Why Are They Spending Trillions on AI?

The numbers and economics behind the AI working-capacity revolution—from electricity and infrastructure to tokens, robots, value, ownership, and agency.

Author: Valto Loikkanen Status: v1.0.1 research package, released 2026-08-14. See CHANGELOG.md for what changed since v1.0.0, and the repository’s Releases page for citable, checksum-verified snapshots of every release. Research initiated: August 12, 2026 Factual source cut-off: 2026-08-13 Licence: CC BY 4.0 for original prose, diagrams, and model structure — see LICENSE for the full scope note (third-party quotes, trademarks, and external data are excluded and remain under their own terms). Suggested attribution: “Research synthesis and modelling by Valto Loikkanen, assisted by AI.” Citation: see CITATION.cff. Checksums: see CHECKSUMS.sha256. Changelog: see CHANGELOG.md.

What this is

A neutral-observer research synthesis reverse-engineering a connected economic model — energy → hardware → compute → models → tokens → AI working capacity → digital work → outcomes → value → agency — from public statements by Jensen Huang (and the Nvidia/Wall Street financing panel), Sam Altman, and Mark Zuckerberg, combined with independently verified hardware specs, pricing, and benchmarks, plus the author’s own prior published work on AI infrastructure ownership and cooperative models.

It is not a vendor report, a political programme, or an investment recommendation. Six reader lenses — Individual, SME/owner-manager, Finance/infrastructure investor, Government/region/community, AI builder/operator, and Educator/researcher/journalist — are given equal weight and no priority ordering, by deliberate editorial decision.

Every substantive claim is tagged with one of five evidence classes (Observed Fact / Attributed Statement / Derived Calculation / Scenario Assumption / Interpretation) so a reader can tell, at a glance, what is independently verified, what is someone’s on-record claim, what is arithmetic, and what is an editable illustration.

Disclosure: the author has commercial and advocacy interests in personal AI infrastructure and cooperative/ownership-based AI models through Prifina, Digiole, ValtoAI, and the PIOS and EIOS reference frameworks. See Method §3.4 of the whitepaper (01-whitepaper.md) for the full disclosure and non-advice boundary.

Reading paths — start here, not with the whole package at once

This package is built so different readers can get a common factual base at different depths, without everyone being forced through the same 89-page document first:

• First-time reader, any lens: start with 00-how-to-use-this-research.md (one page — what this is/is not, what it can/cannot establish), then 12-executive-brief.md.
• The authoritative map — 01-whitepaper.md (+ 01-whitepaper.pdf). The complete model, all evidence-class tagging, all caveats. This is the source of truth every other asset should trace back to. The PDF is self-contained: it embeds the four diagrams most load-bearing for the argument (the end-to-end chain, the ownership stack, the scale spectrum, and the new-value bridge) directly at their point of use, so a reader holding only the downloaded PDF sees them without needing the repository. The remaining diagrams (2–5, 7, 9–10) live only in assets/diagrams/ and the standalone deck/briefs — the PDF is the narrative paper with its most essential visuals embedded, not a complete visual reproduction of the full 11-diagram set.
• The entry point — 12-executive-brief.md. A fast, self-contained summary for readers who want the shape of the argument without the full depth.
• The audit and exploration layer — the companion workbooks (03–08) and 18-companion-data-model.xlsx. Every formula is shown; every assumption is an editable cell. This is where you check the paper’s arithmetic or re-run it with your own numbers.
• The conversation starter — 19-slide-deck.pptx / 13-slide-deck-outline.md and the short-form pieces (14–16). For presenting or sharing the argument, not for auditing it.
• Before relying on any specific figure: check 20-appendix-known-limitations.md — a consolidated list of what this model’s evidence can and cannot currently establish (model performance, utilization, financing, power pricing, adoption timing, and the Professional-tier evidence gap).

Package contents

Rebuilding the Pages site after a content change

Every *.html file at the repo root except index.html and diagrams.html (hand-authored) and perspectives.html (generated by a separate script, below) is generated from its matching .md file — do not hand-edit a generated .html file; edit the markdown source and rebuild:

python3 tools/build-pages.py

This regenerates all pages and rewrites cross-references between them (a link to another tracked .md file becomes a link to that file’s .html page) so reading stays inside the Pages site rather than jumping to GitHub’s markdown viewer. Requires pandoc on PATH.

After adding or editing a perspectives/*.yaml entry, rebuild the Perspectives index:

python3 tools/build-perspectives.py

Requires pyyaml (see tools/requirements.txt).

Running the consistency checker

tools/check-canonical-consistency.py verifies that the whitepaper, the built PDF, the companion Excel workbook, and the slide deck all agree with data/canonical-cost-model.csv — the single source of truth for tier cost figures. It fails closed: a missing dependency, a missing file, or a broken Python environment is reported as a failure, not silently skipped, so a PASS can be trusted to mean every check actually ran.

python3 -m venv .venv
source .venv/bin/activate
pip install -r tools/requirements.txt
python3 tools/check-canonical-consistency.py

Known environment issue on some macOS/Homebrew setups: if the default python3 on your system reports ImportError: ... pyexpat ... or No module named expat, that Python build’s XML support is broken (a Homebrew libexpat version mismatch), and every check that touches the xlsx or pptx will fail — this is a real environment problem the checker is correctly refusing to paper over, not a bug in this script. Use a different Python installation (e.g. a pyenv-managed build, or python3.11 from a fresh Homebrew python@3.11 install) to create the virtualenv above instead.

Verifying checksums

shasum -a 256 -c CHECKSUMS.sha256      # macOS
sha256sum -c CHECKSUMS.sha256          # Linux/GNU coreutils

CHECKSUMS.sha256 contains plain hash/filename lines only — no comments or header text — so both tools above accept it without warnings.

Status and known open items

This is a complete v1.0 draft produced by an AI-assisted research and modelling workflow, with an independent verification pass, per-section fact-consistency audits, and a deterministic (non-AI-regenerated) final assembly step to guarantee no content was dropped or truncated during compilation.

Known items flagged transparently within the package itself:

• Resolved (2026-08-13): the earlier “€42 vs €19.50–25/member/month” cooperative-cost item was not actually a discrepancy — the two figures describe two different hardware tiers (an illustrative €100,000 shared workstation-class machine vs. a pool of $4,699 NVIDIA DGX Spark units), not competing estimates of the same cooperative. See 05-workbook-token-factory-scenarios.md §1a/§1b and 11-appendix-assumption-register.md row 6a for the corrected, fully auditable breakdown.
• Resolved (2026-08-13): the “$0.123/M tokens” GB300-class inference figure is now confirmed directly on NVIDIA’s own site (previously flagged UNVERIFIABLE). It applies specifically to a 72-GPU GB300 NVL72 rack at 116 tokens/sec/user using NVIDIA Dynamo and TensorRT-LLM — it must not be used to price any workstation or desktop-class device. See 03-workbook-global-baseline.md §5.7.
• Several hardware/price inputs (Professional-tier HGX B300 node pricing/power, hyperscale rack capex, the illustrative €100,000 DGX-Station-class and €25,000 humanoid robot prices) have no public source and are flagged ⚠ throughout as placeholders, not verified figures — NVIDIA does not publish retail prices for DGX Station or HGX B300.
• Resolved (2026-08-13): the slide deck, diagrams, and companion workbook are now finished binary files (19-slide-deck.pptx, assets/diagrams/*.png, 18-companion-data-model.xlsx) rather than markdown specs — the workbook’s formulas and the diagrams were independently verified against the source markdown during the build (one hyperscale-utilization formula error and two diagram layout defects were caught and fixed this way).
• Resolved (2026-08-13): a final reconciliation pass found and fixed a cluster of stale Home/Cooperative-tier figures that had not been updated when this paper’s canonical ranges were established — the whitepaper, 05-workbook-token-factory-scenarios.md, 13-slide-deck-outline.md, 19-slide-deck.pptx, and assets/diagrams/diagram-08-scale-spectrum.png all now cite the single canonical Home ($1.37–$11.89/M tokens) and Cooperative ($1.99–$7.62/M tokens) ranges and their correctly-derived $/AI-working-hour figures. The same pass corrected a 1,000x unit error in the Bitcoin-mining energy-monetization comparison (§14 of the whitepaper, §5 of the Token-Factory Scenario Workbook) — the corrected arithmetic inverts that passage’s original illustrative conclusion: valued at the canonical Home-tier production cost, AI tokens’ gross revenue-per-MWh-equivalent exceeds even efficient Bitcoin-mining fleets, not falls below them as an earlier draft claimed. Two prose labels in 18-companion-data-model.xlsx that were stored as broken formulas (would have displayed #NAME?) were also converted to plain text.
• Resolved (2026-08-14): this package is now deployed as a GitHub Pages site at valto.github.io/ai-working-capacity-revolution, tracking main — see the Status line above for how the live site relates to the last tagged release.

Before any public release, a human review pass of the full whitepaper and workbooks is recommended, particularly for the reconciliation and placeholder items above.