Why Are They Spending Trillions on AI?
Executive Brief — A Fast Map of the Full Research Paper
Companion to: “Why Are They Spending Trillions on AI?” (v1.0 research package) Author: Valto Loikkanen Factual source cut-off: 2026-08-13 Licence: CC BY 4.0 — Attribution 4.0 International
The ten-second version
Since 2025–2026, a small group of companies, executives, and financial institutions have committed hundreds of billions — and on their own stated ambitions, potentially trillions — of dollars to AI infrastructure. This brief maps the real question: what would need to be true for that spending to make economic sense, and how much of that is independently checkable today versus asserted by people who benefit from it being believed? It does not assume the spending is either obviously justified or obviously a bubble. Working capacity is not the same as value, and nothing in the numbers proves the underlying bet right.
1. What actually happened (the hard facts)
On August 10, 2026, Nvidia signed agreements with six major financial firms — Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs, and KKR — to mobilize more than $500 billion in third-party capital for AI infrastructure. This is a real, signed, dated commitment [OBSERVED FACT — CNBC/Fortune, Aug 10–11, 2026].
On the same broadcast: - BlackRock CEO Larry Fink said the industry will ultimately need to “raise trillions of dollars over the coming years” [ATTRIBUTED STATEMENT — Fink]. - Nvidia CEO Jensen Huang put a price on the buildout: “each gigawatt is something like $50, $60 billion” [ATTRIBUTED STATEMENT — Huang], a figure at the high end of, but broadly consistent with, independent analyst estimates (roughly $35–49B/GW).
These are real statements from named, financially-interested parties — not independently audited costs. That distinction matters throughout everything below.
2. The chain the whole paper follows
Energy → hardware → compute → models → tokens → AI working capacity → digital work → outcomes → value → agency.
Each arrow is a real, costed conversion. The paper’s central discipline: these layers are never collapsed into one number, and the last step — outcome and value — is never treated as automatically produced by the ones before it. A cheap, fast, wrong answer, scaled across a fleet of AI agents, is still cheap and wrong at scale [INTERPRETATION].
A parallel, much less mature chain applies to physical robots: capital + energy + maintenance + utilisation + orchestration → humanoid physical working capacity.
3. Why the cost of a token is not the value of anything
A “token” (a unit of AI output) is a measure of capacity, like a kilowatt-hour measures energy capacity — not a measure of value produced. The paper draws a hard, repeated line between:
- Owned-production cost (electricity + hardware + facility, running open models on your own machines) — typically $0.05–$12 per million tokens, tier-dependent [DERIVED CALCULATION].
- Retail API price (what OpenAI, Anthropic, Google charge for their finished, proprietary products) — typically $2–$50 per million tokens [OBSERVED FACT / ATTRIBUTED STATEMENT, per each vendor’s pricing page].
These two numbers legitimately differ by 1–2 orders of magnitude because retail pricing bundles R&D, safety work, and margin that a bare hardware calculation doesn’t include. Confusing them was a documented error the underlying research made and then corrected — kept in the paper deliberately as a teaching example.
Converted into $/AI-working-hour (how much AI capacity actually costs per hour of use, depending on whether it’s casual chat use or heavy multi-agent orchestration), costs range from under a cent to roughly $150/hour depending entirely on production tier and usage intensity [DERIVED CALCULATION]. None of these figures says whether the hour produced anything worth having.
From capacity to value — a conceptual bridge, not a new set of numbers [INTERPRETATION]: cheap AI working capacity is a supply of digital work, not automatically growth, income, or social benefit. The full paper’s “From AI working capacity to new value” section frames this as two separate questions: (1) can the capacity be created — the cost/utilization arithmetic above — and (2) can that capacity create and capture new value, which depends on invention, user adoption, trust, and a chosen revenue or public-value model, none of which any $/token figure can supply. The same boundary applies to any future Scenario Explorer built on this paper’s workbooks: it can model cost and capacity, not whether a given deployment actually creates value.
4. Three interested voices, shown side by side — none endorsed
| Executive | Frames AI as… | Representative statement | Class |
|---|---|---|---|
| Jensen Huang (Nvidia) | Compute as a financeable infrastructure asset | Nvidia’s chips are an “investable asset” | ATTRIBUTED |
| Sam Altman (OpenAI) | Intelligence as a utility whose falling cost drives demand | Cost of a given AI capability falls “about 10x every 12 months” | ATTRIBUTED |
| Mark Zuckerberg (Meta) | Superintelligence as something that should be personal, not concentrated | Proposes agents that “share a person’s goals and values, not our company’s” | ATTRIBUTED (that the statement exists) |
Each speaks from the layer of the stack their business occupies. None of the three is independently audited; none is treated here as more correct than the others.
5. Bubble or infrastructure transition? Both readings have support
- Supports “real transition”: the $500B financing is a signed instrument, not a rumor; independent grid-demand forecasts (a reported 50–80 GW US shortfall by 2030) are directionally consistent with the executives’ claims.
- Supports “bubble caution”: the per-gigawatt cost, the power-demand figures, and the “trillions” language all come from people who sell the chips or place the capital; independent estimates cluster somewhat lower than Huang’s figure; hardware itself depreciates fast (Nvidia’s own next-generation chip is marketed as up to 10x more efficient per token).
The paper’s position: these are not competing facts, they’re overlapping ones — the same commitment can be real infrastructure and include mispriced risk [INTERPRETATION].
6. Who owns it — one candidate architecture among several
The paper separately asks who controls the stack (energy, hardware, compute, models, information, agents) once it’s built. Six access models exist side by side today — centralized platforms, enterprise dedicated contracts, open models on owned hardware, private ownership, cooperatives, and local/regional capacity — each trading differently across cost, privacy, resilience, and control. No model is asserted here as universally better.
The author’s own cooperative/ownership proposal (through Prifina, Digiole, ValtoAI, and the PIOS and EIOS reference frameworks) is one candidate among these six, argued on its own numbers — including two cooperative-cost scenarios for two different hardware tiers, shown side by side rather than treated as competing: ~€42/member/month (50 members) for an illustrative shared workstation-class machine, and ~€20–23/member/month (50 members) for a pool of smaller NVIDIA DGX Spark devices. This is disclosed as the author’s own commercial and advocacy interest, not presented as a neutral finding.
7. Six reader lenses — equal weight, no priority order
The full paper is written for six audiences, each pointed to different sections, none ranked above another:
- Individual — what owning vs. renting personal AI actually costs, and what data control means for you.
- SME / owner-manager — whether AI changes how you deliver an existing service, not just what you sell.
- Finance / infrastructure investor — the gap between executive statements and audited figures.
- Government / region / community — grid-demand implications and cooperative vs. centralized infrastructure choices.
- AI builder / operator — real $/token production costs vs. retail pricing traps.
- Educator / researcher / journalist — how to tell an attributed claim from a proven fact.
8. The bottom line
Trillions are being committed because a small number of well-positioned people believe cheap, abundant AI working capacity is coming and will be worth more than it costs to build. Abundant working capacity, if it arrives, does not make the human question easier — it raises the importance of judgment about what to do with it, not less [INTERPRETATION].
Evidence-class legend
OBSERVED FACT — independently checked against a primary source. ATTRIBUTED STATEMENT — what a named person/org said on record; not automatically proven true. DERIVED CALCULATION — arithmetic shown from cited inputs. SCENARIO ASSUMPTION — an editable illustrative parameter, not a market figure. INTERPRETATION — a labelled possible reading, never presented as fact.
Non-advice statement: This brief and the full paper it summarizes are educational research and scenario analysis — not investment, legal, tax, procurement, or policy advice. All cost and scale figures are editable illustrations built on stated assumptions, not forecasts or recommendations.
Licensed under CC BY 4.0. Suggested attribution: Research synthesis and modelling by Valto Loikkanen, assisted by AI.
This executive brief is a standalone deliverable summarizing the full whitepaper “Why Are They Spending Trillions on AI?”