The Risks of AI Data Center Investments

What actually goes wrong in AI data center investments, in rough order of how often — leasing, power, schedule, fees, liquidity and obsolescence — and the eligibility rules that decide what you can buy at all.

Most risk sections in offering documents are written by lawyers to be complete. This one is written to be useful: the failure modes that actually occur, in rough order of how often they decide the outcome, and what each one does to your money.

1. Nobody signs a lease

The most common way a private data center deal disappoints is the simplest: the building gets built and the tenant does not materialise at the assumed rate, or at all.

This risk is concentrated in deals that raise money before a lease is signed. The projected yield in such a deal is arithmetic performed on an assumption. If the building leases at 80% of the assumed rate, roughly 20% of the gross revenue disappears, and because operating costs do not fall in proportion, the hit to net income is worse than 20%. If it leases late, you have paid for capital that earned nothing in the meantime.

Context from the market: 74.3% of North American capacity under construction was already preleased (CBRE, H1 2025). In a market where three quarters of the pipeline has a tenant before completion, a project that reaches construction without one is unusual, and the reason is worth understanding before you fund it. See how data center leases work.

2. The power does not arrive on time

Covered at length in why power is the constraint, and second on this list only because leasing is first. A non-binding utility proposal is not contracted power, and CBRE reports power constraints pushing construction timelines "to 2027 and beyond" (CBRE, Global 2025). Every month of delay is a month without rent against capital that is already spent — and if the project has debt, interest accrues through the delay.

3. Construction and equipment slip

Data center fit-out depends on long-lead items: transformers, switchgear, chillers, and for AI densities the liquid cooling distribution as well. A single-source supplier, particularly an overseas one, concentrates this risk and adds tariff, certification and shipping exposure. A benchmark timeline quoted by a vendor is the vendor's best case.

Ask who bears overrun. If the sponsor has a fixed-price contract with a creditworthy general contractor, that is meaningful. If overruns come out of project cash, they come out of your return.

4. The yield you were shown is not the yield you receive

Not a risk exactly — more an accounting gap that reliably surprises people. Check, in writing, whether the headline figure is before or after:

• asset or property management fees, usually an annual percentage;
• sponsor promote or carried interest, a share of profit above a threshold;
• acquisition, sourcing or financing fees, often paid at closing and earned regardless of how the project performs;
• sponsor units that share in income alongside yours.

A projection described as being "before possible management fees" is explicitly not telling you your return. The phrase is honest; it is just easy to read past.

5. You cannot get out

Private data center investments are illiquid in a way that is hard to appreciate until you want your money. There is normally no market, transfer is restricted by the operating agreement as well as securities law, and the exit depends on an event — a sale, a refinancing, a listing — that the sponsor controls and has not promised.

Be sceptical of exits described as "a planned future offering" or "an anticipated liquidity event." A plan is not a commitment, and intervals of five to ten years are normal. For comparison, the SEC's investor bulletin on non-traded REITs notes that even in that more regulated wrapper, a listing or liquidation may not occur until more than ten years after the investment.

Treat money in a private deal as money you will not see for years, and size the position accordingly.

6. Concentration

One building, in one county, with one power supply, and often one tenant. There is no diversification inside the investment. Everything that happens to that project happens to all of your money in it. This is the structural difference from a REIT, which spreads the same exposure across hundreds of facilities.

7. Technological and functional obsolescence

A facility is engineered for a load profile. If densities keep rising, a building designed for today's racks may need re-engineering to host the next generation; if the cooling approach the industry settles on differs from the one installed, retrofit costs fall on the owner. The shell and the electrical service are long-lived; the mechanical fit-out is not. Models that assume fifteen years of income with no capital replacement reserve are understating cost.

8. Demand disappointing at the industry level

The tail risk that would hurt everything at once: AI capital expenditure slows, hyperscalers pause, and a market with a large forward pipeline turns from shortage to surplus. Note the DOE/LBNL projection range is itself wide — 325 to 580 TWh by 2028, 6.7% to 12% of US electricity (DOE, 2024). The best available forecast spans nearly a factor of two. Position sizes should reflect that nobody knows.

9. Sponsor risk

Someone has to execute. Ask what else they have built, whether it was this asset class, whether it was this size, and whether an investor in their last deal would take their call. A first-time sponsor in a capital-intensive, power-constrained, technically demanding asset class is taking on a hard job.

10. Tax and structure

Pass-through structures generate tax reporting that arrives when it arrives, and can create filing obligations in the project's state. REIT dividends are generally taxed as ordinary income. None of this is a reason to avoid an investment, and all of it is a reason to ask your own tax adviser before rather than after.

Who can invest

Eligibility is not a risk, but it determines which of the above you are exposed to, and it surprises people who have already decided to invest.

Regulation A offerings are open to non-accredited investors. In a Tier 2 offering a non-accredited natural person may invest no more than the greater of 10% of annual income or 10% of net worth, and Tier 2 is capped at $75 million raised in any 12-month period (Tier 1: $20 million) (SEC, Regulation A; SEC, Guidance for Issuers).

Private placements under Rule 506(b) may be sold to unlimited accredited investors and no more than 35 non-accredited investors in any 90-calendar-day period (SEC) — in practice, accredited only.

A natural person is accredited by, among other routes, net worth over $1 million excluding the primary residence (SEC), or income over $200,000 individually or $300,000 jointly in each of the two most recent years with a reasonable expectation of the same this year, or holding a Series 7, 65 or 82 licence in good standing (SEC).

Many AI data center equity offerings require accredited status and set minimums in five figures. If that is not you, the honest answer is that most of this category is closed, and what you can actually buy is a shorter list than the advertising suggests.