By late June 2026, the AI capex boom 2026 families must underwrite is no longer a story about a single chipmaker. It is a concentration file: listed hyperscalers spending on data centres and accelerators, a private-cloud and colocation sleeve that often rents the same demand, and family books that already own the cluster three times. This note is general information for families and family offices. It is not a valuation, a trading recommendation, or personalised advice. Structures depend on facts and on counsel.
The listed side is unusually well documented. NVIDIA’s Form 10-K for the fiscal year ended 25 January 2026, filed with the U.S. Securities and Exchange Commission in February 2026, is a public description of data-centre demand, customer concentration, export controls, and supply. Calendar-year 2025 annual reports from other large U.S. technology issuers, also filed in early 2026, describe capital expenditure on cloud, AI training, and inference as a first-order use of cash. The family-office mistake is to treat those filings as a scoreboard of “winners” and then ignore look-through. Vellum Finance treats the boom as a budget: how much of the family’s listed, private, and operating-company risk is the same capex cycle.
AI capex boom 2026 families: listed concentration, not a beauty parade
Concentration is the first fact. The IMF Global Financial Stability Report of April 2026 describes how a surge in capital expenditure by large global technology firms in artificial intelligence, the so-called hyperscalers, supported equity-market leadership, and how a few large firms dominate market valuations. IMF staff also flag that sizeable AI-related investment can raise debt and interconnectedness, and that circular financing along the AI value chain is a vulnerability if investment slows. This article will not invent a family-office valuation or a price target. It will use that official language: concentration, interconnectedness, and the possibility that mega-cap names fail to generate the returns already in the price.
The same GFSR is careful about what it does not know. Conflict and energy shocks can slow AI investment. Earnings of hyperscalers had, in the staff’s telling to March 2026, kept pace with capex, with free cash flow still high. That is a system-level observation, not a reason for a family to increase a single-name weight. A family that already holds a global equity tracker, a “quality” UCITS, and a technology SMA often owns the same issuers three times. Look-through weights, not share-class names, are the book. Adjacent reading sits in Vellum’s 2026 note on public markets quality after 2022-2026.
Read the filings, not the slogan. NVIDIA’s investor and SEC EDGAR 10-K archive are the primary documents for the accelerator layer. Microsoft, Amazon, Alphabet, and Meta publish capital-expenditure and cloud commentary in their own 10-K and 10-Q packs on SEC EDGAR. Microsoft’s fiscal year ends in June; by the date of this article the relevant public pack is the FY2025 10-K and the FY2026 quarterly reports, not a later annual report that has not been filed. Date the document. Do not mix fiscal years.
What a 10-K actually tells a family office
A Form 10-K is a risk and business description under U.S. securities law. Item 1 describes the business. Item 1A lists risk factors: customer concentration, export controls, supply, competition, and regulation. Item 7 discusses liquidity and capital resources, including capex. Families who skip to a revenue infographic miss the only pages that matter for a family book: who the customers are, how concentrated they are, what happens if a government restricts shipments, and how much cash is already committed to facilities and equipment.
Customer concentration is the family-office translation of “AI winner.” If a handful of cloud providers buy a large share of accelerators, the chipmaker’s cycle and the hyperscaler’s capex cycle are not two independent bets. They are a chain. If those same hyperscalers also appear in the family’s listed equity sleeve, the chain is on both sides of the invoice. Export-control risk is a second translation. A 10-K that spends pages on licensing, denied parties, and geographic restrictions is telling you that policy can change the addressable market faster than a product cycle. That is a governance input, not a reason to day-trade the name.
Capex in the buyer’s 10-K is the other side of the same chain. A cloud issuer that describes multi-year data-centre and GPU commitments is telling you that cash will leave the firm whether or not the family’s private-cloud GP has a good quarter. Depreciation, power contracts, and residual value of specialised chips sit in that cash story. This article will not invent a capex dollar total for any issuer. The family’s job is to read the issuer’s own table and the cash-flow statement, then ask how much of the family book is levered to that table.
Private-cloud, colocation, and the same tenant
Private markets often sell “AI infrastructure” as diversification from listed tech. Sometimes it is. A European colocation site with a diversified tenant roster, contracted power, and a conservative loan-to-value can be a real-estate and power business that happens to host servers. Sometimes it is not. A development that is pre-leased to a single hyperscaler, financed with construction debt, and located in a power-constrained market is a tenant-concentration and completion risk wearing an AI label. The April 2026 GFSR, in its commercial-real-estate discussion, notes that data centres have been the outperforming CRE subsector, with hyperscaler tenant concentration, power constraints, and speculative development among the vulnerabilities. That is an official caution, not a marketing rebuttal.
Family offices that commit to private-cloud or data-centre funds should demand the same look-through they would demand on a listed name: tenant list, remaining lease term, power status, construction milestones, and overlap with listed hyperscalers already owned. Open architecture matters here. A house fund labelled “digital infrastructure” can be a concentrated bet on the same four customers the public book already holds. The 2026 test of independence is in Vellum’s note on open architecture versus in-house funds.
Circular financing and why the family should care
IMF staff in April 2026 pointed to circular financing along the AI value chain: arrangements in which suppliers, customers, and financiers recycle exposure through one another. The family does not need a forensic reconstruction of every round-trip. It needs to know whether a private commitment, a listed holding, and a vendor-financing note in an operating company are the same names. If they are, a slowdown in capex is not three uncorrelated disappointments. It is one. Governance is a single look-through of issuers and counterparties, updated when a 10-Q arrives, not a new narrative every time a conference keynote lands.
Power, water, and the European family constraint
Data centres consume power and, in some designs, water. European families who own land, industrial sites, or municipal-adjacent property hear “AI campus” as a planning and grid file long before they hear it as a return. Grid connection queues, local opposition, and environmental permits are not footnotes. They are the difference between a leased facility and a delayed SPV. The European Central Bank treats energy prices and investment as part of the euro-area activity and inflation file. A family that underwrites a private data-centre commitment without a power-availability paper is underwriting a construction story, not an AI story.
Listed energy and utility names can be a hedge, a duplicate, or a red herring. A utility that is already in the public book because of the energy-security sleeve is not automatically a diversifier for an AI-capex sleeve. It may be the other side of the same power constraint. Map it. Do not assume the labels “tech” and “utilities” are orthogonal.
How to budget concentration without inventing a fair value
A family office does not need a proprietary discounted-cash-flow for NVIDIA or a hyperscaler in order to act. It needs ceilings. An issuer ceiling, a sector ceiling, a listed-private overlap ceiling, and a currency ceiling (because the cluster is still heavily U.S. dollar). Those ceilings belong in the investment policy, with a look-through report at least quarterly. Rebalancing is the mechanism. Without a written trigger, AI capex becomes a one-way weight that only grows when the names rise.
Implementation can be passive (accept the index weight and write that down), active (underweight the cluster on purpose), or custom (direct indexing that caps issuers the family already owns privately or through a founder line). Each path is honest if the family can say what it owns. None of them requires a house valuation that this article will not produce. Advice and look-through sit on the Vellum Finance services map. Vellum is a fee-only multi-family office. The family pays for the map, not for a product that packages the boom.
Conclusion
The AI capex boom 2026 families actually face is a chain: hyperscaler spending, accelerator suppliers, private data-centre and cloud vehicles, and public indexes that already concentrate in the same issuers. NVIDIA’s February 2026 10-K and the other large issuers’ EDGAR packs are the listed evidence. The IMF’s April 2026 GFSR is the system-level caution on concentration, circular financing, and the possibility that investment slows. Private-cloud is diversification only when the tenant, the power, and the financing are not the listed book in another wrapper. Budget ceilings, look-through, and dated filings beat a beauty parade. This article will not invent a fair value. The family’s job is to refuse to own the same capex cycle four times without naming it.
Discretion. Stability. Prosperity.
Team Vellum
A team of passionate professionals who combine their expertise to bring knowledge through Vellum Finance & Patrimoine blog articles. Each member writes about their own field of expertise, cross referencing with our colleagues own fields to ensure the highest quality of information possible in all our content.
