I've had a version of the same conversation with family office principals about fifteen times in the past two years. They've heard about the AI infrastructure buildout. They understand that data centers are the physical substrate of the AI economy. They want exposure. They have capital. And they've arrived at the conversation with an underwriting model that doesn't quite fit what they're actually being asked to buy.
This isn't a criticism — it's a structural problem. Family offices are sophisticated allocators, but their frameworks were built for asset classes with longer track records: real estate, private equity, direct lending, operating businesses. AI infrastructure development at the powered-land stage is none of those things exactly. It borrows elements from all of them, which means assuming it works like any one of them leads to predictable blind spots.
Three mistakes come up consistently enough that they're worth naming clearly.
Mistake 01: Treating It Like Real Estate
The most common frame family offices bring to AI infrastructure is real estate — which makes intuitive sense. The asset is land. There's power. There might be a structure. The development involves permitting, construction, and ultimately a lease or sale to an end user. The vocabulary overlaps.
But the risk profile doesn't. Commercial real estate — even development-stage real estate — has a relatively stable demand picture and a relatively legible path from raw land to stabilized asset. Zoning, construction costs, and absorption rates are uncertain, but they're uncertain within a range that decades of comparable transactions have calibrated.
Powered land for AI infrastructure doesn't have that comparable history. The technical requirements are moving fast — the cooling and power density specs for a GPU-dense AI cluster in 2026 are fundamentally different from what a hyperscale cloud tenant needed in 2019. The jurisdictional landscape is shifting, with moratoriums and regulatory changes occurring faster than traditional real estate underwriting cycles can track. And the tenant universe is narrower and more sophisticated than most real estate developers are used to — hyperscalers and large AI operators negotiate leases from a position of significant information advantage over first-time data center landlords.
The implication: underwriting powered land like stabilized real estate misprices the development risk. Family offices that apply real estate cap rates to early-stage AI infrastructure sites are buying a fundamentally different risk profile than they think they are — often at valuations that only make sense if everything goes right.
Mistake 02: Underestimating the Timeline and Capital Intensity
The second mistake is a calibration error on development timeline and the capital required at each stage. Family offices that have done real estate development are accustomed to construction timelines of 12–36 months from permitted land to occupancy. AI infrastructure development — at the scale required to attract institutional tenants — doesn't fit that timeline.
The sequencing from raw powered land to a facility capable of supporting serious compute density typically looks something like this:
- 01Utility interconnection study and queue position confirmation — 6 to 18 months, independent of anything else you do
- 02Environmental review, permitting, and entitlements — 12 to 36 months in most jurisdictions, longer in contested ones
- 03Infrastructure buildout (substation, fiber, cooling systems, civil work) — 18 to 30 months after permits clear
- 04Tenant lease negotiation and fit-out — 6 to 18 months, with hyperscale tenants running their own procurement process on their own timeline
These stages don't run fully in parallel. The total timeline from acquisition of raw powered land to first tenant occupancy is realistically four to seven years in most markets, with significant capital calls at each stage and no revenue until you're much further along than most family office models assume.
The capital intensity is also higher than it looks from the outside. Power infrastructure upgrades, fiber pulls, cooling system installation, and substation builds can each run into eight figures before a single server rack is installed. Family offices that underwrite AI infrastructure with a real estate development capital model — land acquisition plus vertical construction — routinely discover they've undercapitalized for the infrastructure layer.
Mistake 03: Assuming Tenant Demand Is Self-Evident
The third mistake is the most seductive, because it's grounded in something true. AI compute demand is real, it's large, and it's growing. The hyperscalers are spending hundreds of billions on infrastructure. Everyone needs more data centers. Therefore, if you build one, someone will come.
This reasoning is correct at the macro level and unreliable at the asset level. Hyperscalers and large AI operators are not buying undifferentiated data center capacity from first-time developers. They have technical specifications — power density requirements, cooling approaches, fiber diversity standards, security protocols — that most early-stage sites can't meet without significant additional investment. They have preferred developer relationships built over years of repeated transactions. And they have the leverage to dictate lease terms, construction specifications, and development timelines to a degree that often surprises sponsors who've never negotiated with them before.
Smaller AI operators and colocation customers are a more accessible tenant universe, but they don't provide the anchor lease that makes large-scale development economics work. The math on a 50MW facility is very different depending on whether you're leasing to a hyperscaler at 10-year triple-net terms or to a mix of colocation tenants at shorter durations with higher operational overhead.
What sophisticated AI infrastructure developers do — and what most family offices don't account for in their initial underwriting — is pre-negotiate tenant interest before breaking ground, or at least before committing to the full capital stack. Build-to-suit structures, letters of intent, and anchor pre-leases reduce the tenant demand risk substantially. Speculative development into an undifferentiated market is a fundamentally higher-risk posture than it appears from the macro demand numbers.
What Family Offices Should Do Instead
None of this means family offices shouldn't be in AI infrastructure. The opportunity is real and the allocation makes sense for a multi-generational capital base that can tolerate development timelines. But the approach needs to match the actual risk profile of the asset class.
The family offices that are doing this well share a few common practices. First, they're accessing pre-diligenced deal flow rather than evaluating raw sites themselves — they rely on advisors who have already done the technical assessment, jurisdictional analysis, and capital sequencing work before the opportunity reaches them. Second, they're structuring in at the right stage for their risk tolerance — some are comfortable with early-stage development risk; others are better suited to entering at a later milestone when the permitting is clear and the tenant conversation is advanced. Third, they're building relationships with developers who have done this before, rather than being the sophisticated capital behind a first-time sponsor.
The asset class rewards family offices who approach it with the same rigor they'd bring to a complex direct investment in an operating business — deep diligence, stage-appropriate entry, and alignment with sponsors who understand the specific technical and political requirements of the market. It penalizes those who treat it as a real estate trade with a better narrative.
I work with family offices on exactly this problem: getting the framework right before the capital is committed, and making sure the sites they're evaluating have been assessed against the actual requirements of institutional AI infrastructure development — not the assumptions that come from applying a different asset class's lens to a new one.