Mapping the full AI stack to money
agent37.club
Every session we’ve done — agents, models, hyperscalers, chips, energy — maps directly to companies, valuations, and risks. The question isn’t whether AI is big. It’s which layer captures the value, and for how long.
This session synthesizes everything into investment frameworks: S-curves to find positioning, Wright’s Law to model costs, bull/bear cases to stress-test conviction, and historical parallels to calibrate expectations.
Value concentrates at chokepoints — layers where one company controls the bottleneck. The biggest investment returns come from identifying chokepoints before the market prices them in.
Where each layer sits on the adoption curve
Where on the S-curve is each layer? Growth phase = buy the picks and shovels. Maturity = the value migrates to whoever builds on top.
Most investors buy at the top of the growth phase, mistaking maturity signals for continued acceleration. Understanding the stack tells you when to shift.
| Layer | Phase | Signal | Investment Implication |
|---|---|---|---|
| Applications | Early Growth | Cursor: $0 → $500M ARR in 18 months | Highest upside, highest risk. Most startups will fail. |
| Models | Growth | Training costs falling 50× faster than Moore’s Law | Commoditizing fast. Value moving to inference & deployment. |
| Hyperscalers | Peak Growth | $650B capex in 2026, 36% YoY increase | Revenue growth strong, but capex intensity is a drag on FCF. |
| Chips | Growth | NVIDIA 62% YoY revenue growth, TSMC 30% | Picks and shovels. Best risk-adjusted returns in the stack. |
| Energy | Early Growth | Power transformer lead time: 128 weeks | Physical bottleneck. Long-duration thesis with real scarcity. |
The most powerful force in technology economics
For every cumulative doubling of units produced, costs decline by a constant percentage — the “learning rate.”
// Wright's Law (1936) Cost_n = Cost_1 * n^(-b) // where b = -log2(learning_rate) // learning_rate = cost retained // after each doubling
Unlike Moore’s Law (time-based), Wright’s Law is production-based — costs fall because you build more, not because the calendar advances.
| Technology | Learning Rate | Cost Drop / Doubling |
|---|---|---|
| GPU Compute | 37.5% | 37.5% cheaper |
| Li-ion Batteries | 28% | 28% cheaper |
| Solar PV | 24% | 24% cheaper |
| Semiconductors | ~20-25% | ~20-25% cheaper |
GPU compute costs are declining 48% per year (ARK). Technologies on steep learning curves create exponential value — the question is who captures it.
The paradox: Unit cost of compute is plummeting, but frontier model training costs are skyrocketing — GPT-4 cost ~$100M, projected >$1B by 2027. Both statements are true: models are scaling faster than costs are falling. More compute per dollar, but we want exponentially more compute.
| HBM Gen | Year | $/GB | Capacity |
|---|---|---|---|
| HBM2 | 2016 | ~$20 | 4-8 GB/stack |
| HBM2e | 2020 | ~$10-15 | 16-32 GB |
| HBM3 | 2022 | ~$10-12 | 16-24 GB |
| HBM3e | 2024 | ~$12-15 | 24-36 GB |
| HBM4 | 2026 | ~$15-18 | 36+ GB |
HBM3e prices rose ~20% for 2026 orders. HBM4 production costs ~30% more than HBM3e. Demand so far outpaces supply that Wright’s Law can’t kick in.
HBM accounts for ~$3,200 of the B200’s ~$6,400 manufacturing cost. SK Hynix controls 57-62% of the market. This is the most profitable chokepoint in the stack.
AI data centers need megawatts to gigawatts. Cheap solar + batteries make remote data centers viable. Nuclear SMRs ($89-102/MWh) are being pursued for baseload — Microsoft restarted Three Mile Island.
Cumulative global power-generation capex must roughly triple to ~$10 trillion by 2030. Stationary energy storage must expand ~19×. Power is the binding constraint for AI scaling.
Stress-testing the thesis for every major player
The thesis: NVIDIA wins as long as AI spending accelerates. The risk is not that AI fades — it’s that the monopoly erodes from custom silicon below and open-source software above.
Advanced nodes (7nm and below): 77% of revenue. 3nm alone is 28%. TSMC is the most important company in the world that most people can’t name.
ASML has the most durable monopoly in tech. There is literally no alternative to EUV. Every sub-7nm chip on Earth requires their machines.
The key question: Is HBM structural or cyclical? If structural (AI demand never stops growing), SK Hynix at 7× forward is a generational opportunity. If cyclical business-as-usual, the market is right to be skeptical.
AMD & Broadcom are the credible challengers. AMD: OpenAI chose MI450 for 1 GW deployment. Broadcom: $73B AI backlog in custom silicon. Both at ~33× forward P/E.
Power is the real bottleneck. Not money, not demand, not land. Every watt of AI compute requires ~1.2-1.5 watts of total facility power.
The non-financial risks that could reshape everything
The uncomfortable truth: The entire AI revolution depends on a handful of companies in a handful of countries. A disruption to any one of these would cascade through the entire stack. This concentration is both the investment opportunity (monopoly pricing power) and the tail risk (geopolitical disruption).
TSMC produces >90% of the world’s most advanced chips. Advanced nodes (7nm and below) = 77% of TSMC revenue. 3nm alone is 28%.
Thesis: TSMC’s irreplaceability makes Taiwan too valuable to attack. But — as US, Japan, and EU build domestic fabs, the shield weakens. Taiwan fears “America First” policies may erode it.
Taiwan risk is unhedgeable at portfolio level — it would crash every tech stock simultaneously. The rational approach: size positions appropriately and diversify across the stack.
For investors: Export controls create both risk and opportunity. NVIDIA benefits from eased China access (H20, RTX PRO sales). But the regime is volatile — policy can shift with every administration. Companies with diversified geographic revenue (TSMC, ASML) are better positioned than those dependent on a single market.
| Project | Award | Status | Timeline |
|---|---|---|---|
| TSMC Arizona | $6.6B | Fab 1 at full capacity | 4nm shipping; 3nm/2nm Fab 2 installing |
| Intel Ohio | $8.5B | Delayed to 2030 | AZ Fab 52 hit milestone (18A process) |
| Samsung TX | $6.4B | Mass prod. late 2026 | Focused on 2nm GAA |
| TSMC Kumamoto | Japan-funded | Fab 1 operational | Fab 2 upgraded: 3nm by late 2027 |
| Rapidus (Japan) | $12B (gov’t) | 2nm pilot line active | Mass production target: 2027 |
| Intel Magdeburg (EU) | EUR 30B | Canceled (Aug 2025) | — |
Reality check: Japan has committed ~$65B to semiconductors — larger than the US CHIPS Act as a share of GDP. The EU’s 20% market share target by 2030 is “very unlikely” (EU Court of Auditors) — projected at ~11.7%. TSMC Arizona is the only unqualified success so far.
a16z, ARK, Sequoia — where they agree and disagree
| Metric | Projection |
|---|---|
| AI Software Market (2030) | $14 trillion |
| Data Center Capex (2030) | $1.4T/year |
| GPU Cost Decline (annual) | 48% compound |
| Inference Cost Decline (2024) | 99% in one year |
| Software Token Cost Drop | $3.50 → $0.32/M tokens |
| Robotics TAM | $26+ trillion |
| Energy Storage Expansion | ~19× needed |
Training costs improving at 50× Moore’s Law speed. But frontier training costs are rising: GPT-3 ($4.6M) → GPT-4 (~$100M) → projected >$1B by 2027. Models scale faster than costs fall.
“Driven by AI, today’s technology investment cycle could prove to be the biggest in history.”
2023: GPU supply was the bottleneck (NVIDIA monopoly). 2024: HBM supply (SK Hynix). 2025-26: Power & energy. 2027+: possibly application revenue (Sequoia’s $600B question).
Each earnings cycle reveals which bottleneck is binding. TSMC capex guidance, NVIDIA order backlog, hyperscaler capex announcements — these are the signals. The stack gives you the framework; the data gives you the timing.
| Phase | Railways (1840s) | Telecom (1990s) | AI (2020s) |
|---|---|---|---|
| Trigger | Steam locomotive | Internet + fiber | Transformer + GPUs |
| Peak investment | ~1/3 of UK GDP | >$1T debt | ~$650B/yr capex |
| Market concentration | 71% of stock market | Telecom dominated Nasdaq | Mag 7 dominate S&P |
| Overcapacity | 1/3 of lines never built | <5% of fiber lit | $600B revenue gap |
| Bust | 50% share decline | $2T market cap destroyed | TBD |
| Long-term outcome | Rails enabled industrialization | Fiber enabled the internet | Infra enables AI economy? |
The pattern: Overbuild → bust → the infrastructure enables something bigger than anyone imagined. Key difference: this cycle is funded by cash-rich hyperscalers, not leveraged startups. The bust may be less acute but the overbuilding more persistent.
Investment Frameworks — agent37.club
agent37.club