Top artificial intelligence labs hide training costs and hardware details for new models

Major artificial intelligence laboratories have abruptly stopped revealing the technical details behind their flagship models, according to an analysis published by Superintel. In previous years, tech companies routinely published paper documentation detailing how many processing chips, training hours, and dataset tokens were required to build a new system. Recent releases from developers such as Google, OpenAI, DeepSeek, Meta, and Moonshot omit hardware budgets, training time, and dataset sizes entirely.

This drop in public transparency comes at a time when physical computing power is surging. Research group Epoch AI estimated that the largest observed artificial intelligence computing cluster grew roughly tenfold between mid-2024 and early 2026. Because developers no longer report their hardware metrics, journalists, academic researchers, and government officials can no longer directly observe the true scale of the newest frontier runs.

It remains uncertain whether firms are staying quiet to shield trade secrets from competitors or to head off government oversight. However, this lack of transparency creates a fundamental policy challenge: when public oversight and regulatory rules rely heavily on hardware usage limits, keeping cluster sizes private makes standard AI safety enforcement virtually impossible.