AI generates insight in seconds. The question is which insight to believe.

NASA’s Commercial Lunar Payload Services task-order awards alone sum to approximately $1.04 billion, directed at lunar infrastructure with identified mission requirements and contracted deliverables.^[1] Against this, the largest publicly known raise for a water-first asteroid-mining venture is Karman+’s $20 million seed round in February 2025.^[2] The asymmetry is not accidental. Capital follows demand signals, and lunar infrastructure has them. Asteroid mining has spectral surveys and white papers.

The mechanism is procurement-driven capital allocation. Institutional investors require demand proof – binding offtake, anchor customers, contracted throughput – before committing growth-stage capital. Without demand proof, ventures remain at seed stage regardless of technical progress. The $20 million seed is not a stepping stone; it is a ceiling imposed by the absence of anyone willing to commit to buying what these ventures propose to sell.^[3]

Per-kilogram launch costs to LEO declined at roughly 10.4% annually from 2000 to 2020.^[5] Reusable super-heavy lift will accelerate that curve. Even if the actual price lands at five times SpaceX’s target, Earth-launched water to LEO undercuts any plausible asteroid-water cost through at least 2035.^[6] You do not need Starship at $200 per kilogram to undercut asteroid water. You need it at $2,000 per kilogram, which falls within the range of near-term improvements to existing heavy-lift vehicles.

Governance verification flagged a historical pattern match: in 1980 Exxon committed approximately $5 billion (inflation-adjusted) to the Colony Oil Shale Project at Parachute Creek, Colorado.^[7] The resource was enormous and verified. On May 2, 1982 – “Black Sunday” – Exxon shut the project down and laid off 2,200 workers. The kerogen did not disappear. What disappeared was the demand environment.^[3]

The COTS program awarded approximately $800 million in funded Space Act Agreements starting in 2006. CRS contracts followed, exceeding $3 billion in initial phases.^[9] Six years after the first COTS award, SpaceX delivered cargo to the ISS. NASA did not invent commercial cargo; it signed purchase orders, and the purchase orders built the industry.

A binding offtake contract changes the incentive structure through a specific chain: contracted revenue creates a bankable projection, which funds hardware development, which enables depot infrastructure, which establishes a price signal that secondary buyers can reference. The price signal converts a one-customer market into a multi-customer market. Without it, the asteroid-water sector has the policy equivalent of an LOI – legal authorization is not financial commitment.^[10]

Google AI Overviews reaches more than 1.5 billion monthly users. ChatGPT has 400 million weekly actives. Meta AI is approaching one billion monthly.^[1] The numbers suggest a transition that has already happened. They haven’t.

AI as a primary news source sits at low single digits to roughly 7%.^[2] Search’s zero-click rate was already approximately 58.5% before AI answers existed. What the data describes is a two-layer market: mass reach without behavioural substitution. People have access to AI-synthesised answers. They have not yet stopped consulting shared sources. The intervention window is structurally open, but the adoption curve is steepening.

Users report feeling more informed after AI explanations. But peer-reviewed evidence for actual comprehension gains is sparse.^[3] VALIS cross-validated the claim against the medical and policy literature and found the confidence-competence gap is documented, but the functional-literacy claims it rests on remain unverified at scale.

The measurable divergence is specific: 0–8 percentage points on hard factual items, 10–30 percentage points on value-laden framing, and 5–15 percentage points on hedging and confidence.^[4] Moral and value framing shows the strongest, most robust effects. This is not an inevitability. It is a design problem, and design problems have design solutions.

OS assistants, browser defaults, search defaults, messaging hubs, enterprise suites: these are the real chokepoints. The most consequential path dependence in AI is not in the models. It is in the distribution layer that decides which model’s answer a person sees first.^[5]

The EU and China already have binding countermeasures in place: choice screens, interoperability requirements, transparency obligations, provenance labelling. The US, UK, and OECD still rely on voluntary standards and competition tools. Open-weight models reduced API-only moderation effectiveness but did not make centralised governance obsolete; distribution remains hub-centric.^[6] The governance room is shrinking fastest at this layer, and it is the layer where intervention still works.