The Insurance Gatekeepers

Data centers need more than power and chips. They need someone willing to insure them.

Image credit: Photo by İsmail Enes Ayhan on Unsplash

When Meta needed $27.3 billion to build its Hyperion data center in Louisiana, it did not write a check or take out a loan. It built a different vehicle called Beignet Investor, transferred 80 percent of the equity to Blue Owl Capital, split the lease into a four-year renewable term instead of one twenty-year lease, and attached a $28 billion "residual value guarantee" put into the small footnotes of its annual report.[1] [2] The framework existed for one reason: to convince rating agencies, insurers, and bond investors that a facility the size of a small city was a manageable risk. The fact that a company with Meta's cash reserves needed this much financial engineering to build one data center reveals another constraint on the AI boom. It is not only GPUs. It is not only power. It is also whether insurers and investors will agree to take on the risk.

The Physical Risk No One Accounted For A single data center campus can now cost $10 billion to $20 billion to build, double that once servers are set up.[3] Swiss Re Institute estimates that global insurance premiums tied to data centers will multiply, from $10.6 billion to $24.2 billion by 2030.[4] Tom Harper, the data center practice leader at insurance broker Gallagher, told CNBC that putting $10 to $20 billion into a single location "creates capacity issues in the marketplace" and that it has become "nearly impossible to reasonably insure" a $20 billion property through regular means.[5] [6] It is a capacity problem, not a pricing problem. Insurers can price catastrophe risk, but no single carrier wants that much risk concentrated in one place, especially when many of these centers sit in hurricane- or wildfire-prone regions.[7] [8]

This is important because insurance has historically been the precondition for large infrastructure to be built, not some random formality. Nuclear power, maritime trade, and commercial aviation all scaled only after insurers built the actuarial tools to price them.[9] AI data centers have no comparable actuarial history for projects of this scale and novelty. Moody's reporting shows the gap at the facility level: Meta's $30 billion Hyperion campus in Louisiana carries only about $4 billion in coverage, leaving roughly $26 billion uninsured at a single site.[10] Lenders are increasingly demanding insurance guarantees before releasing capital they otherwise would have provided.

Why Lenders Now Answer to Underwriters

Once insurance becomes scarce, it stops being a cost line and starts gating financing outright. This is not hypothetical. In March 2026, KKR and Blackstone turned down specific data center debt deals because the insurance on offer was not enough. The coverage available only protected against maximum foreseeable loss, not full replacement value, and neither firm was willing to carry the uninsured residual risk.[11] Lenders financing new AI campuses are now requiring large insurance limits before committing any capital. This reverses the normal order of project finance: instead of insurance following the money, the money now waits for insurance. Whoever controls insurance capacity, a small set of global reinsurers and underwriters, has power over which AI infrastructure gets built and where.

The Accounting Trick That Made the Risk Invisible

The insurance blockage has collided with a similar phenomenon: the underlying debt is disclosed in annual reports, but structured in ways that keep it off the balance sheet lines insurers and rating agencies typically scan first. Five major tech companies, Microsoft, Google, Amazon, Meta, and Oracle, have used SPVs, finance leases, residual value guarantees, and credit-wrap derivatives. The effect is that an estimated $2.3 trillion of data center-related obligations sits outside the balance sheet, roughly 1.7 times their combined reported liabilities.[12]Microsoft's finance lease liabilities more than doubled from $27.1 billion to $66.6 billion in two years without appearing on the "debt" line of its balance sheet. [13] Moody's has warned that these structures mean "disclosures may not show the full picture," and that hyperscaler lease commitments across the industry doubled to $969 billion last year, with more than two-thirds still absent from real balance sheets.[14] [15]

This connects directly back to the insurance problem. Underwriters are being asked to price risks that are already difficult to model because the physical assets are new, highly concentrated, and expensive. When the full financial obligation is also harder to see, insurers have less information about the total exposure attached to a project. That opacity makes the risk harder to price and further reduces the amount of insurance capacity available. The accounting structure therefore does not sit separately from the insurance constraint. It makes the insurance constraint worse.

Who Gains, Who Loses Private credit firms like Blue Owl, Pimco, Apollo, and BlackRock gain enormous new asset classes, causing huge returns on paper.[16] [17] Rating agencies gain more influence, since a single investment-grade opinion from S&P or Moody's can unlock tens of billions in financing that otherwise would require a sign off from an insurer.[18] Reinsurers and specialty underwriters gain leverage because their scarce capacity is now a precondition before construction. The losses, if they come, will land on people who never chose this exposure and won't see it coming: public market investors and policyholders whose capital sits inside pension funds and insurance reserves, now concentrated in AI infrastructure risk that has never been properly modeled. [19]

The Counterargument, and Why It Falls Short Financial experts argue this is simply project finance doing what it has always done, isolating asset-specific risk so a single failure doesn't sink the parent company, exactly as it did for pipelines and power plants.[20] That argument holds when the underlying asset has decades of actuarial history. AI data centers lack exactly that. GPU hardware loses its value far faster than the twenty- to thirty-year leases financing it, creating a mismatch between how long the debt lasts and how long the asset keeps its worth. Traditional finance also assumes transparent, consolidated disclosure. Here, the structures have the effect of keeping obligations off the books rating agencies and insurers rely on. [21]

What it Reveals The AI boom has been narrated as a story about computers, energy, and geopolitics. It is also becoming a story about who is willing to hold the most risk on trillion-dollar bets nobody has actuarial data for. Insurers and reinsurers are becoming one of the rate-limiting institutions for how fast hyperscalers can build, because no amount of capital or political will can force an underwriter to accept a risk it cannot price. Watch three numbers over the next year: whether specialty capacity from programs like Aon's DCLP and FM's Intellium grows faster than hyperscaler capex, whether Moody's next gap update shows the shortfall narrowing or widening, and how credit rating agencies treat off-balance-sheet lease commitments once renewal season forces insurers to reprice accumulation risk. [22] If capacity keeps losing that race, the AI expansion won't stall because of a chip shortage or a policy fight. It will stall because someone, somewhere, decided the risk was uninsurable.

---

Notes

[1] Meta Platforms, Inc., "Annual Report 2025, Form 10-K," January 29, 2026. https://www.sec.gov/Archives/edgar/data/1326801/000162828026025534/meta-12312025x10kars.htm

[2] "New Capital Strategies in the AI Era: Meta Secures $27.3 Billion for Data Center Financing Without Adding Balance Sheet Debt," Moomoo News, October 29, 2025. https://www.moomoo.com/news/post/60508112/new-capital-strategies-in-the-ai-era-meta-secures-27

[3] Anchen, J., & Finucane, J. (2026, March 27). Sigma insights 07/2026: Insuring AI: Data centre value accumulation risks. Swiss Re Institute. https://www.swissre.com/institute/research/sigma-research/sigma-insights-07-2026-insuring-ai-data-centre-risks.html

[4] Recamara, J. (2026, April 6). AI data centres put insurers' capacity and credit risk under stress test: Report. Insurance Business. https://www.insurancebusinessmag.com/ca/news/technology/ai-data-centres-put-insurers-capacity-and-credit-risk-under-stress-test-report-570834.aspx

[5] Ibid. [4]

[6] Data Centers Powering AI Create Unprecedented Risk Accumulation Challenges for Insurers. https://riskandinsurance.com/data-centers-powering-ai-create-unprecedented-risk-accumulation-challenges-for-insurers/

[7] Ibid. [3]

[8] Ibid. [4]

[9] McGregor, S., Brundage, M., Cooper, A. F., Paskov, P., Ecoffet, A., Bucknall, B., Wei, K., Anderljung, M., Szpruch, L., Treece, B., Zick, T., Weil, G., Ozer, U., Kalinich, K., Gonzalez, J., Baranov, V., Koren, M., Laban, G., Arazi, G., Winand, H., Blum, D., Clowes, T., Kleinman, A., & Srinivasan, A. (2026, July). Underwriting the Agent Economy: The Blueprint for an AI Insurance Stack. arXiv. https://arxiv.org/abs/2607.11999

[10] The big deal about data centers (May 28, 2026) https://www.moodys.com/web/en/us/insights/insurance/the-big-deal-about-data-centers.html

[11] Michelle Chan and Lee Harris, "Lenders struggle to find insurance cover for mega data center projects," Financial Times, March 17, 2026. https://www.ft.com/content/5ba0cf1a-0d81-4479-a58c-3c8b5b088682?syn-25a6b1a6=1

[12] Bank of America analyst Savita Subramanian, cited in "A $1.6 Trillion Hidden Debt Problem Is Clouding the AI Boom," NAI 500, August 13, 2026. https://nai500.com/blog/2026/08/a-1-6-trillion-hidden-debt-problem-is-clouding-the-ai-boom/

[13] Ibid. [1]

[14] "Moody's opinion threatens to derail off-balance-sheet data centre deals," IFR, February 26, 2026. https://www.ifre.com/bonds/2390941/moodys-opinion-threatens-to-derail-off-balance-sheet-data-centre-deals

[15] Reuters. (2026, July 26). Meta, Microsoft and Amazon test limits of investor appetite for AI spending. Yahoo Finance. https://finance.yahoo.com/technology/ai/articles/meta-microsoft-amazon-test-limits-040100994.html

[16] Ibid. [2]

[17] "Data centres seek credit ratings to unlock billions in financing," Financial Times, February 21, 2026. https://www.ft.com/content/e0d9d5f2-c09d-426e-af03-193b488b7b1e

[18] Ibid. [17]

[19] "Meta's $27 billion bet turns AI compute into Wall Street's hottest new investment," Fortune, October 31, 2025. https://fortune.com/2025/10/31/metas-27-billion-bet-turns-ai-compute-into-wall-streets-hottest-new-investment/

[20] Ibid. [2]

[21] "Building on the Edge of a Cliff: The Unspoken Debt of Google and Meta," Odaily, July 28, 2026. https://www.odaily.news/en/post/5212208

[22] Aon plc, "Aon Expands Data Center Lifecycle Insurance Program to $5 Billion with Reliable by Design Approach to Digital Infrastructure," press release, July 20, 2026, https://www.prnewswire.com/news-releases/aon-expands-data-center-lifecycle-insurance-program-to-5-billion-with-reliable-by-design-approach-to-digital-infrastructure-302829377.html FM, "Markets/Coverages: FM Bumps Up Limits for Data Centers to $5B," Insurance Journal, January 13, 2026, https://www.insurancejournal.com/news/national/2026/01/13/854094.htm.