The Annual AI Slowdown Panic is Here

What It Covers

The AI Breakdown examines the recurring pattern of summer AI slowdown panic arriving early in 2025, driven by token shortages, Uber's ROI concerns, and a VS Code install plateau, while contrasting these narratives against surging GPU rental prices, 10x annual token demand growth, and record revenues at OpenAI and Anthropic.

Key Questions Answered

• •Benchmark validity: DeepSWE, built by DataCurve, addresses benchmark gaming by creating tasks from scratch rather than scraping GitHub issues. GPT-5.5 scored 70% versus DeepSeek V4's 8%, revealing a 30+ percentage point gap between frontier and Chinese models that existing benchmarks like SWE-Bench completely obscured. Self-verification behavior — models writing their own tests — was the clearest differentiator between top and weaker performers.
• •Token supply vs. demand math: Global inference capacity is expanding roughly 3x annually, while token demand is growing approximately 10x per year according to EpicAI research. GPU rental prices have doubled in four months. This supply-demand imbalance means OpenAI and Anthropic face no near-term revenue pressure, making the bubble narrative structurally inconsistent with basic commodity pricing signals.
• •IDE market share shift: A plateau in VS Code AI extension installs reflects platform migration, not declining adoption. OpenAI Codex CLI installs grew from 100,000 per day in January to over 1.5 million per day recently, as developers move to terminal interfaces and desktop apps. Tracking only IDE metrics produces a systematically misleading picture of actual coding agent adoption.
• •Agent debt management: Rapidly assembled agent workflows accumulate "agent debt" — conflicting system prompts, polluted memory, and overlapping tools that produce unpredictable behavior months later. Treating agent infrastructure with the same discipline applied to technical debt — regular cleanup, clear tool boundaries, and documented system prompts — becomes a necessary operational practice as agentic deployments scale inside organizations.
• •AI job displacement recalibration: Sam Altman acknowledged miscalculating how quickly AI would eliminate entry-level white-collar roles. Goldman Sachs CEO David Solomon separately estimated AI has displaced 16% of entry-level tasks internally while arguing productivity gains historically expand total employment. The practical friction of organizational AI deployment creates a natural speed limit that theoretical task-automation models consistently underestimate.