## Part I: The Constraint
### The Bottleneck Has Moved
For two years, the constraint on AI scaling was chips. NVIDIA couldn't fabricate H100s fast enough. That constraint is easing. TSMC is building fabs. Blackwell is ramping. Within 2-3 years, chip supply will better match demand.
The new constraint is power. Unlike chips, power doesn't respond to capital the same way.
A new transmission line takes 5-10 years. A grid interconnection study takes 18-36 months. These timelines aren't about money. They're about physics, regulation, and sequential dependencies that can't be parallelized. You can't write a check to skip the queue.
In 2025, the four largest hyperscalers will spend roughly $405 billion on capex. In 2026, that rises to $600 billion. Roughly 75% goes to AI infrastructure. It's still not enough. Amazon, Microsoft, Google, and Meta are all reporting the same problem: they can build data centers faster than they can power them.
Total U.S. data center power consumption was roughly 17 GW in 2022. By 2030, projections range from 35-50 GW. That's the equivalent of adding 20-35 nuclear plants in eight years. No one is building 20 nuclear plants.
### The Moat Is Measured in Years
The companies that understood this early, that acquired land with grid access, began interconnection studies years ago, and built utility relationships, now hold positions that can't be replicated on relevant timelines.
Galaxy Digital acquired the Helios campus in December 2022 for $65 million during a crypto crisis. What they actually bought was grid interconnection, three years before AI infrastructure shortage became obvious. A competitor starting today faces the same capital requirements but can't recover those three years. The moat is temporal.
Applied Digital began site development at Ellendale, North Dakota years before hyperscaler demand materialized. By 2025, when CoreWeave needed capacity urgently, Applied Digital had power available on timelines no new entrant could match. The contracts went to the company that started early.
The same pattern repeats across Cipher, IREN, and TeraWulf. Operators who understood grid access was the scarce asset acquired it before scarcity became obvious.
---
## Part II: The Lock-In
### Three Layers of Cascading Obligations
The demand structure has three layers. Each layer's commitment locks in the layer below.
**Layer 1: Enterprise to Hyperscaler.** A Fortune 500 CIO signs a $500M, 5-year Azure commitment for AI workloads. Microsoft now has a contractual obligation to deliver that compute.
**Layer 2: Hyperscaler to Operator.** Microsoft, needing infrastructure to fulfill its enterprise commitments, signs a lease with an operator who can deliver power. The operator now has a contractual obligation to build and operate the facility.
**Layer 3: Operator to Physical.** The operator finalizes grid interconnection, purchases equipment, begins construction. The capital is deployed. The asset is real.
Each layer enables the layer below to secure financing and execute. The enterprise commitment makes the hyperscaler commitment rational. The hyperscaler commitment makes operator financing possible. The operator financing makes the physical asset exist.
### The Distinction That Matters
Not all operators occupy the same position in the stack.
**Colocation (APLD, GLXY, CIFR, WULF):** Build the facility, provide power and cooling, lease space to hyperscalers who bring their own equipment. Capital-light relative to revenue. Lower revenue per MW, but customers bear technology obsolescence risk.
**GPU Cloud (IREN, NBIS):** Own the GPUs, sell compute services. IREN's Microsoft contract isn't a lease. It's a compute services agreement. Revenue per MW is 3-4x higher, but they bear hardware obsolescence risk.
This is why contracted revenue matters more than pipeline. Contracted revenue represents obligations that have cascaded through all three layers. Pipeline represents hope that they might.
---
## Part III: The Mispricing
### The Inversion Problem
The market has inverted risk and reward.
Companies with zero revenue and speculative timelines trade at premium valuations. Companies with billions in contracted revenue trade at discounts because of "execution risk."
Speculative AI plays with no infrastructure, no contracts, and delivery timelines in the 2030s trade at $15-20+ billion market caps. Operators with $57 billion in combined contracted revenue from hyperscalers trade at $70-75 billion combined market cap.
Either the contracts are worthless, or the market is mispricing who benefits from AI infrastructure buildout.
### Category Errors
The market sees what it expects to see.
Galaxy Digital is covered by crypto analysts who track Bitcoin holdings and trading volumes. The Helios data center, 800 MW contracted to CoreWeave with projected revenue exceeding $1 billion annually, gets a paragraph in reports dominated by crypto outlook. When Bitcoin falls 10%, Galaxy falls with it, even though CoreWeave's contract doesn't change.
IREN was Iris Energy, a Bitcoin miner. Now it has a $9.7 billion contract with Microsoft. Cipher still mines Bitcoin but has $9.3 billion in contracts with AWS and Fluidstack. TeraWulf was mining at a coal plant site. Now it has $17 billion in Google-backed contracts and a hyperscaler as a 14% equity holder.
The transformations are real. The categorizations haven't caught up.
---
## Part IV: The Operators
### Core Four (Infrastructure/Colocation)
**Applied Digital (APLD):** $16B contracted revenue
- 400 MW with CoreWeave (~$11B), 200 MW with undisclosed hyperscaler (~$5B)
- North Dakota sites offer structural cooling cost advantages
- Only operator with delivered AI infrastructure at scale (100 MW operational)
- Macquarie committed $5B in perpetual preferred equity
- Primary risk: CoreWeave concentration
**Galaxy Digital (GLXY):** $15B contracted revenue
- Helios campus: 800 MW contracted with CoreWeave
- Most severe category error. Covered as crypto, valued as crypto.
- Texas location (ERCOT) offers faster interconnection
- $1.4B project financing secured against CoreWeave commitment
- Primary risk: 100% CoreWeave concentration, crypto correlation in trading
**Cipher Mining (CIFR):** $9.3B contracted revenue
- Most diversified customer base: AWS (200 MW), Fluidstack/Google (300 MW)
- 3.2 GW total power pipeline across multiple sites
- Only operator with direct AWS relationship
- Primary risk: Multi-site execution complexity
**TeraWulf (WULF):** $17B contracted revenue
- Lake Mariner and Cayuga sites: inherited infrastructure from retired coal plants
- Google took 14% equity stake through Fluidstack relationship
- When a hyperscaler converts from counterparty to co-owner, incentives align
- Primary risk: Fluidstack intermediary dependency (mitigated by Google backing)
### Adjacent Two (GPU Cloud)
**IREN:** $9.7B contracted revenue
- Direct Microsoft relationship, 5-year compute services agreement
- 200 MW at Childress, Texas with liquid cooling at scale
- Highest revenue density per MW among all operators
- Primary risk: Hardware obsolescence, Microsoft concentration
**Nebius (NBIS):** $20.4B contracted revenue
- Largest contracted backlog of any operator in this thesis
- Microsoft: $17.4B. Meta: $3B.
- Two hyperscaler relationships provide diversification
- Geopolitical discount (Yandex heritage) that counterparties don't share
- Primary risk: Hardware obsolescence, execution at scale
---
## Part V: What Breaks It
### The Bear Cases
**Demand disappoints.** AI revenue remains a fraction of AI infrastructure spending. If enterprise adoption slows, hyperscalers have data centers they don't need. Historical parallel: telecom buildout 1998-2001.
**Power supply catches up.** The interconnection queues are long but full of projects. Grid operators are accelerating approvals. If power abundance replaces scarcity, the moat disappears.
**Execution fails.** These are young companies attempting industrial-scale construction for the first time. The contracts include termination provisions for missed timelines. None have built at this scale before.
**Customer concentration destroys value.** APLD and GLXY depend heavily on CoreWeave. WULF and CIFR depend on Fluidstack. IREN depends on Microsoft. If counterparties struggle, contracted revenue evaporates.
**Hardware obsolescence (adjacent positions).** IREN and NBIS own the GPUs. If next-gen chips make current hardware uncompetitive, they have stranded assets. The 5-year contracts provide term protection, but re-contracting risk is real.
### The Rebuttals
**On demand:** The hyperscalers are capacity-constrained now. Every earnings call confirms it. They're signing multi-billion-dollar, 15-year leases with any operator who can deliver. Near-term demand is proven.
**On power supply:** New entrants face the same queue. Money doesn't skip physics. The 2-4 year advantage persists until competitors complete their own interconnection timelines.
**On execution:** This is real and irreducible. Applied Digital has delivered 100 MW on time. Others are still in construction. Watch the delivery milestones.
**On concentration:** IREN's direct Microsoft relationship is cleanest. Cipher has AWS diversification. TeraWulf has Google equity alignment. The risk is structural but varies by operator.
**On obsolescence:** This is why IREN and NBIS are "adjacent" rather than "core." The risk is real. Contracts provide term protection. Size positions accordingly.
### What I Might Be Wrong About
- CoreWeave's durability (affects APLD and GLXY)
- Fluidstack's execution (affects WULF and CIFR)
- Contract enforceability (relying on operator representations)
- Execution at scale (extrapolating from limited track records)
- The demand trajectory (could be at peak AI capex now)
---
## The Numbers
| Operator | Contracted Revenue | Market Cap | Ratio |
|----------|-------------------|------------|-------|
| APLD | $16B | ~$10-11B | ~0.7x |
| GLXY | $15B | ~$8-9B | ~0.6x |
| CIFR | $9.3B | ~$5-6B | ~0.6x |
| WULF | $17B | ~$4-5B | ~0.3x |
| **Core Four** | **$57B** | **~$28-31B** | **~0.5x** |
| IREN | $9.7B | ~$13-14B | ~1.4x |
| NBIS | $20.4B | ~$25-27B | ~1.3x |
| **All Six** | **~$87B** | **~$70-75B** | **~0.8x** |
The core four trade at roughly half their contracted revenue. The adjacent two trade at premiums reflecting higher revenue per MW but also higher obsolescence risk.
---
## The Bottom Line
The constraint is physical. Power takes years to secure. The companies that positioned early own something that can't be replicated on relevant timelines.
The obligations have cascaded. Enterprises committed to hyperscalers. Hyperscalers committed to operators. The contracts are signed. The revenue is locked.
The market is mispricing the result. Speculation trades at premiums. Infrastructure trades at discounts. Either the contracts are worthless, or the market is catching up to who actually benefits.
The thesis is testable. Watch the delivery milestones. Watch the revenue recognition. The information that proves or disproves this will arrive over the next 12-24 months.
Read the full thesis: www.followthewatts.com