Jensen Huang reveals that Anthropic drove 100 percent of TPU growth, underscoring the AI chip market's extreme concentration. The Nvidia CEO frames this as validation of the company's strategic focus on large language model providers rather than broad-based demand. Huang also highlights Nvidia's supply-chain advantages, arguing they prove harder to replicate than raw benchmark performance.

Meanwhile, Simon Willison identifies November 2025 as the inflection point where AI agents transitioned from prototype stage to operational deployment. He highlights "dark factories" - autonomous systems operating with minimal human oversight - and describes a shift toward agentic engineering as the dominant development paradigm.

The headline hints at a deeper issue in AI development: when models train other models without human inspection, the process becomes opaque. This matters because model-to-model knowledge transfer happens at scale, often without interpretable outputs. Engineers can't easily verify what gets learned or why. This creates a transparency problem within the AI research pipeline itself.

The concentration of TPU demand around Anthropic signals that compute capacity remains the limiting factor for frontier model development. Nvidia controls the supply. Anthropic controls the primary demand. This two-player dynamic shapes the entire AI infrastructure landscape.

Willison's timing observation carries weight. November 2025 marked when agentic systems moved beyond academic interest into production workloads. That shift requires different engineering practices, safety considerations, and operational frameworks than earlier models.

The dark factories concept compounds the transparency issue. When autonomous systems operate with minimal human oversight, and when training happens model-to-model in opaque ways, the industry faces a growing knowledge gap about how these systems actually behave at scale. No one fully sees inside the training process or the operational decisions these agents make.

This convergence - concentrated compute supply, black-box training methods, dark operational factories - defines the current AI moment. The