EPOCHAL(00)

Constraint Alpha

Capital crowds into the story. Returns accrue to the bottleneck.

01 / THE NAME

Why “Constraint”

The defining event of the coming decade is a collision: the fastest compounding process ever created is meeting the slowest supply chains in the industrial economy. We invest in the collision.

THE ENGINE — THE INTELLIGENCE CLOCK

AI research now manufactures its own scarcest input. Capability compounds quarterly — and consumes the physical world as fuel.

THE QUEUE — THE STEEL CLOCK

Four-year transformer lead times, sold-out turbine slots, decade-long licenses, interconnection queues holding multiples of installed capacity.

MANDATE — AI infrastructure · deep tech · critical-resource supply chains · North America & Europe
02 / FOUNDER

From Operator to Allocator

2014–2019

Soar Aviation

Scaled 1 → 55 aircraft on prepaid demand and internal cashflow. Sold 50% in July 2018 at a ≈65M enterprise valuation.

LESSON: Unit economics determine outcomes.
2020–2023

Stratton Car Finance

Acquired ~1/3 ownership. Revenue grew 45M → 82M. Exited to a listed buyer at a 121M enterprise valuation.

LESSON: Restraint. Reduce friction, not add activity.
2022–PRESENT

IREN

Accumulated through deep dislocation to become one of the company’s largest shareholders. Realized partially as the business matured.

LESSON: Temperament. Patience is the scarcest asset.
03 / CAPITAL COMPOUNDING

Track Record

/ CUMULATIVE · 1 JAN 2019 – 17 JUN 2026
7.07x
+606.5% CUMULATIVE
29.8%
CAGR · ~7.5 YRS
(38%)
MAX DD · 2 NEG YRS OF 7
/ CUMULATIVE GROWTH OF 1.00 — QUARTERLY CLOSES, SUPPLIED SERIES
1x3x5x7x201920202021202220232024202520267.07x AT 17 JUN 2026

Proprietary capital of Neel Khokhani · manager-supplied series, unaudited · not the performance of any fund — Epochal: Constraint Alpha Inc. (est. 2026) has no performance history · full performance record available for due-diligence review.

04 / METHOD

The Research Process

  1. /01Primary sources over consensusNameplates, permit databases, satellite passes, grid filings. One transformer wound for 415V — the hyperscalers’ global spec — told us what a Texas site was, for whom, and on what schedule, before any disclosure did.
  2. /02Model everything ourselvesOwn build-ups of megawatts, wafer starts, and cash flows — including reverse-engineering what the current price implies.
  3. /03Follow the capital structureWhen bonds and equity price the same physical asset differently, one market is mispricing the scarcity. The gap itself is the trade.
  4. /04Stress-test the bear caseA written drawdown protocol: numbered stages, six hypothesis categories, run against the strongest available bear material.
  5. /05Governance as researchCompensation campaigns, regulator submissions, board correspondence. Acting as an owner sharpens the read on management quality.
EVERY IDEA CLEARS A 17–20% REAL HURDLE, OR IT DOES NOT ENTER THE BOOK.
05 / PORTFOLIO THEMES

Four Themes, One Stack

WHY NOW — Peak Scarcity arrives 2027–2030, when strategic demand meets pre-automation supply. Today’s prices are still set by commercial bidders. The window is the entry.

THEME 01

Datacenters

AI-native operators and powered-shell landlords — the new REITs.

CONSTRAINT: ENERGIZED MW + INTERCONNECT
THEME 02

Datacenter EPC

The engineering and electrical contractors gating every buildout.

CONSTRAINT: GRID EQUIPMENT + SKILLED BUILD
THEME 03

Memory

High-bandwidth memory structurally short into 2027–28.

CONSTRAINT: HBM SUPPLY
THEME 04

Niche Semiconductors

Epitaxial wafers — the choke point beneath the entire stack.

CONSTRAINT: EPI WAFER CAPACITY
THEME 01 / DATACENTERS

The binding variable is not chips, capital, or demand. It is energized megawatts with secured interconnect.

Compute is fungible and financeable. A substation position in an interconnection queue is not.

CHIPS — depreciating, substitutable, collateralizable
CAPITAL — abundant at every layer of the structure
POWER — THE 2026 CLEARING PRICE AT THE FRONT OF THE QUEUE: ~$50M PER MEGAWATT-YEAR
THEME 01 / DATACENTERS

AI-Native Operators

Operators that own the full stack — land, power, interconnect, and the compute on top. The equity behaves like an infrastructure developer, priced like a tech stock in both directions.

The opportunity recurs wherever the market prices the business model and ignores the physical asset underneath it.

CASE STUDY / IREN · ENTERED LATE 2022
MARKET SAW
A bitcoin miner at a $100–200M market cap, deeply out of favour
WE UNDERWROTE
Land, energized power, and secured interconnect across Texas and British Columbia
OWNERSHIP
Accumulated patiently to become one of the largest shareholders; active on compensation, regulator submissions, board correspondence
OUTCOME
Compounded materially as the buildout energized; partially realized, still a core holding
THEME 01 / DATACENTERS

Powered Shells — The New REITs

Own the land, the substation, and the shell. Lease it to a hyperscaler on a 10–15 year contract. Collect rent.

These are real estate cash flows with investment-grade tenants — the same economics that built every listed REIT. The market has not yet re-filed them under that heading.

When it does, the repricing is the return on top of the yield.

15%+
UNLEVERED IRR, MID-TEENS PER ANNUM
THE BUILD-UP, ILLUSTRATIVE PER MW
All-in capex — land, substation, shell
~$10–12M
Triple-net rent, 10–15 yr lease, IG tenant
~$1.5–1.8M / YR
Cash yield on cost, before any leverage
MID-TEENS
THEME 01 / DATACENTERS

The Commercial Building Comparison

PRIME COMMERCIAL BUILDINGPOWERED SHELL
TENANTInvestment-grade corporateInvestment-grade hyperscaler
LEASE TERM10–15 years10–15 years
ASSETLocation, steel, glassLand, substation, secured interconnect
SUPPLY RESPONSECapital builds more within 2–3 yearsGated by multi-year interconnect queues
UNLEVERED IRR5–6%Mid-teens

A commercial building offering a mid-teens unlevered yield would be arbitraged away within weeks. Powered shells offer it today — because the scarce input cannot be replicated by capital alone.

THEME 01 / DATACENTERS

Regulation as a Tradeable Constraint

In 2025 the Texas grid operator proposed reclassifying its interconnection queue — ERCOT PGRR 145, “Batch Zero.” A paragraph of planning-guide language that redraws who gets energized, and when.

We treated it as a portfolio-level event: every holding’s queue position mapped, grandfathered versus exposed, with the P&L implications modeled before the market reacted.

/ WHY THIS WORKS

Most investors do not read planning-guide revision requests.

SAME TOOLKIT: PERMIT DATABASES · NAMEPLATE REGISTRIES · SATELLITE IMAGERY
THEME 02 / DATACENTER EPC

The Builders Behind the Buildout

Every gigawatt of datacenter requires substations, high-voltage work, and mechanical-electrical fit-out — regardless of which operator or chip wins.

Some queues are slow because of knowledge — software compresses those. Some are slow because of matter and institutions: forging capacity, thermal-cycling tests, licensing calendars. No amount of intelligence shortens them; they inherit the scarcity draining out of everything else.

We own the matter-limited side: contractors and grid-equipment makers with contracted, power-related backlog — the theme without underwriting any single tenant.

THE QUEUE, MEASURED · MID-2026 ORDER-TO-DELIVERY
Large power transformer, 500kV class
3–5 YEARS ↑
Heavy-duty gas turbine delivery slot
SOLD OUT TO ~2030 ↑
Grain-oriented electrical steel allocation
12–24 MONTHS ↑
US grid interconnection study, major ISOs
MULTI-YEAR ↑
FROM OUR OWN PROCUREMENT CONVERSATIONS, MANUFACTURER DISCLOSURES & INDUSTRY REPORTING · ALL LENGTHENING SINCE 2024
THEME 03 / MEMORY

The HBM Shortage

High-bandwidth memory is the one component of the AI server that cannot be over-ordered into existence. Supply is structurally short into 2027–28; new fabs arrive too late to break the cycle.

Our position: SK Hynix, the share leader, under long-term agreements with the hyperscalers — entered at 4–5x annualized operating profit during the drawdown.

We own the Seoul ordinaries and the Frankfurt GDR, deliberately avoiding a US listing that traded at an unarbitrageable ~22% premium after its Nasdaq debut. Even the choice of line is constraint work.

~60%
HBM SHARE HELD BY THE LEADER
4–5x
ENTRY MULTIPLE, ANNUALIZED OP
2027–28
SUPPLY REMAINS STRUCTURALLY SHORT
THEME 03 / MEMORY

A Thesis Under Stress

−55%
SK HYNIX DRAWDOWN, JUN–JUL 2026
Added
POSITION INCREASED INTO THE TROUGH
  1. S.1Ran the written drawdown protocol — numbered stages, six hypothesis categories — against the strongest available bear material and the earnings call.
  2. S.2Audited the bulls’ own client model — and found a 30-point HBM share-erosion assumption their published articles never confront.
  3. S.3Reverse-engineered the trough price: it implied trough-quarter profits held in perpetuity, with no cycle recovery.
  4. S.4Verdict: multiple compression and technical flow unwind — not thesis impairment. Long-term agreements with ~10 customers directly rebutted the bear case.
THEME 04 / NICHE SEMICONDUCTORS

The Epi Wafer Basket

Every leading-edge chip begins as an epitaxial wafer. Four companies make them at scale. Capacity takes years to build, new fabs are paused or underutilized, and long-term agreements reset in 2027.

Consensus models flat pricing into that reset. The supply-demand work says otherwise — and that delta is the trade.

We hold three of the four global makers as a basket: the choke point, not the logos.

250k 1M
LEADING-EDGE EPI DEMAND, WPM · CY23 → CY28

A 50GW AI-buildout scenario implies ~1.5M wpm — the entire epi supply that exists today.

THEME 04 / NICHE SEMICONDUCTORS

Sizing the Basket

THE CONVEXITY LEG

A new fab running at ~20% utilization with heavy fixed depreciation. If pricing pulls fill rates to 60–70%, EBITDA roughly triples. Binary, mathematically extreme.

HIGHEST-CONVICTION ADD · 2–5x RANGE
THE PURE-PLAY

~60% share in leading-edge epi, structurally under-built into the shortage, with the key customer agreement expiring exactly at the 2027 pricing reset.

CLEANEST EXPRESSION · 2–2.5x
THE SHARE LEADER

Most swing capacity, present in all five geographies, government-subsidy de-risked. Best risk-adjusted leg of the three.

LARGEST BASKET POSITION · 1.5–2.5x

The initial size-up was cut from 65–75% to 35–40% once each leg was scored on entry point, not thesis. A fourth maker was evaluated and deliberately skipped.

Sizing follows the hierarchy, not the story.

06 / RISK

The Operating System

Concentration with explicit discipline: high-conviction positions governed by written rules, not judgment in the moment.

ORIGINATION PROTOCOLS

What qualifies for research, written down before the idea arrives.

SIZING RULES & CEILINGS

Delta-adjusted single-name limits, checked against every statement.

MANDATORY FACTOR HEDGES

A standing short book against the factor exposure the longs create.

DRAWDOWN PLAYBOOK

The written protocol from the Hynix review — run every time, not when convenient.

TAIL INSURANCE

Index puts held permanently against the concentrated book.

17–20% REAL HURDLE

The filter at the door. Ideas that clear it are rare; that is the point.

07 / THE RECORD, FORWARD

The Ledger

Ten dated, falsifiable expectations, graded publicly every August. A map you cannot be wrong about is marketing.

  1. 01END-2027Agent task horizons exceed one work-month; a frontier lab attributes most research-engineering output to its own models
  2. 022028A G7 or Gulf state treats compute as formal strategic infrastructure — allocation authority, not merely subsidy
  3. 032027–30A >40% compute-complex drawdown on an efficiency or geopolitical shock — then new highs in utilization and pricing within eight quarters
  4. 042029Distributed training at frontier scale; the “coherent campus” premium compresses materially
  5. 052030Matter-limited chokepoints outperform knowledge-limited infrastructure; a celebrated 2026 category is revealed as a false chokepoint
  6. 062030Verification, assurance, and AI-liability capacity emerges as a recognized asset class
  7. 072031Robots cross one million units annually; an actuation-chain component repeats the 2024 transformer repricing
  8. 082031Automated-lab throughput becomes a named institutional category; trial capacity is the acknowledged bottleneck of AI-era medicine
  9. 092032Robotic construction and machine-run qualification compress a flagship project schedule by more than a third vs 2025
  10. 102033–35Automated production adds marginal capacity in a former chokepoint; pricing collapses toward energy-plus-materials cost
First annual grading: August 2027
EPOCHAL(00)/ CLOSE

We invest with patience, conviction, and purpose.