Three Forces, One Periodic Table

The Manifesto: Three Forces, One Periodic Table

A force, in this casefile, is a structural demand pattern that compounds on its own clock and does not retreat with normal economic cycles. Four such forces are visible in 2026.

Force 1 — Recursive AI compute growth. Each token-cost halving opens a new tier of economically viable workloads. Customer support routing arrived at $10 per million tokens; pull-request-level code review at $1; persistent edge agents at $0.10. Each tier compounds inference volume into the same constrained physical stack. Software adoption moves in weeks; the periodic table moves in years.

Force 2 — National security demand floor. Three statutory instruments put a price-insensitive bid under defense, energy, and quantum buys. NDAA Section 854 phases in procurement prohibitions on covered critical minerals from countries of concern through January 1, 2030.¹⁵ DPA Title III (Defense Production Act) has produced direct federal investment in rare-earth processing capacity (MP Materials), lithium (Lithium Nevada), and germanium processing (5N Plus).¹⁶ The CHIPS Act $39 billion in manufacturing incentives carries a 10-year clawback that locks recipients to compliance with country-of-concern restrictions.¹⁴ These are not subsidies that get repealed in budget bills. They are procurement and capital structures the federal government enforces by statute.

Force 3 — Energy transition, quantum-national-security, and reshoring. The third-force pull comes from grid hardening (transformers, GOES (grain-oriented electrical steel), copper), advanced nuclear (HALEU (High-Assay Low-Enriched Uranium), zirconium), and quantum-national-security stack (cryogenic dilution refrigerators, low-noise amplifiers, post-quantum cryptography hardware). Each is funded by a different statutory or executive vehicle — IRA (Inflation Reduction Act) Sections 30D and 45X for batteries and minerals, DOE LPO (Loan Programs Office) for nuclear and grid, IARPA / NSA classified programs for quantum. Different communities, different conferences, same upstream materials.

The grid-hardening pull is now visible at the RTO level. PJM Interconnection — the largest US RTO, serving 65M people across 13 states + DC — held its 2027/28 Base Residual Auction in December 2025 and was unable to procure enough capacity to meet the reliability target, falling 6,623 MW short. PJM CEO David Mills, in a May 8 2026 stakeholder letter, used regulator-grade language: "an era of scarcity," "structurally different from prior periods of tightness," "the current situation is not tenable." The same MISO 2025/26 capacity auction cleared at .50/MW-day, 22x year-over-year. NYISO winter capacity-energy risk is documented. ERCOT has 2,000+ interconnection requests in queue.²³ The energy-transition demand wave is binding before new generation is built — meaning grid-material chokepoints (large transformers, GOES, copper, NdFeB magnets) are getting absorbed by AI data center load before energy transition even competes for them. This sharpens master claim 3F-M1: the materials don't sequence by demand domain. They bind simultaneously.

The asymmetric counter-force — China’s counter-flywheel. Three years of observable data show that each major Western reshoring announcement provokes a calibrated PRC export-control response on the next-most-strategic upstream material. August 2023 brought gallium and germanium controls. December 2024 brought graphite and antimony tightening. 2025–2026 brought tungsten product-form restrictions and rare-earth processing technology controls. The bifurcation cuts both ways: every Western response widens the moat for non-China qualified suppliers.

The policy world treats these as four separate supply chain challenges. The dependency graph shows they are one challenge viewed from four angles. Naming the same material in four different policy memos does not make it four problems. The convergent periodic table — fifteen materials, four demand columns, sixty cells — is the figure that makes this concrete.

The rest of this casefile builds that figure, tests it against falsifiable claims, and identifies the public companies that sit at the convergent neck.

The Hourglass Topology

The structural image for every constrained material in this casefile is an hourglass. Upstream supply is wide. Downstream demand is wide. The neck — the qualified, accredited, slow-to-expand producer or processor — stays fixed while the bottom keeps growing. New downstream consumers get added every year; new qualified upstream suppliers get added on a 7-to-15-year mining-and-permitting timescale.

The pattern is observable across material families that share no chemistry, no geography, and no end-market.

InP (indium phosphide) substrates: indium feedstock is geographically concentrated (China refines roughly 60%); the qualified merchant neck in the West is essentially one company (AXTI, with 18-to-24-month customer qualification cycles); downstream demand has expanded from telecom lasers (2022) to AI datacenter optical interconnects at 400G / 800G / 1.6T (2025) to projected co-packaged optics, photonic compute, and defense LiDAR (2028).

UHP HF (ultra-high-purity hydrofluoric acid): fluorspar feedstock has 60%-plus China production share; the qualified merchant neck for the highest semiconductor purity grades is two Japanese suppliers (Stella Chemifa and Kanto Denka); downstream demand has expanded from legacy-node cleaning (2022) to EUV multi-patterning etch for AI-node chips (2025) to four projected CHIPS Act fabs at full operation (2028).

NdFeB sintered magnets: rare-earth oxide feedstock is 85-to-90% China-refined (neodymium, praseodymium, dysprosium, terbium); the qualified magnet-manufacturing neck is 90%-plus China today, with MP Materials and Lynas building Western capacity; downstream demand has expanded from wind turbines, hard drives, and defense (2022) to datacenter cooling fans and EV drivetrain motors (2025) to projected humanoid actuators, AUKUS submarine components, and hypersonic guidance systems (2028).

The figure below shows all three side-by-side. The necks stay the same width. The bottoms get wider. The same pattern is true of the other twelve materials in the convergent periodic table grid (Section 7); these three are shown because they sit in different chemistries and on different policy clocks, and they make the topology concrete.

Two reinforcing loops run through every neck simultaneously.

The cost-demand spiral. Falling unit costs (a token, a watt-hour, a megawatt of grid capacity, a magnet kilogram) drive more deployment. More deployment drives more compute investment, more turbine installation, more datacenter buildout, more humanoid prototyping. Scale economies in manufacturing reduce per-unit costs further, which drives more deployment. The loop is bounded by the neck — physical capacity cannot expand at software speed — but within that constraint it is self-reinforcing. Every halving of token cost, every dollar-per-watt drop in solar PV, every dollar-per-kilogram fall in NdFeB has produced a measurable expansion of the downstream addressable market.

The security-investment spiral. Geopolitical tension produces export controls. Export controls produce reshoring mandates. Reshoring mandates attract government capital — the CHIPS Act, the EU Chips Act, NDAA Section 854 procurement-prohibition phasing, DPA Title III invocations. Government capital accelerates buildout. More buildout creates more demand on the same constrained producers. Further supply-chain awareness produces more policy. The loop is lumpy and politically variable, but it creates demand floors that persist across normal commercial cycles. When the commercial AI cycle dips, the floor holds — that is the entire point.

The hourglass is the master metaphor for the rest of the casefile. Every claim, every grid cell, every bear case below maps to a specific neck whose width is changing on a different clock than the demand pouring through it.

Force 1 — Recursive AI Compute Growth

AI is the first technology that generates demand for more of itself. Oil does not discover new uses for oil. Copper does not design systems that need more copper. AI agents are increasingly constructing the software, workflows, scientific experiments, and business processes that demand more AI.

This is not infinite recursion. It is a step-function multiplier. Each cost halving opens a new tier of economically viable workloads. At $10 per million tokens: customer support routing, document summarization, basic code generation. At $1: continuous code review, monitoring, testing, documentation at the pull-request level. At $0.10: embedded in every microservice, every API gateway, every operational alert. At $0.01: persistent agents running autonomously at the edge. Each threshold unlocks a new addressable market that was previously too expensive to serve. The total multiplier across those broadening markets is multi-fold across the current cycle, bounded by physical throughput capacity at every layer of the stack.

Depth of useful agent chains is bounded by practical limits: error accumulates by hop four or five in most production workflows, latency budgets cap at 200 to 500 milliseconds for user-facing applications, and cost ceilings force chain termination before infinite recursion. The real driver is not depth of chains — it is breadth of use cases. Every workload tier that crosses the economic threshold adds to aggregate inference volume regardless of whether any individual chain runs deep.

The demand curve is logistic with stepped plateaus, not exponential. Each plateau corresponds to a physical capacity ceiling. Fab construction runs three to five years from groundbreaking to production volume. HBM (High-Bandwidth Memory) allocation cycles run eighteen months. Power infrastructure runs thirty-six to forty-eight months. Substrate qualification cycles run eighteen to twenty-four months. The volume does not compound indefinitely in a single curve — it compounds in steps, and each step encounters the same physical stack of constrained materials.

The constrained physical stack is well-mapped. Token Value Compounding walks through eight layers in detail: compute (GPUs, captive ASICs), memory (HBM, DDR5), advanced packaging (CoWoS (Chip on Wafer on Substrate)), power (transformers, switchgear), cooling (liquid cooling, NdFeB fan motors), substrates (InP, GaAs), photoresist and ultra-high-purity chemicals, and EDA (Electronic Design Automation) software. Each layer has a constrained neck. Each constrained neck has a small set of qualified merchant suppliers. Each set of qualified merchant suppliers has investable public-company equivalents. The eight-layer breakdown lives there; this page abstracts upward to the cross-domain pattern.

Two implications follow from the recursive pattern. First, AI demand for the upstream stack does not behave like ordinary commercial demand: it has the structure of a feedback loop, not an isolated end-market. When a frontier model crosses a cost threshold and unlocks a new workload tier, that workload tier proceeds to build the next frontier model, the next training run, the next inference fleet. The demand pulls itself forward. Second, the pull rate exceeds the qualified-supplier expansion rate by an order of magnitude on multiple necks. New CoWoS lines, new HBM fabs, new InP substrate capacity, new NdFeB magnet plants — each takes years to bring online while the demand-side workload tier compounds quarterly.

The fastest-moving technology in human history is systematically increasing the value of the slowest-moving assets in the economy.

Force 2 — National Security Demand Floor (Defense + Energy + Quantum)

Commercial demand is price-elastic. If token costs rise or AI adoption slows, enterprises defer deployments. National security demand is largely price-inelastic: a defense program manager complying with a statutory procurement requirement buys the compliant material at whatever the market-clearing price is. A DOE LPO loan recipient bringing a new SMR (small modular reactor) online buys HALEU at whatever Centrus charges. A NASA-and-NSA quantum hardware program buys cryogenic dilution refrigerators at whatever Bluefors and Oxford Instruments quote.

Commercial AI demand is price-elastic. National security procurement is not. When both compete for the same constrained input, the bottleneck gets a demand floor.

Tier 1 — Legislatively durable, bipartisan.

The CHIPS and Science Act (Public Law 117-167, August 2022) committed $39 billion in manufacturing incentives.¹⁴ The clawback provision ties recipients to a 10-year period of restrictions on expanding capacity in countries of concern. The fabs currently under construction — TSMC Arizona, Samsung Taylor, Intel Ohio, Micron Boise — lock in multi-decade materials demand regardless of any single administration’s policy preferences. Concrete has been poured. Agreements have been signed. The clawback gives the government legal standing to enforce compliance for a decade.

NDAA Section 854 prohibits Department of Defense procurement of covered critical minerals from countries of concern, with effective dates phasing in through January 1, 2027, and expanding through January 1, 2030.¹⁵ This is not a subsidy that gets repealed in a budget bill. It is a procurement prohibition baked into the annual defense authorization, which has passed every year since 1961 without missing a single fiscal year. The DoD must source compliant materials regardless of cost differential. That is a price-insensitive demand floor by statute, with the institutional weight of an annual must-pass bill behind it.

Defense Production Act Title III — permanent authority since 1950 and reauthorized continuously — has recently produced direct federal investment in rare-earth processing capacity at MP Materials ($35 million), a separate lithium processing award of $11.8 million to Lithium Nevada, and germanium processing capacity expansion at 5N Plus.¹⁶ These invocations represent the federal government taking a direct economic position in critical-material supply chains, at whatever cost is necessary to achieve domestic security of supply. The instrument is permanent; the allocations get refreshed each fiscal year.

The HALEU Availability Program at the Department of Energy committed $700 million to bootstrap a non-Russian supply of high-assay low-enriched uranium for advanced reactors.²¹ Centrus Energy, the only US company with a deployed HALEU enrichment cascade, sits at this neck.

Tier 2 — Incentive-driven, politically variable.

The IRA critical-mineral provisions (sections 30D and 45X) represent the largest single demand instrument for several material categories — battery-grade lithium, cobalt, nickel, graphite — but their durability depends on legislative and executive priorities that move with election cycles. The DOE Loan Programs Office carries $400 billion-plus in lending authority; execution depends on administration staffing and priorities. These instruments are significant but should be weighted differently than the statutory floors above.

The asymmetry between Tier 1 and Tier 2 is the operating distinction this casefile makes. Statutory floors carry the demand through political cycles; incentive programs amplify it during friendly cycles and recede during hostile ones. The investable thesis depends primarily on the Tier 1 floor; Tier 2 is the upside multiplier. Bear D below tests the political-fracture scenario where the Tier 1 floor partially erodes — the asymmetric falsifier this casefile is most concerned with.

The figure shows commercial AI buildout, defense procurement, energy transition, and quantum-national-security stacked across years. The qualified-supplier ceiling does not move on the same clock. When commercial demand dips, the floor holds.

Force 3 — China’s Counter-Flywheel

The asymmetric force in this casefile is not a Western policy instrument. It is China’s response function to Western reshoring. Three years of observable data show a calibrated escalation pattern, not a categorical denial pattern.

August 2023: gallium and germanium export licensing tightened.¹⁷ December 2024: graphite and antimony added to the controlled list. 2025: tungsten product-form restrictions and rare-earth processing technology export controls. 2026: indium tightening indicators and rare-earth deeper-stage processing controls under discussion. Each new line item lands on the next-most-strategic upstream material that Western capacity cannot yet replicate at scale. The escalation is calibrated — partial restrictions, license-based rather than embargoes, leaving room for diplomatic adjustment — and it is observable as a recurring pattern.

The regime assumption this casefile operates under: Persistent US-China technology competition, no full decoupling resolution, no Chinese advanced-technology breakthrough, no US policy reversal. If any of these four assumptions break, cascade-model claim 3F-L8 below catches it (bipartisan congressional consensus on technology competition holds through end-2027), and the entire investable window framework changes shape. Cascade claim 3F-M2 tracks the escalation pattern directly.

The diagram above traces the loop: PRC export controls tighten → Western reshoring announces ($CHIPS, MP NdFeB build, Centrus HALEU, Lynas heavy-rare-earth processing in Texas) → reshoring demand spike on remaining qualified Western suppliers → PRC adds the next control layer. The loop runs in both directions. Bifurcation cuts both ways.

For Western non-China qualified suppliers, the counter-flywheel produces something the textbook commodity model does not: a structural moat that widens with each escalation. A new control event in 2026 raises the perceived value of every existing non-China qualified producer in that material category. The 5N Plus (VNP) germanium processing capacity becomes more valuable when China tightens germanium quotas. The MP Materials NdFeB magnet ramp at Independence, Texas becomes more valuable when China processes a new rare-earth restriction. AXTI’s InP substrate position becomes more valuable when indium controls tighten. The pattern is observable retrospectively across every major control event since August 2023.

For Chinese producers, the counter-flywheel produces a slow-motion concentration risk. As Western capacity ramps in adjacent materials, China’s leverage in any single material weakens. The strategic response — which is what 2025–2026 has shown — is to escalate to deeper-stage processing controls (refining technology, separation chemistry, sintered-magnet manufacturing know-how) rather than headline material flows. This is observable in the recent shift from raw-material restrictions to processing-technology export controls.

The investable implication is direct: every Western reshoring announcement in a new material category should be tracked for the calibrated PRC response in the same or adjacent category within 18 months. Cascade claim 3F-M2 monitors this. Bear C below tests what happens if the loop breaks — if China backs down or the West backs down. Through 2027, the base case is the loop continues.

The Convergent Periodic Table

This is the load-bearing figure of the casefile. Fifteen materials. Four demand columns. Sixty cells. Each cell is encoded by status: red (binding constraint today), amber (tightening, 12-to-24-month visible squeeze), blue (demand visible, neck not yet binding), gray (domain does not pull this material). Each red cell carries a chokepoint company tooltip — the public-company equivalent that captures the position implication.

The policy world treats these as four separate supply chain challenges. The dependency graph shows they are one challenge viewed from four angles. The grid is what that sentence means in practice.

The China-controlled top. Rows 1 to 7 are materials where China controls 60-to-90% of refined or processed supply. Gallium, germanium, NdFeB rare earths (neodymium, praseodymium, dysprosium, terbium), indium, cobalt, samarium-cobalt, and tungsten are all in this band. AI, defense, and quantum-NatSec all draw on at least three of these; energy draws on most of them. The convergence is not coincidence. The Chinese strategic mineral policy of the past decade picked precisely the upstream materials that the four demand domains all need.

The single-source middle. HALEU is a row of its own — Russia controlled 95% of pre-war commercial supply, the US is bootstrapping with Centrus as the sole deployed cascade, and the demand pull comes from advanced reactors (energy), Navy reactors (defense), microreactor datacenter pilots (AI), and quantum cryogenic infrastructure (NatSec). The chokepoint is not a Chinese export control. It is a single qualified producer in a domain where qualification cycles run a decade.

The Western chokepoints inverse pattern. Rows 14 and 15 — InP substrates and EUV photoresists — show the inverse pattern. China refines under 10% of indium phosphide substrates; under 5% of EUV-grade photoresists. The Western neck is the constraint. AXTI is the sole US merchant InP supplier. JSR, Tokyo Ohka Kogyo (TOK), Shin-Etsu, and Fujifilm collectively dominate EUV photoresist. The hourglass topology operates in both directions: any qualified merchant neck — Western or Chinese — pulled by all four domains is a chokepoint.

The amber middle. Copper, lithium, tantalum, and silicon-carbide (SiC) substrates sit in the tightening band. China-share is moderate (25-to-60%); demand is growing at 12-to-24-month visible squeeze rate; the neck is holding but the gap is closing. These are watch-list materials, not yet binding for all four domains today, likely binding for at least two domains by end-2027.

Of the 60 cells in the grid, roughly one third are red. Each red cell is a position implication. Each position implication has a chokepoint company. Each chokepoint company has its own playbook in our database; the cascade hypothesis registry below tracks the falsification conditions that would break the convergence thesis for each.

The Investable Window: 2026–2032

Three clocks tick at different speeds. Software adoption moves in weeks. Physical capacity moves in years. Policy moves in election cycles. The investable thesis lives in the gap between the fastest clock and the slowest.

The window opens in 2026 with NDAA Section 854 effective dates phasing in.¹⁵ It widens through 2027 as the first round of CHIPS Act fab pilots — TSMC Arizona, Samsung Taylor — reach early production volume. It widens further in 2028 as Section 854 expansion provisions trigger and DOE LPO HALEU production targets get tested against actual delivery. It begins to close around 2030 as Section 854 reaches full-rate effective and the bulk of CHIPS-funded capacity is operational. By 2032, either reshoring catches up enough that the necks widen materially — in which case the convergence thesis weakens and the position implications change — or it does not, and the same chokepoint companies remain investable.

Token Value Compounding tracks the AI side of the window in detail. NDAA Rare Earth tracks the Section 854 timeline. Tungsten Squeeze tracks W-specific dynamics. The 13 Atoms tracks the structural quad-convergence framework. This casefile covers the convergence layer that sits above all three.

The temporal mismatch is the structural source of the thesis. NEPA (National Environmental Policy Act) permitting for new domestic mines runs 7 to 15 years from filing to production.¹⁹ AI doubling cadence runs 12 to 24 months. Defense statutory deadlines run 36 to 60 months from authorization. The qualified-supplier expansion clock cannot match the demand-pull clock on any single material. That is the entire content of the hourglass topology.

What moves the window? On the bullish side: any new statutory floor (a Magnet Investment Tax Credit, an expanded HALEU procurement, a quantum-supply-chain authorization). On the bearish side: any partial erosion of an existing floor (a CHIPS clawback waiver, an NDAA Section 854 carve-out, a Senate floor vote against the consensus). Cascade claims 3F-L1 through 3F-L8 monitor the specific events that would shift the window’s open and close dates. The Pro-gated registry below contains the live source URLs and current confidence values.

Bear Cases

Every casefile in this house lists the conditions under which it would be wrong. Four bears sit between the manifesto and the hypothesis registry. Each is tied to a specific cascade-claim falsification trigger.

Bear A — Substitution wins

A qualified alternative material appears inside one of the fifteen grid rows before the investable window closes — peer-reviewed or commercial-scale, removing 20% or more of demand on a binding material in any single domain. Cascade claim 3F-L6 monitors this. Probability is low for 2026 to 2028 (qualification cycles run 18 to 36 months; we would already see qualification work in motion if it were going to land before 2028) and rising for 2029 to 2032. Position implication on confirmation: trim the affected chokepoint by 25 to 50%, maintain optionality on the remaining materials. Bear A tests the assumption that the hourglass neck cannot be widened by a chemistry workaround on the demand side.

Bear B — The convergence is illusion

The four demand domains are coincidentally pulling the same materials, and the structural co-dependency is a narrative fit, not a causal map. Falsified by year-end-2027 grid analysis showing fewer than two cells where any single material is binding for three or more domains simultaneously — verified against published USGS Mineral Commodity Summaries, Adamas Intelligence quarterly tracking, and DOE Critical Materials Assessment data so the test is externally checkable, not internally judged. Probability moderate. This is the strongest skeptic frame and the most honest counter to the casefile. Position implication on confirmation: re-evaluate the cross-domain optionality premium. Single-domain plays still hold, but the manifesto-level thesis weakens to a collection of separate theses.

Bear C — China backs down or the West backs down

A negotiated settlement reduces export controls, or US political consensus on China policy shifts producing a partial floor erosion. Cascade claim 3F-L8 catches a Senate-floor rollback vote (the floor, not committee). Probability low-moderate for 2026 to 2028. Position implication on confirmation: reduce defense and NatSec demand-floor weight in the thesis. Commercial AI thesis still holds intact (the TVC casefile remains valid). Bear C tests the assumption that the security-investment spiral keeps running.

Bear D — National Security Consensus Fractures

The asymmetric falsifier. A major political shift fractures the bipartisan consensus underpinning NDAA Section 854, DPA Title III, and CHIPS clawback enforcement. The statutory floor partially erodes — not all-or-nothing, slow-motion, no black-swan event required. Triggered by cascade claims 3F-L1 (NDAA Section 854 carve-out) or 3F-L8 (Senate floor vote against the consensus). Probability low through 2027; rises post-2028 election cycles. Position implication on confirmation: statutory-floor weight in the thesis weakens; commercial AI demand still holds; reduce pure NatSec exposures (Centrus AMRC, defense suppliers without dual-use commercial revenue). Bear D is the specific scenario this casefile is most concerned with — the West unilaterally erodes the floor it built.

Pro: Hypothesis Registry Across Domains

This casefile is monitored by a 10-claim cascade model named three-forces-convergence. Two master claims (3F-M1, 3F-M2) and eight linked claims (3F-L1 through 3F-L8) are each tagged with: a falsification condition, a primary-source set, a price-sensitivity rating from 1 to 5, a position implication, and a current verdict updated by the cascade runner each night at 02:30 UTC.

Two master claims visible to all readers below. The remaining eight linked claims, with current source URLs, current confidence values, and exact position implications by ticker, are gated to Pro subscribers. Each claim links to its current verdict on /performance/.