Frontier models are now the least transparent — labs stopped disclosing training compute, data, and parameters (Stanford AI Index 2026)
Stanford HAI's AI Index 2026 finds industry produced over 90% of notable AI models in 2025, yet 'the most capable models are now the least transparent': training code, parameter counts, dataset sizes, and training duration are no longer disclosed for the most resource-intensive systems, including those from OpenAI, Anthropic, and Google. Public knowledge of how frontier models are built is shrinking as the models get more capable — and AI data-center power capacity reached ~29.6 GW by Q4 2025.
Why it matters (spine). This is the counterpart to the inference-cost card (inference-cost-collapse) and the data-exhaustion card (data-exhaustion-2026-2032). As capability rises, the public record of how these systems are built is going dark — the AI Index reports that training code, parameter counts, dataset sizes, and training duration "are no longer disclosed for several of the most resource-intensive systems, including those from OpenAI, Anthropic, and Google," and that parameter counts have "stayed near 1 trillion for three years, though reporting from frontier labs has stopped." Cheap inference plus opaque training is precisely the world RDKG bet on: when you can no longer see (or trust) what went into a model, the scarce, defensible thing becomes verified, independently-measured ground truth with a receipt. Money + metacrisis lever: an auditable, provenance-stamped knowledge layer is worth more, not less, as the frontier closes.
What would prove it wrong (falsifier). Frontier labs resume disclosing training compute, dataset sizes, and parameter counts on a capability-equivalent basis (a transparency reversal), or a credible independent-audit regime makes model internals verifiable without first-party disclosure — either of which would weaken the case that external verified ground truth is the scarce input.
Live-status check + triangulation. Quoted from the AI Index 2026 Report, Chapter 1 (Research and Development), Stanford HAI (report published 13 April 2026; data window through Q4 2025): "Industry produced over 90% of notable AI models in 2025, but the most capable models are now the least transparent." Data-center figure: "Total AI data center power capacity reached approximately 29.6 GW by Q4 2025, enough to power all of New York state at peak demand," of which "roughly 11.8 GW" is chip thermal design power. Fetched from the primary chapter PDF in-session.
Correction carried forward (discipline). A figure circulating in secondary summaries — "DeepSeek-V3 trained for ~$5.6 million" — was not used here: the AI Index's "5.6 million" refers to "5.6 million projects on GitHub" (open-source growth), a different sentence entirely. Only the primary-quoted transparency and power-capacity figures are carded as verified.
Lean the moat harder into ground truth we measure ourselves. As public disclosure of how frontier models are trained disappears, the defensible input is not model access but independently-verified, freshly-dated, falsifiable data — the thing a closed lab cannot retroactively publish. Make 'measured by us, sourced, dated' the RDKG differentiator.
domains: machine-layer · transparency · training-disclosure · ai-index · moat · compute