Constraint Alpha
Capital crowds into the story. Returns accrue to the bottleneck.
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.
AI research now manufactures its own scarcest input. Capability compounds quarterly — and consumes the physical world as fuel.
Four-year transformer lead times, sold-out turbine slots, decade-long licenses, interconnection queues holding multiples of installed capacity.
From Operator to Allocator
Soar Aviation
Scaled 1 → 55 aircraft on prepaid demand and internal cashflow. Sold 50% in July 2018 at a ≈65M enterprise valuation.
Stratton Car Finance
Acquired ~1/3 ownership. Revenue grew 45M → 82M. Exited to a listed buyer at a 121M enterprise valuation.
IREN
Accumulated through deep dislocation to become one of the company’s largest shareholders. Realized partially as the business matured.
Track Record
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.
The Research Process
- /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.
- /02Model everything ourselvesOwn build-ups of megawatts, wafer starts, and cash flows — including reverse-engineering what the current price implies.
- /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.
- /04Stress-test the bear caseA written drawdown protocol: numbered stages, six hypothesis categories, run against the strongest available bear material.
- /05Governance as researchCompensation campaigns, regulator submissions, board correspondence. Acting as an owner sharpens the read on management quality.
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.
Datacenters
AI-native operators and powered-shell landlords — the new REITs.
Datacenter EPC
The engineering and electrical contractors gating every buildout.
Memory
High-bandwidth memory structurally short into 2027–28.
Niche Semiconductors
Epitaxial wafers — the choke point beneath the entire stack.
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.
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.
- 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
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.
- 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
The Commercial Building Comparison
| PRIME COMMERCIAL BUILDING | POWERED SHELL | |
|---|---|---|
| TENANT | Investment-grade corporate | Investment-grade hyperscaler |
| LEASE TERM | 10–15 years | 10–15 years |
| ASSET | Location, steel, glass | Land, substation, secured interconnect |
| SUPPLY RESPONSE | Capital builds more within 2–3 years | Gated by multi-year interconnect queues |
| UNLEVERED IRR | 5–6% | 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.
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.
Most investors do not read planning-guide revision requests.
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.
- 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 ↑
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.
A Thesis Under Stress
- S.1Ran the written drawdown protocol — numbered stages, six hypothesis categories — against the strongest available bear material and the earnings call.
- S.2Audited the bulls’ own client model — and found a 30-point HBM share-erosion assumption their published articles never confront.
- S.3Reverse-engineered the trough price: it implied trough-quarter profits held in perpetuity, with no cycle recovery.
- S.4Verdict: multiple compression and technical flow unwind — not thesis impairment. Long-term agreements with ~10 customers directly rebutted the bear case.
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.
A 50GW AI-buildout scenario implies ~1.5M wpm — the entire epi supply that exists today.
Sizing the Basket
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.
~60% share in leading-edge epi, structurally under-built into the shortage, with the key customer agreement expiring exactly at the 2027 pricing reset.
Most swing capacity, present in all five geographies, government-subsidy de-risked. Best risk-adjusted leg of the three.
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.
The Operating System
Concentration with explicit discipline: high-conviction positions governed by written rules, not judgment in the moment.
What qualifies for research, written down before the idea arrives.
Delta-adjusted single-name limits, checked against every statement.
A standing short book against the factor exposure the longs create.
The written protocol from the Hynix review — run every time, not when convenient.
Index puts held permanently against the concentrated book.
The filter at the door. Ideas that clear it are rare; that is the point.
The Ledger
Ten dated, falsifiable expectations, graded publicly every August. A map you cannot be wrong about is marketing.
- 01END-2027Agent task horizons exceed one work-month; a frontier lab attributes most research-engineering output to its own models
- 022028A G7 or Gulf state treats compute as formal strategic infrastructure — allocation authority, not merely subsidy
- 032027–30A >40% compute-complex drawdown on an efficiency or geopolitical shock — then new highs in utilization and pricing within eight quarters
- 042029Distributed training at frontier scale; the “coherent campus” premium compresses materially
- 052030Matter-limited chokepoints outperform knowledge-limited infrastructure; a celebrated 2026 category is revealed as a false chokepoint
- 062030Verification, assurance, and AI-liability capacity emerges as a recognized asset class
- 072031Robots cross one million units annually; an actuation-chain component repeats the 2024 transformer repricing
- 082031Automated-lab throughput becomes a named institutional category; trial capacity is the acknowledged bottleneck of AI-era medicine
- 092032Robotic construction and machine-run qualification compress a flagship project schedule by more than a third vs 2025
- 102033–35Automated production adds marginal capacity in a former chokepoint; pricing collapses toward energy-plus-materials cost