Slide Deck Outline — “Why Are They Spending Trillions on AI?”

Companion deliverable (Release Asset #5) — 26 slides, following the whitepaper’s own Front Matter → Closing structure. Evidence-class tags used throughout: [FACT] = Observed Fact · [ATTR] = Attributed Statement · [CALC] = Derived Calculation · [ASSUMPTION] = Scenario Assumption · [INTERP] = Interpretation.

Slide 1 — Title - “Why Are They Spending Trillions on AI?” — The numbers and economics behind the AI working-capacity revolution - Author: Valto Loikkanen · v1.0 research package · Factual source cut-off 2026-08-13 - Licence: CC BY 4.0 — Attribution 4.0 International - Speaker note: This is a neutral research synthesis, not a vendor report or investment recommendation — state that framing before anything else.

Slide 2 — Method: Five Evidence Classes - Table: OBSERVED FACT (primary source, checked on/before 2026-08-12) · ATTRIBUTED STATEMENT (on-record claim, not auto-proven) · DERIVED CALCULATION (formula always shown) · SCENARIO ASSUMPTION (editable, not a market figure) · INTERPRETATION (labelled reading, never fact) - Every claim in this deck carries one of these five tags, exactly as in the source paper. - No fabricated quotes, figures, or URLs; unverifiable items are labelled UNVERIFIABLE rather than smoothed over. - Speaker note: This slide is the contract with the audience — every number that follows is traceable to one of these five buckets.

Slide 3 — How to Use This Research: Six Equal Lenses - Individual · SME/owner-manager · Finance/infrastructure investor · Government/region/community · AI builder/operator · Educator/researcher/journalist - No lens is prioritized over any other — a deliberate editorial stance held throughout, not an oversight [INTERP, stated editorial policy]. - Each lens has a different entry point into the same chain: energy → hardware → compute → models → tokens → AI working capacity → work → outcomes → value → agency. - Speaker note: Tell the room which lens they’re likely reading from, but stress that the deck deliberately refuses to rank the six.

Slide 4 — Executive Summary: The Chain - Chain: energy → hardware → compute → models → tokens → AI working capacity → digital work → outcomes → value → agency (Front Matter §1). - Parallel physical chain: capital + energy + maintenance + utilisation + orchestration → humanoid working capacity (Part VI). - Headline: Aug 10, 2026 — Nvidia + 6 firms (Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs, KKR) MOUs to mobilize >$500B [FACT]; Fink: “trillions of dollars over the coming years” [ATTR, Fink]. - Ten-second takeaway: trillions are committed on a belief that cheap AI working capacity is coming — working capacity is not value, and nothing here proves that belief right. - Speaker note: This slide is the whole paper in miniature; everything after it is evidence for or against pieces of this one paragraph.

Slide 5 — Part I: Why Are They Spending Trillions? - Aug 10, 2026: Nvidia signs MOUs with six asset managers/banks to mobilize >$500B in third-party AI-infrastructure capital [FACT, CNBC/Fortune]. - Huang: “each gigawatt is something like $50, $60 billion” [ATTR, Huang]. - Three structurally different, differently-incentivized answers on offer: compute as infrastructure (Huang), intelligence as utility (Altman), superintelligence as personal (Zuckerberg) — shown side by side, none endorsed. - Speaker note: Frame this as the reader’s actual question, not a rhetorical one — three interested parties, three different stakes.

Slide 6 — Three Voices, Side by Side - Table:

• None of the three claims is independently audited; that they were said is confirmed, whether they’re true is not.
• Speaker note: Resist any temptation to declare a winner here — the paper’s whole editorial stance depends on not doing so.

Slide 7 — What Would Need to Be True: Bubble, Infrastructure Transition, or Both? - Supports “transition”: signed $500B MOUs are real [FACT]; independent power-demand forecasts (BCG 50–80GW, S&P Global 64.4GW in 2025) are directionally consistent [ATTR, low-confidence sourcing]. - Supports “bubble caution”: Huang/Fink’s own per-GW and demand figures are self-interested ATTR statements; independent analyst estimates ($35–49B/GW) sit below Huang’s $50–60B [ATTR]. - Disciplined position: both readings draw on overlapping, not competing, facts [INTERP] — the same MOU can be real infrastructure and include mispriced capital. - Speaker note: The point is not to pick a side — it’s to show the same facts support two different narratives simultaneously.

Slide 8: Part II — Electricity, Power, and Tokens per Watt - GB300 NVL72 rack: 135kW TDP / 155kW peak per OEM spec [FACT, Lenovo]; NVIDIA itself publishes no power figure. - MLPerf v6.0: “2.5M tokens/sec” is a 4-rack/288-GPU aggregate, not a single-rack figure — a common misattribution corrected here [FACT]. - Tokens/watt is not one number per chip — it ranges roughly 1.7M–5.2M tokens/sec/MW depending on latency tolerance [CALC/INTERP]. - Speaker note: Use this slide to show the paper’s discipline in action — catching and correcting a benchmark misattribution rather than repeating it.

Slide 9 — Hardware as Capital Equipment, and Its Financing - Huang frames Nvidia’s platform as an “investable asset” [ATTR, Huang, CNBC Aug 10 2026]. - Capital stack observed vs. assumed: DGX Spark $4,699 MSRP [FACT]; HGX B300 has no public price [ASSUMPTION]; GB300 rack ~$4M is an illustrative analyst estimate [ASSUMPTION]. - Depreciation risk: Vera Rubin claimed “up to 10x more tokens/MW” vs GB200 [ATTR, vendor]; SemiAnalysis’s independent figure is 2–5.4x [FACT-adjacent]. - Speaker note: Emphasize that today’s capital is a bet against a hardware curve both the vendor and independent analysts agree is moving fast.

Slide 10 — The Eight Cost Layers (Never Collapsed Into One Number) - Layers: raw energy cost → hardware-amortized cost → financed asset cost → full operating infrastructure cost → capacity/utilisation cost → token production cost → workload/AI-working-capacity cost → outcome and value (never mechanically derived). - Worked EUR example: financing term alone moves cost from €0.688/M tokens to €1.435/M tokens at 50% utilisation [CALC]. - Utilisation is usually the single largest lever at every scale [INTERP, cross-workbook pattern]. - Speaker note: This ladder is the spine of the entire cost analysis — every dollar figure later in the deck sits on one specific rung, never blended with another.

Slide 11 — Owned Production Cost vs. Retail API Price - Self-correction preserved deliberately: retail lab prices are a finished, marked-up product price, not a cost basis for owned hardware [method note]. - Retail benchmark: Claude Sonnet 5 $2/$10 per M tokens [FACT]; GPT-5.6 Terra ~$2/$12 [ATTR, via intermediary fetch]; Gemini 3.1 Pro Preview $2/$12 [FACT]. - Owned production: Home $1.37–$11.89/M tokens (canonical); Hyperscale ~$0.091–0.312/M tokens, canonical full-layer, mid $0.133/M [CALC] — Hyperscale sits 1–2 orders of magnitude below retail; Home sits closer to, and at low utilization inside, the cheapest retail floor. The gap is real everywhere but its size is tier- and utilization-dependent, by design, not error. - Speaker note: This gap is the point, not a mistake to reconcile — “generate your own power” vs. “buy from the grid.”

Slide 12 — Part III: Usage-Intensity Bands - Four bands (tokens/AI-working-hour, all SCENARIO ASSUMPTION): Chat/advisor (10K–30K) · Active copilot (60K–120K) · Delegated agent (200K–600K) · Heavy multi-agent orchestration (1M–12M+). - OpenAI’s own unaudited telemetry: Codex = 99.8% of weekly internal output tokens; 99th-percentile users ran 60+ agent-hours/day [ATTR, OpenAI, self-reported]. - These describe OpenAI staff using OpenAI’s own product — not a general-population statistic [caveat, stated explicitly]. - Orchestration does not create a different kind of hour — it creates multiple working hours in parallel: 1 hour = 1 agent/person × 1h; 10-team hour = 10 agents/people × 1h = 10 agent/human-hours; 1 day = 8 agent/human-hours; 10-team day = 80 agent/human-hours. Team capacity is workers × hours, on both sides, with no exception. - Speaker note: Band 4’s open-ended ceiling is informed by, not proven by, this one company’s internal data. The 1-vs-10 table is the guardrail against reading a band-4 figure as “one very fast hour” instead of “many parallel ordinary hours.”

Slide 13 — From Tokens to $/AI-Working-Hour - Formula: /AI − working − hour = (tokens/hour ÷ 1, 000, 000) × (/M tokens at that tier) [CALC]. - At delegated-agent band, central case: Hyperscale $0.0532/hr vs. Home $1.268/hr vs. Retail-API $3.60/hr [CALC] — a ~68x spread across tiers at the same usage band. - Self-hosting is not automatically the cheapest path — its case rests on control and independence, not undercutting retail [INTERP]. - Speaker note: The spread is arithmetic, not opinion — but which column matters depends entirely on which lens is asking.

Slide 14 — Quality Multipliers and the Limits of Human/AI Comparison - Eight factors shape effective working capacity: capability, reliability, initiative, judgment, creativity, context, tools, information [INTERP, author’s own framework]. - Reliability and information act as multipliers, not add-ons — unreliable capacity caps how much can safely be delegated. - What IS comparable: throughput, marginal cost, capacity elasticity. What is NOT: workload complexity, quality/correctness, supervision burden, context completeness — kept visibly separate, not folded into one multiplier. - The employer/billable split: owned AI pairs with human employer cost (fully loaded — wage+benefits+overhead); retail AI pairs with human billable/externally-purchased rate (agency/contractor/consultancy, which embeds its own margin and overhead on top). Comparing owned AI against a billable rate — or retail AI against employer cost — mixes the internal/external layers exactly as Part III warns against mixing owned-production cost with retail API price [INTERP]. - Speaker note: This is the guardrail slide against naive “$/hour AI vs. $/hour human” headlines — and against the narrower mistake of picking whichever AI number and whichever human number make the best-sounding comparison.

Slide 15 — Part IV: Information as AI’s Operating System - Author’s own framing: information is “the operating system for AI” [FACT that phrase appears; INTERP as underlying claim] — disclosed author interest in this framing (Prifina/Digiole/ValtoAI/PIOS/EIOS). - Same model + different information = very different effective working capacity. - Zuckerberg: “fully private mode…even Meta cannot see” [ATTR]; Altman’s usage anecdote implies ~1,000,000x token growth per heaviest user, 2019→2026 [CALC on ATTR inputs]. - Speaker note: Flag the disclosure here explicitly — this section sits closest to the author’s own commercial stake.

Slide 16 — AI Maturity and the Shifting Human Role - Two axes: how work gets done (Advise → Cowork → Delegate → Lead) and scale of orchestration (individual → team → workforce) [INTERP, author’s framework, partially unverified public form]. - Three framings of the same shift: Altman on freedom/agency [ATTR]; Zuckerberg on balance of power [FACT/ATTR]; Huang on capital allocation [ATTR] — none more correct than the others. - As execution gets cheap, “what should we do” replaces “how do we get it done” as the load-bearing human question [INTERP]. - Speaker note: Each of the six lenses experiences this shift differently — no priority ordering intended.

Slide 17 — Value Can Be Positive, Zero, or Negative - Working capacity ≠ value. Orchestration multiplies capacity, not judgment [INTERP]: “a cheap, fast, wrong answer scaled across a multi-agent team is still cheap and wrong at scale.” - Three human analogies: brilliant professor vs. street-smart entrepreneur; new CEO vs. veteran employee; world-class investor vs. empathetic salesperson — each shows capability ≠ fit ≠ value. - No production tier or ownership architecture is shown by this paper’s evidence to be inherently more likely to produce positive value than another. - Speaker note: This is the hinge slide — everything before it is cost/capacity; everything after assumes this boundary holds.

Slide 18 — From AI Working Capacity to New Value - Conceptual bridge, not a numerical finding [INTERP throughout]: capacity created (Parts I–IV) is not automatically growth, income, or social benefit. - Two separate questions: (1) can the capacity be created — this paper’s cost/utilization model; (2) can that capacity create and capture new value — invention, adoption, trust, revenue model, none of which any $/token figure supplies. - Visual: “Creating Growth Based on New Value” (Diagram 11) — two columns converging on a shared “New Value” band: Idea/Invention → Innovation → Verifying New Value descending on the left; Productivity → Revenue Model → Value Proposition & Communication ascending on the right. - Note: the same boundary applies to any future Scenario Explorer built on this paper’s workbooks — it models cost and capacity, not whether a deployment creates value. - Speaker note: This slide is the second hinge — it stops the room from reading “cheaper tokens” as “more growth” without the missing middle step.

Slide 19 — Part V: The Ownership Stack - Layers, each separately ownable: energy, hardware, compute (access), models, information, agents, identity, governance. - Concentration varies sharply by layer: hardware manufacturing is extremely concentrated [FACT/INTERP]; information is the least standardized/portable layer [INTERP]. - No single actor owns “AI” end to end — ownership questions must be asked layer by layer. - Speaker note: Use this to reframe “who owns AI” as a bad question — it’s really six or seven different questions.

Slide 20 — Six Coexisting Access Models - Table: Centralized platform · Enterprise dedicated (e.g., IBM Cloud/Together AI $240M deal [FACT]) · Open models on owned hardware · Private/individual ownership · Cooperative (€19.50–23/member/month for a DGX Spark pool, or €42/member/month for a shared workstation-class machine — two hardware tiers, not one contested figure [CALC/FACT]) · Local/regional capacity. - None asserted as universally superior — trade-offs differ across scale, cost, privacy, resilience, convenience, control [INTERP]. - Ownership models only make arithmetic sense once usage sits consistently at copilot-band intensity or above [CALC-supported break-even]. - Speaker note: The two cooperative figures describe two different hardware tiers (a DGX Spark pool vs. a shared workstation-class machine), not a discrepancy in one estimate — shown explicitly as a worked example of assumption transparency.

Slide 21 — Electricity-Grid Analogy and Digital Sovereignty Beyond Geography - Analogy: home solar → cooperative → commercial producer → utility-scale plant → grid exchange, mapped onto AI compute tiers [INTERP, author’s own commentary, not independently sourced verbatim]. - Weakest link: no frictionless spot market for AI compute yet exists, unlike electricity or Bitcoin hashrate [FACT/INTERP]. - Sovereignty reframed via five questions: control, portability, exit, governance, continuity — not just server geography [INTERP, framework proposed in this paper]. - Speaker note: Note explicitly that this analogy is the author’s own structural comparison, flagged as such rather than presented as settled.

Slide 22 — Part VI: Scale Scenarios, Tier by Tier - Table: Home ($4,699, $1.37–11.89/M tokens) · Cooperative (~$47K, $1.99–7.62/M) · Professional (unofficial/ASSUMPTION, $0.044–0.146/M, no opex — least evidenced) · Hyperscale (~$16M/4-rack, canonical full-layer $0.091–0.312/M, mid $0.133/M) [CALC throughout]. - Governing levers: utilisation (3.5–4x swing), financing term, interactivity setting — all editable, none a forecast. - Cost-per-hour of capacity is not value — a 500–800x swing between production cost and retail pricing shows how meaningless a bare figure is without its pricing assumption. - Speaker note: State the non-advice boundary explicitly here — every number is a scenario, not a recommendation.

Slide 23 — Humanoid Robots as Embodied Working Capacity - Real market range: Unitree G1 from $13.5K [FACT] to Agility Digit ~$250K [ATTR, secondary]; illustrative €25K base case sits within this spread [INTERP]. - Same utilisation pattern as digital tiers: 2,000→8,000 hrs/yr cuts cost/hour ~3.5–4x [CALC]; electricity is not the binding cost here (unlike token production). - Scope explicitly narrow: excludes supervision, insurance, downtime — every figure is a cost floor, not a full commercial cost. - Speaker note: Flag this as a “later, illustrative extension,” not a general robotics-economics claim.

Slide 24 — Part VII: Six Lenses, Equal Weight, No Ranking - Individual: rent vs. own break-even sits near the copilot band, not chat band. - SME: delivery innovation, not product innovation, is the accessible opportunity [ATTR, author’s prior work]. - Investor/Government/Builder/Educator: same MOU and cost figures, read through different risk lenses — interested-party statements (Huang, Fink) flagged consistently as ATTR, not audited fact. - Speaker note: Reiterate — these are grouped for slide economy only, not because any lens outranks another.

Slide 25 — Closing: Opportunity, Not Inevitability - Capital and infrastructure figures (Parts II–IV) support reading this as opportunity, not destiny [INTERP, editorial choice]. - The unknown variable: adoption, organizational change, and time — not model capability — appear to be the binding constraint [INTERP]. - Closing synthesis: “how do we get it done” is getting cheaper faster than most say plainly; “what should we do” has not gotten any easier at all. - Speaker note: End on the open question, not a resolved answer — the paper deliberately does not decide this for the reader.

Slide 26 — Licence and Attribution - Licensed under CC BY 4.0. Suggested attribution: Research synthesis and modelling by Valto Loikkanen, assisted by AI. - This material is educational research and scenario analysis — not investment, legal, tax, procurement, or policy advice. - Companion workbooks (Global Baseline, AI Working-Capacity Conversion, Token-Factory Scenarios, Investment-Thesis Notes, Humanoid Working-Capacity, Localized/EUR-Finland template) available for readers to rebuild any model with their own inputs. - Speaker note: Close by pointing to the workbooks as the place where the audience can go build their own numbers.

Licensed under CC BY 4.0. Suggested attribution: Research synthesis and modelling by Valto Loikkanen, assisted by AI.