Myth ledger: 'Inference is two-thirds of all AI compute in 2026'

The circulating claim. The two-thirds/half/third progression appears near-verbatim across at least four venues, anchored by Deloitte's 2026 TMT Predictions.

What the citation chain shows (traced link-by-link, capture 2026-07-06): Deloitte's own endnote for the statistic cites a World Economic Forum contributor op-ed by Rodrigo Liang — CEO of SambaNova Systems, an AI inference-chip vendor — a source with a direct commercial stake in "inference dominates" being believed, and no compute-measurement methodology anywhere in the chain. The WEF piece itself returned 403; even Deloitte's attribution of the figure to it is unverified word-for-word.

What runs against it. The one real spending datapoint in the vault (Epoch AI's curation of OpenAI 2024 figures, Tier 2): training $3B, inference $1.8B — training ABOVE inference at a frontier lab, with research compute the largest single category. Epoch's theoretical allocation work predicts comparable training/inference spending, not inference dominance.

Status history.

• 2026-07-07 (later, cycle 5) — contested REINFORCED: the ruling-7 re-source run checked the two venues most likely to hold a primary measurement (Stanford HAI AI Index, Epoch AI) — neither publishes the two-thirds figure or any industry-wide inference-share measurement; the primary-adjacent estimates that do exist disagree with each other and with the dominance narrative. The gap in the public record is now double-confirmed.
• 2026-07-07 — opened as contested (queen cycle 4): not merely unverified — a checkable counter-datapoint exists and the chain's root is incentive-disqualified for a quantitative claim under sources.md's own floor logic.

Receipts. Capture 2026-07-06-is-training-vs-inference (all URLs, fetch statuses, and the 403 disclosed); claim-inference-dominant-ai-compute-2026 (carries the [unverified-quant] flag + audit V-002/V-011); claim-training-inference-compute-asymmetry-mechanism (the solid half of the seam).