Year 6 Survival Model: AI Data Center Investment Risk and Why Lease Duration Breaks the Math
- Tony Grayson
- 1 hour ago
- 14 min read
A Nuclear Submarine Commander's Framework for Underwriting AI Infrastructure
By Tony Grayson, President & GM of Northstar Enterprise + Defense | Former U.S. Navy Nuclear Submarine Commander | Stockdale Award Recipient | Veterans Chair, Infrastructure Masons
Published: January 23, 2026
TL;DR
Today, I read an analyst note predicting a 3x stock surge for a company solely because they're about to sign a lease. That euphoria is a warning sign. A 5-year contract on a gigawatt site is a short-duration revenue instrument sitting on long-duration infrastructure. The gap between contract life and asset life is where capital gets mispriced. Unless you have a credible re-tenanting thesis—not a spreadsheet assumption—you aren't underwriting. You're gambling.
In 30 Seconds
The Problem: Lease terms have extended to 10-15 years, but revenue is tied to hardware cycles that obsolete every 2-3 years
The Scale: Q3 2025 saw 7.4 GW of data center capacity leased—more than all of 2024 combined. $61 billion flowed into data center deals in 2025.
The Mismatch: A 1 GW campus has 30-year civil works, but GPU generations that refresh every 18-24 months.
The Precedent: Global Crossing built fiber networks on "demand will grow forever"—they were right about demand, wrong about asset value
The Test: What is this asset worth the day the contract expires if renewal doesn't happen?
The Rule: If your Year 6 thesis is "they'll renew," you don't have a thesis
What Is the Year 6 Survival Model for AI Data Center Investment Risk
The Year 6 Survival Model is a capital allocation framework that assumes non-renewal as the base case for AI data center leases. It requires explicit modeling of retrofit costs, dark period cash burn, and residual value to non-strategic buyers.
The model emerged from a simple observation: most AI data center investments are underwritten on the assumption that tenants will renew. But when GPU power requirements double every two years and cooling technology shifts from air to liquid, the facility you built in Year 1 may be physically incapable of serving the tenant's needs in Year 6. Renewal isn't a negotiation—it's an exit.
"I call this the Year 6 Survival Model: the practice of underwriting AI infrastructure as if renewal won't happen, because often it can't."
— Tony Grayson
Key Concepts for AI Data Center Investment Risk
Year 6 Survival Model: A capital allocation framework that assumes non-renewal as the base case for AI data center leases and requires explicit modeling of retrofit costs, dark period cash burn, and residual value to non-strategic buyers.
Encumbered Power: Grid interconnect capacity that cannot be easily monetized because the facility requires a nine-figure renovation to support next-generation density.
Kill Cost: The specific cost to strip a facility's interior MEP (mechanical, electrical, plumbing) to the bare walls to unlock power value for re-tenanting.
Dark Period: The 12-18-month period during which a facility generates zero revenue while undergoing retrofit for next-generation workloads.
Stranded Asset: Infrastructure that becomes economically unviable before the end of its physical life because technology requirements advance faster than facilities can be retrofitted.
Commander's Intent
Purpose: Equip capital allocators, operators, and boards with a framework for underwriting AI infrastructure that survives contract expiration. This covers the AI Data Center Investment Risk.
End State: Readers understand why contract duration has become mispriced, can identify the retrofit cost trap, and know what questions to ask before committing capital to gigawatt-scale projects.
Objectives:
1. Explain why the traditional "sign a hyperscaler, finance it like a utility" model is dead
2. Identify the specific technical mismatches that create stranded asset risk
3. Provide historical precedents that validate the pattern
4. Deliver a concrete framework—the Year 6 Survival Model—for honest underwriting
A Note on Consensus: This analysis directly contradicts the prevailing view from JLL, CBRE, and most sell-side analysts who project a $3 trillion "infrastructure supercycle" without addressing contract duration risk. The bulls are pricing these assets as if renewal is guaranteed. I'm arguing that's the central question they're not asking.
Why AI Data Center Financing Models Are Broken
For the last decade, the market treated hyperscaler contracts like annuities. Sign a big name. Lock a term. Finance it like a utility. Assume renewal. Print money.
I think that mental model is dead.
We are seeing companies sign short-term contracts and claim victory. But a 5-year contract on a gigawatt site is a short-duration revenue instrument sitting on long-duration infrastructure.
The gap between contract life and asset life is where capital gets mispriced. Unless you have a credible re-tenanting plan, you aren't underwriting—you're gambling.
"During my time managing $1.3 billion in infrastructure at Oracle, I watched lease committees approve 15-year commitments without ever asking what the asset was worth in Year 6. The question wasn't on the diligence checklist. It should have been."
— Tony Grayson
Consider the numbers: Oracle's rent-payment commitments to data center landlords typically span 15-19 years. But the cloud-computing services sold to clients operate on much shorter cycles. As JLL noted, "if you're only getting a third of that guaranteed by a credit entity, it's really viewed as a five-year lease."
I'm not alone in this concern. Michael Johnston, a partner at Menlo Equities focused on data center investments, put it bluntly: "It's fair to say that no one is sure about the renewal probabilities in West Texas 15 years from now. There isn't a data point."
Key Takeaway: The structural mismatch isn't a risk factor to monitor—it's the central underwriting question. A 15-year lease with a 5-year backstop is a 5-year lease.
The Stranded Asset Trap: 30-Year Infrastructure vs. 3-Year Tech Cycles
A 1 GW campus is the definition of heavy, long-cycle infrastructure. The interconnects and civil works are designed for at least a thirty-year horizon.
But the revenue stream is tied to hardware cycles that obsolete themselves every 2-3 years. That platform replacement cadence is accelerating. As a result, density assumptions change faster than mechanical and electrical systems can be rebuilt.
A five-year contract may appear secure, yet the hardware within the facility is likely to see at least one—and possibly two—complete refreshes before the lease concludes.
The NVIDIA GB200 NVL72 illustrates the problem. This rack-scale system—weighing 1.36 metric tons (3,000 lbs)—uses more than 2 miles of copper cabling to mesh 72 Blackwell GPUs together, with hard distance limits that constrain facility design. A facility designed for traditional hot/cold aisles may not physically support these architectures.
Meanwhile, NVIDIA's GPU roadmap shows power consumption reaching 1,400W per GPU with Blackwell Ultra in 2025, scaling to 1,800W with Vera Rubin in 2026. Feynman GPUs arrive in 2028 with specifications not yet disclosed—but the trajectory is clear. A facility designed for today's thermal envelope becomes obsolete before the lease expires.
This isn't theoretical. In December 2025, Fermi Inc.—the data center power developer co-founded by former Texas Governor Rick Perry proposing to build 11 GW of capacity—disclosed that an investment-grade tenant had terminated a $150 million agreement tied to the project. The non-renewal risk is already materializing.
Historical Precedents: Global Crossing and the Office Tower Collapse
The belief that creditworthy counterparties guarantee returns is a dangerous way to do business.
The Fiber Optic Collapse
In the late 1990s, Global Crossing built massive fiber networks on a simple thesis: bandwidth demand will grow forever. They weren't wrong about the demand—they were wrong about the asset's value.
Wharton's analysis captured the dynamic: "People were building fibers believing that was the right forecast. Plus, they were dumping it wherever it was cheap to dump it—not where it would actually get traffic."
When Global Crossing filed for bankruptcy in January 2002, it reported $22.4 billion in assets and $12.4 billion in debt. The company never had a profitable year. By 2004, only about one-tenth of installed fiber was actually "lit."
The telecom crash wasn't caused by wrong demand forecasts—it was caused by asset value collapse when competition made the infrastructure commodity.
The Office Tower Implosion
We saw it again with office towers during COVID. Long leases to Fortune 500 tenants were considered bond-like until remote work structurally impaired the underlying demand driver.
Today, office vacancy rates have hit a record 19.6% in Q1 2025—the highest on record. CMBS office delinquency rates reached 11.8%. The leases were still there. The demand driver had shifted.
The lesson is consistent: Contract duration becomes fiction when the asset's utility shifts.
I believe AI data centers are just the next iteration of this pattern.
Key Takeaway: Global Crossing's fiber was fine. The business case collapsed. Office towers still stand. The tenants left. AI data centers face the same structural risk—not from demand disappearing, but from technology dislocation, making the physical plant unfit for purpose.
Why Traditional Data Centers Can't Handle AI Workloads
In traditional leases, if you had power, cooling, and a roof, it was good enough.
AI infrastructure is different. Customers aren't renting square footage—they are renting the ability to run specific chipsets at specific temperatures.
The Thermal Envelope Problem
A facility designed for air cooling cannot support direct-to-chip liquid cooling without a large retrofit. Air cooling fails above 41.3kW per rack while liquid cooling handles 200kW+. Retrofit costs run $2-3 million per megawatt.
Data Center Dynamics reports that retrofitting can reduce capital expenditure by 20-40% compared to new builds, but that still means 60-80% of new-build costs. And those retrofits take 12-18 months during which the facility generates zero revenue.
The Power Architecture Problem
We are hitting the physical limits of amperage on standard copper cabling. Moving from 480V AC to 800V DC isn't a plug-and-play upgrade—it is a topology break. The NVL72 GB300 requires 140kW per rack—versus 10kW average for traditional CPU-based cloud servers just a few years ago.
The Fabric Topology Problem
New architectures like the Blackwell NVL72 have hard distance limits for copper cabling. The entire rack uses 5,000+ copper cables precisely cut and measured. A facility designed for long hot/cold aisles might not be able to support these requirements without complete reconfiguration.
If your facility falls short on any dimension by Year 6, renewal isn't negotiable. It's an exit.
AI Data Center Investment: You're Buying Equipment, Not Real Estate
Based on what I'm reading in the news, I feel like some investors are missing this shift.
In a warehouse, the shell is the value. In an AI data center, the cooling loops, PDUs, switchgear, and backup gen-sets represent the majority of the capital stack.
To call this "real estate" is an error. You are leasing rapidly depreciating industrial machinery inside a concrete box. If that machinery becomes unfit for purpose in five years, you haven't lost "some value." You've lost your basis.
Consider the capital breakdown: JLL's own numbers project $1.2 trillion in "real estate asset value" but up to $2 trillion in "tenant IT fit-out." Nearly two-thirds of the capital in this "infrastructure supercycle" isn't infrastructure at all—it's hardware with a 3-year shelf life.
Key Takeaway: We are financing servers as if they were skyscrapers. When your financing terms are 5x longer than your technology's useful life, you don't have a real estate asset. You have a depreciating equipment loan dressed up as infrastructure.
The Encumbered Power Problem: Why Grid Access Isn't Enough
Bulls argue that power scarcity protects the asset. "Even if the tenant leaves, I have the grid connection."
This logic ignores the cost of access. If your gigawatt is tied to a facility that requires a nine-figure renovation to be usable for next-gen density, that interconnect isn't a clean option. It is encumbered power.
"The bulls say power scarcity protects the asset. I say encumbered power is worse than no power—you're paying to maintain an option you can't exercise without nine figures of retrofit capital."
— Tony Grayson
In Year 6, you aren't competing against other old buildings. You are competing against greenfield sites designed for modern density. Facilities like those being purpose-built in 2026 around AI workloads—with liquid cooling, higher-density racks, new UPS systems—from day one.
If you own a brownfield redevelopment site, you must price it accordingly. That seems like a bet with very bad odds.
The numbers bear this out: Construction costs are rising 7% annually. Every year you wait to retrofit, the math gets worse. And during that 12-18 month retrofit period, your asset produces zero revenue while competitors with greenfield AI-native facilities capture the market.
How to Apply the Year 6 Survival Model: AI Data Center Underwriting Framework
The alternative to perpetuity thinking isn't paralysis. It's being honest with ourselves.
For the capital partner, the diligence standard must change. I think boards should ask for a "Year-X Survival Model" that hard-codes the cost of relevance:
Retrofit CapEx Reserve: Explicitly carry $150-200 per kW in retrofit capex to maintain fitness for next-generation workloads
Kill Cost Analysis: Price the cost to strip the interior MEP to unlock the power value for re-tenanting
Dark Period Survival: Model whether the capital stack can survive the cash burn of 12-18 months the facility sits dark during reconstruction
Competitive Position: Identify greenfield competitors entering your market during the contract period and model their cost advantage
If the deal breaks under those scenarios, you aren't underwriting an asset. You are underwriting a wish.
Takeaway: The Year 6 Survival Model forces honest answers. Can this facility support next-gen workloads without nine-figure retrofits? Can the capital stack survive the dark period? What's the asset worth to a buyer who didn't build it?
The Bottom Line for AI Data Center Investors
We used to assume that infrastructure demand was predictable and technology cycles were slower than the amortization schedule.
I don't think that is true anymore.
The question for any committee evaluating a gigawatt-scale project—or really any project—is simple:
What is this asset worth the day the contract expires if renewal doesn't happen?
If the answer relies on a spreadsheet assumption rather than a specific resale thesis, the deal isn't conservative.
It's just speculation dressed up as analysis.
If your Year 6 thesis is "they'll renew," you don't have a thesis.
Want to discuss how your organization can apply the Year 6 Survival Model? I consult with capital allocators, boards, and infrastructure teams on AI data center underwriting. Connect with me on LinkedIn or reach out through Northstar Enterprise + Defense.
Frequently Asked Questions
What is the Year 6 Survival Model?
The Year 6 Survival Model is a framework Tony Grayson developed for underwriting AI data center investments. It requires capital allocators to explicitly model what happens when the initial lease expires: retrofit costs ($150-200/kW), the "kill cost" to strip MEP for re-tenanting, and whether the capital stack survives 12-18 months of zero revenue during reconstruction. The model forces honest answers about asset value beyond contract duration.
Why is AI data center lease duration a risk?
AI data center lease duration creates risk because 5-year contracts sit on 30-year infrastructure. The revenue stream is tied to hardware cycles (GPUs, cooling systems) that obsolete every 2-3 years, while civil works are designed for three decades. This mismatch means facilities may become unfit for next-generation workloads before the lease expires, turning renewal into an exit rather than a negotiation.
Is the AI data center boom a bubble?
The comparison to a bubble is incomplete. Demand for AI compute is real and growing—Q3 2025 saw 7.4 GW leased in a single quarter, more than all of 2024. The risk isn't that demand disappears; it's that the physical infrastructure becomes obsolete faster than the financing assumes. Global Crossing's fiber wasn't a demand bubble—it was an asset value collapse. AI data centers face the same pattern: demand may persist while the assets built to serve it become stranded.
What is stranded asset risk in data centers?
Stranded asset risk in data centers refers to infrastructure that becomes economically unviable before the end of its physical life. This occurs when technology requirements (power density, cooling capacity, network topology) advance faster than facilities can be retrofitted. A data center designed for 15kW racks cannot serve tenants requiring 120kW racks without nine-figure renovation—making the existing infrastructure "stranded" despite having years of physical life remaining.
How long do AI data center leases last?
AI data center lease terms vary significantly. JLL reports that headline lease terms have extended to 10-15 years, but the credit-backed portion is often only 5 years. Oracle's commitments span 15-19 years on paper, but analysts note "if you're only getting a third of that guaranteed by a credit entity, it's really viewed as a five-year lease." This gap between headline term and credit-backed term is the central underwriting risk Tony Grayson's Year 6 Survival Model addresses.
What happened to Global Crossing and why is it relevant?
Global Crossing built massive fiber networks in the late 1990s, reaching a $47 billion valuation. They filed for bankruptcy in January 2002 with $12.4 billion in debt, never having been profitable. The company was right about bandwidth demand growing—they were wrong about their asset's value. By 2004, only 10% of installed fiber was "lit." Tony Grayson sees AI data centers following the same pattern: demand may grow, but asset value can still collapse.
Why can't traditional data centers handle AI workloads?
Traditional data centers were designed for 10-15kW per rack with air cooling. AI workloads require 50-140kW per rack with liquid cooling. The NVIDIA GB200 NVL72 consumes 120kW and requires direct-to-chip liquid cooling, 5,000+ precisely cut copper cables, and specific distance constraints that traditional hot/cold aisle designs cannot accommodate.
How much does it cost to retrofit a data center for AI?
Retrofitting a data center for AI workloads costs $2-3 million per megawatt for liquid cooling installations. Data Center Dynamics reports retrofits can reduce CapEx by 20-40% compared to new builds—meaning they still cost 60-80% of greenfield construction. The retrofit also requires 12-18 months of downtime during which the facility generates zero revenue.
What is encumbered power in data center investing?
Encumbered power refers to grid interconnect capacity that cannot be easily monetized because the facility requires nine-figure renovation to support next-generation density. Bulls argue power scarcity protects assets, but Tony Grayson counters that if your gigawatt requires $100M+ to become usable for modern workloads, that interconnect isn't a clean option—it's encumbered by the retrofit cost and 12-18 month reconstruction timeline.
How fast are GPU power requirements increasing?
NVIDIA's GPU roadmap shows power consumption increasing rapidly: Blackwell Ultra (GB300) reached approximately 1,400W per GPU in 2025, and Vera Rubin chips are projected at 1,800W in 2026. This acceleration means facilities designed for today's thermal envelope become obsolete before a 5-year lease expires, creating the Year 6 problem Tony Grayson identifies.
What questions should boards ask about AI data center investments?
Boards should ask: (1) What is the explicit retrofit CapEx reserve to maintain fitness for next-gen workloads? (2) What is the "kill cost" to strip interior MEP and re-tenant? (3) Can the capital stack survive 12-18 months of zero revenue during reconstruction? (4) What is the asset worth to a buyer who didn't build it? (5) What greenfield competitors will enter this market before contract expiration?
Who is Tony Grayson?
Tony Grayson is President and General Manager of Northstar Enterprise + Defense, which designs and manufactures modular, AI-optimized data centers. He previously served as SVP of Physical Infrastructure at Oracle (managing a $1.3B budget), with prior executive roles at AWS and Meta. Tony is a former U.S. Navy nuclear submarine commander who served 21 years, including commanding USS Providence (SSN-719), and received the 2015 Vice Admiral James Bond Stockdale Award for Inspirational Leadership.
How does the Year 6 Survival Model differ from standard data center underwriting?
Standard data center underwriting assumes contract renewal and prices assets on steady-state cash flows. The Year 6 Survival Model assumes non-renewal as the base case and requires explicit modeling of: retrofit costs to achieve next-gen capability, dark period cash burn, competitive position versus greenfield entrants, and residual value to a non-strategic buyer. If the deal breaks under these scenarios, it's speculation, not underwriting.
Sources
1. CoStar - "Oracle is on the hook for $248 billion in data center leases" (January 2026)
2. JLL - "Data Center Market Defies Early 2025 Turbulence" (October 2025)
3. Bloomberg - "Global AI Data Center Dominance Shifts Away From Big Tech" (December 2025)
4. CNBC - "Data center deals hit record $61 billion in 2025" (December 2025)
5. Wikipedia - Global Crossing
6. Wharton - "Factors Behind Global Crossing's Failure" (February 2002)
7. NVIDIA Developer Blog - "NVIDIA GB200 NVL72 Designs" (November 2024)
8. Tom's Hardware - "The data center cooling state of play (2025)" (December 2025)
9. SitusAMC - "Data Centers in an AI-Driven Era: Key Trends Reshaping Valuation" (2025)
10. Data Center Dynamics - "Retrofitting liquid cooling for AI data centers" (September 2025)
Related Reading from The Control Room
About the Author
Tony Grayson is President and General Manager of Northstar Enterprise + Defense, which designs and manufactures modular, AI-optimized data centers using proprietary fiber-reinforced polymer composite technology.
Tony served 21 years in the U.S. Navy's nuclear submarine force, finishing as Commanding Officer of USS Providence (SSN-719). He received the 2015 Vice Admiral James Bond Stockdale Award for Inspirational Leadership and is a Naval Academy graduate with degrees in Control Systems Engineering and Engineering Management.
His career is defined by building at scale: he led global infrastructure strategy as a Senior Vice President for AWS, Meta, and Oracle (managing a $1.3B budget and 1,000+ person team) before taking over a failing modular data center company, building it to $200M+ in contracts, and selling it to Northstar. He serves on advisory boards for TerraPower and Holtec International in the nuclear energy sector.
Connect: LinkedIn | The Control Room Blog
Comments