’s Kiro, Tessl’s Spec Registry) while Cursor rocketed to a $9.9B valuation and Tessl raised $125M but only shipped a beta registry months late. The rise of SDD reframes the hype: it institutionalizes writing detailed, machine-executable specs before generation to reduce hallucinations (Tessl’s registry lists 10,000+ library specs) and to provide audit trails that satisfy enterprise procurement and compliance needs (SOC 2, ISO standards). Yet the market shows bubble dynamics—AI coding firms trading at 25–70x ARR (Codeium ~70x) versus the dot‑com peak of 18x—despite real adoption (84% of developers use AI tools; Copilot ~$400M ARR). Technically, SDD flips the workflow: spec → implementation plan → generated code, which mitigates issues Karpathy flagged as “vibe coding” (security audits found 170/1,645 apps with vulnerabilities). Empirical evidence tempers sweeping automation claims: the oft-cited “60–70%” comes from McKinsey’s cross-occupation potential, while METR found experienced devs ran 19% slower with AI despite perceived speedups. AI shines at boilerplate generation (50–80% savings), docs (60–70%), code translation (40–60%) and simple fixes (30–50%), but struggles with architecture, complex logic, legacy integration and team coordination—greenfield projects can approach near‑complete automation, team/legacy contexts often drop below 30%. Expect a market correction in 18–24 months, with consolidation around disciplined vendors that combine product-market fit, governance, and operational rigor.