Investment Governance · 11/3/2026 · Updated 11/12/2026
AI Concentration Risk: Why 'Diversified' Is Now One Trade
AI concentration risk: chips, utilities, private credit and infrastructure funds now ride one AI spending cycle. How a family sees it across every bank.
By Pedro Souto
AI concentration risk is the hidden correlation in most “diversified” portfolios: chips, utilities, private credit and infrastructure funds depend on a few buyers spending on AI. AI-related firms are about half the S&P 500, up from a quarter in 2022 (Bank of England, July 2026). At €10–100M across several banks and funds, no statement adds it up. PWA’s consolidated reporting builds that view; talk to Pedro about AI concentration risk.
Three quarterly statements arrive in the same week. The first bank’s discretionary mandate reports global equities “diversified across eleven sectors” and a book of investment-grade corporate bonds. The second bank’s advisory account holds a structured note on a semiconductor basket and another on US utilities. The infrastructure fund’s letter celebrates a new data-centre platform and a gas-fired plant with a long-term offtake contract; the private-credit fund reports “senior secured loans to digital infrastructure, fully performing”. Each document is accurate. None mentions the others, and every one of them is paid, at the end of the chain, by the same handful of companies deciding how much to spend on AI next year.
Is my portfolio really diversified if everything depends on AI spending?
A portfolio is not diversified if its holdings lose money for the same reason. Chips, power equipment, data-centre landlords, their lenders and about half of the S&P 500 by value (Bank of England, July 2026) share one such reason: a slowdown in AI spending by a small group of buyers. The labels differ; the cash-flow source does not.
Correlation is a property of cash flows, not of sector labels. A utility and a chipmaker sit in different index sectors, but if both are paid from the same capital budget, they will fall together when that budget is cut. That is the working definition we use: AI concentration risk is the share of a family’s total wealth whose value or income depends, directly or through a chain of customers, lenders and guarantors, on continued AI investment by a few buyers.
The regulator for UK pension schemes now says the same thing in plain words. On 22 October 2026 The Pensions Regulator told trustees that “even a diversified portfolio may have hidden concentration risk because AI is increasingly embedded across sectors”, and that losses “can emerge simultaneously across apparently unrelated holdings and asset classes”. Its prescription is a look-through across every asset class, a severe-but-plausible AI downturn scenario, and harder questions to managers. A family has the same problem with fewer people to solve it.
The best map of the chain comes from the industry itself. In an essay published on NVIDIA’s blog on 10 March 2026 (older than six months, like every source marked “older” below), Jensen Huang described AI as a five-layer stack: “Energy → chips → infrastructure → models → applications”, adding that “every successful application pulls on every layer beneath it, all the way down to the power plant”. Read as an investor, that sentence is a dependency chart. Each layer is a customer of the one below, so each layer shares the same ultimate demand.
The five-layer AI stack: who depends on whom, company by company
Every layer of the AI stack reports a different product and the same customer. The latest filings show chip revenue, turbine backlogs, data-centre leases and model funding all tracing back to a few companies’ capital budgets, and those budgets are still rising. Where a company does not disclose what share of its business is AI-driven, we say so rather than estimate.
Market expectations have moved fast. The Bank of England’s July 2026 Financial Stability Report noted the Philadelphia Semiconductor Index was up 97% since the start of the year, and that Bloomberg consensus for AI hyperscalers’ 2028 capital expenditure had moved from under $600 billion at its December 2025 report to over $1 trillion. That is an observation about expectations, dated, not a forecast.
| Layer | Company | What the latest filing or report shows | Source, date |
|---|---|---|---|
| Energy | GE Vernova | Backlog $197bn; gas backlog and slot reservations 122 GW. AI share not disclosed | 8-K, 28 Oct 2026 |
| Energy | Siemens Energy | Record order backlog of €162bn. AI share not disclosed | Q3 FY2026 release, 5 Aug 2026 |
| Energy | PJM (US grid operator) | 2028/29 capacity cleared at the $325/MW-day cap, 6,831 MW short of the reliability requirement | PJM, 14 Jul 2026 |
| Energy | Vistra, Constellation, Schneider Electric, Eaton, Vertiv | AI share not disclosed in the sources we checked | — |
| Chips | NVIDIA | Data Center $89.0bn of $96.2bn quarterly revenue (92.5%); five direct customers hold 70% of receivables | 8-K, 26 Aug 2026; 10-Q, quarter to 26 Jul 2026 |
| Chips | Broadcom | AI semiconductors $16.7bn of $29.6bn quarterly revenue (56%) | 8-K, 2 Sep 2026 |
| Chips | TSMC | Revenue $46.72bn, up 41.2%; 79% of wafer revenue at 7nm and below. AI share not disclosed in the release | 6-K, 15 Oct 2026 |
| Chips | AMD | Warrants for up to 160 million shares each to OpenAI and Meta, vesting on GPU purchases; none vested | 10-Q, quarter to 27 Jun 2026 |
| Chips | ASML | €1.3bn for about 11% of Mistral AI, its largest shareholder | Reuters, 9 Sep 2025 (older) |
| Chips | Intel | US government bought a 9.9% stake for $8.9bn. AI share not disclosed | 8-K, 22 Aug 2025 (older) |
| Infrastructure | Amazon, Alphabet, Meta | 2026 capex guidance of about $220bn, $195–205bn and $135–145bn | Reuters and Bloomberg, Jul–Oct 2026 |
| Infrastructure | Microsoft | Quarterly additions to property and equipment $39.4bn, against $19.4bn a year earlier | 10-Q, quarter to 30 Sep 2026 |
| Infrastructure | Oracle | Remaining performance obligations $664bn; quarterly capex $28.5bn against $8.5bn; about $125bn of notes | 10-Q, quarter to 31 Aug 2026 |
| Infrastructure | CoreWeave | $8.5bn facility rated A3, secured by GPU servers and a customer contract | 8-K, 30 Mar 2026 (older) |
| Infrastructure | Digital Realty, Equinix | Digital Realty: record backlog of $1.8bn at its share. AI share not disclosed by either | 8-K, 29 Oct 2026 |
| Models | OpenAI | $122bn round at $852bn; Amazon $50bn, NVIDIA $30bn, SoftBank $30bn | Bloomberg, 31 Mar 2026 (older) |
| Models | Anthropic | $65bn at $965bn, with Google contributing under an earlier commitment | Bloomberg, 28 May 2026 |
| Models | xAI, Mistral | xAI acquired by SpaceX at a $250bn valuation; Mistral’s largest shareholder is ASML | Reuters, 3 Feb 2026 (older); Reuters, 9 Sep 2025 (older) |
| Applications | Software borrowers | Business development companies lent about $115bn to software firms, a fifth of their lending | BIS Bulletin 128, 14 Jul 2026 |
Chips: NVIDIA, AMD, Broadcom, TSMC, ASML and Intel
The chip layer is the most visibly concentrated, and the filings show it. NVIDIA’s Data Center segment produced $89.0 billion of its $96.2 billion revenue in the quarter to 26 July 2026, 92.5% of the total. Its 10-Q shows five direct customers holding 22%, 14%, 13%, 11% and 10% of accounts receivable, 70% between them. Broadcom reported AI semiconductor revenue of $16.7 billion in its quarter to 2 August 2026, 56% of total revenue, and guided to $21.7 billion for the next quarter.
TSMC does not break out an AI share in its quarterly release, but its revenue rose 41.2% year on year to $46.72 billion in the third quarter of 2026, with 79% of wafer revenue from 7-nanometre and more advanced processes. AMD’s exposure is written into its share register: warrants for up to 160 million AMD shares each to OpenAI and Meta, at $0.01, vesting as those customers buy AMD GPUs. ASML, the lithography supplier to the whole layer, became the largest shareholder in the model-maker Mistral AI in September 2025 (Reuters; older). Intel has counted the US government as a 9.9% shareholder since August 2025 (Intel 8-K; older), which puts a sovereign budget inside the same chain.
Power and grid: GE Vernova, Siemens Energy, Vistra, Constellation, Schneider, Eaton, Vertiv
The power layer sells turbines, transformers, cooling and electricity, and its order books have grown with data-centre demand. GE Vernova’s backlog reached $197 billion at 30 September 2026, with gas equipment backlog and slot reservations at 122 GW; its chief executive cited demand from “utility and data center customers”. Siemens Energy reported a record €162 billion order backlog on 5 August 2026. Neither discloses what share of that backlog is data-centre driven, and nor do Vistra, Constellation, Schneider Electric, Eaton or Vertiv in the sources we checked.
The grid shows the pressure in prices. PJM, a US regional grid operator, cleared its 2028/29 capacity auction at the $325 per MW-day cap on 14 July 2026, 6,831 MW short of its reliability requirement. That price is paid across the region, not only by data centres. For the wider energy picture, see our note on the energy shock 2026.
Infrastructure: hyperscaler capex, CoreWeave, data-centre REITs and their lenders
Hyperscaler capex in 2026 is the single largest source of demand for every layer below it. Amazon raised its annual capital spending forecast to $220 billion (Reuters, 30 July 2026); Alphabet set a 2026 range of $195–205 billion (Bloomberg, 23 July 2026); Meta raised the floor of its range to $135–145 billion (Reuters, 27 October 2026). Those three alone plan $550–570 billion this year. Microsoft’s additions to property and equipment were $39.4 billion in the quarter to 30 September 2026, double the $19.4 billion a year earlier; Alphabet’s were $51.4 billion, against $24.0 billion.
The builders now need outside money. Meta raised its planned equity offerings to $97.5 billion in October 2026 to fund its build-out, after reporting its first negative free cash flow in over a decade (Reuters, 27 October 2026). Oracle’s remaining performance obligations reached $664 billion at 31 August 2026, against quarterly capex of $28.5 billion and about $125 billion of senior notes. CoreWeave financed GPUs with an $8.5 billion facility rated A3 by Moody’s, secured by the servers and a customer contract (CoreWeave 8-K, March 2026; older). Digital Realty reported a record backlog on 29 October 2026; the private-credit and infrastructure funds that own and lend to these buildings are the next link, covered below.
Models: OpenAI, Anthropic, xAI, Mistral and who funds them
The model layer is funded largely by the layers it buys from. OpenAI closed a $122 billion round at an $852 billion valuation on 31 March 2026, with $50 billion from Amazon (of which $35 billion is contingent on an IPO or a technological milestone), $30 billion from NVIDIA and $30 billion from SoftBank (Bloomberg; older). Anthropic raised $65 billion at a $965 billion valuation on 28 May 2026, with Google contributing under an earlier commitment (Bloomberg). SpaceX acquired xAI in February 2026 at a $250 billion valuation for xAI (Reuters; older). Mistral’s largest shareholder, with about 11%, is ASML from the chip layer (Reuters, September 2025; older).
Cloud providers and chipmakers are therefore both suppliers and shareholders of their largest customers. When a model company’s valuation is reset, the reset appears in its investors’ balance sheets as well as in their order books.
Applications: SaaS repricing itself as AI
The application layer is exposed to AI from both sides, and that exposure already sits in private credit. Software companies either sell “AI” features whose cost depends on the model layer, or face competition from AI that compresses their pricing. The BIS found on 14 July 2026 that business development companies had lent about $115 billion to software firms, about a fifth of all their lending and over 80% of their technology portfolios; it noted that AI-driven revenue uncertainty “has not affected these loans yet”. Few application companies disclose an AI share of revenue, so we record it as not disclosed.
If your family also owns an operating business, it may be a further AI exposure of its own; our research on the industries most affected by AI covers that side.
Circular financing in AI: how vendor deals turn separate counterparties into one credit
Circular financing is when a supplier invests in, lends to or guarantees a customer that buys its products. It turns what look like separate counterparties into one credit, because the investor’s equity, the supplier’s revenue, the lender’s collateral and the guarantor’s contingent liability all depend on the same end-buyer continuing to pay. The BIS measured how common it is on 1 October 2026: 55.2% of incoming investment into AI firms in 2021–2025 came from other AI firms, and 46.4% of AI-to-AI deal value also involved a commercial supply relationship.
Most of the headline deals were announced in 2025 or early 2026, including the Microsoft–NVIDIA–Anthropic partnership (November 2025; older) and OpenAI’s Oracle contract (September 2025; older); the table records each one’s latest status from filings or wire reporting.
| # | Money flows | What comes back | Latest status, source, date |
|---|---|---|---|
| 1 | Amazon invests up to $50bn in OpenAI | OpenAI commits $100bn to Amazon computing over eight years, on top of a $38bn contract | $35bn contingent on IPO or milestone; FT, 27 Feb 2026 (older); Bloomberg, 31 Mar 2026 (older) |
| 2 | NVIDIA invests $30bn in OpenAI | Much of the round goes on chips and cloud services from its investors; the September 2025 plan for up to $100bn was abandoned | FT, 27 Feb 2026 (older); Bloomberg, 31 Mar 2026 (older) |
| 3 | NVIDIA guarantees up to $105bn of lease and power payments for an OpenAI affiliate’s Ohio campus | The site will exclusively host NVIDIA AI infrastructure, with limited exceptions | NVIDIA 10-Q (guarantees entered August 2026); Reuters, 31 Aug 2026 |
| 4 | OpenAI agrees about $300bn of Oracle computing over roughly five years | Oracle borrows and spends on the chip and building layers below it | Reuters citing WSJ, 11 Sep 2025 (older); Oracle RPO $664bn, 10-Q, 31 Aug 2026 |
| 5 | NVIDIA and Microsoft commit up to $10bn and $5bn to Anthropic | Anthropic commits to buy $30bn of Azure compute capacity | Microsoft, 18 Nov 2025 (older) |
| 6 | AMD grants OpenAI and Meta warrants for up to 160m shares each at $0.01 | Tranches vest as they buy AMD Instinct GPUs | None vested at 27 Jun 2026; AMD 10-Q |
| 7 | NVIDIA and AMD buy $3bn and $1bn of Meta’s equity offering | Meta funds a capex budget spent partly on their products | Reuters, 27 Oct 2026 |
| 8 | NVIDIA commits $36bn to cloud service agreements with AI clouds | Those AI clouds buy NVIDIA data-centre infrastructure; NVIDIA also commits $25bn of equity to model makers and infrastructure financiers | NVIDIA 10-Q, quarter to 26 Jul 2026 |
The reason circularity matters for a family is mechanical. In a normal supply chain, a supplier’s risk is its customer’s ability to pay. In a circular one, the supplier also owns the customer’s equity, guarantees its rent or buys back its capacity, so one disappointment hits the supplier’s revenue, investment portfolio and contingent liabilities at once. The BIS put it on 27 October 2026: “off-balance sheet commitments from a few large firms help to transform the credit rating of the debt”. The rating of the vehicle is, in effect, the rating of the guarantor’s continued appetite.
That is why a family cannot net these exposures by holding NVIDIA in one bank, a data-centre loan fund in another and an infrastructure fund in a third. They are three tickets on one credit. BIS Bulletin 137 also notes these relationships “increase opacity”, because many of the firms involved are private and disclose little.
Holding the chipmaker, its customer and the customer’s lender in three different banks?
Send Pedro the list of banks, funds and vehicles. The first conversation traces which of them sit on the same chain, not what to do about it.
Talk to Pedro about circular exposureHow much of a “balanced” portfolio is one bet on AI capex?
In a hypothetical €50M family with three banks, two private-equity funds and an infrastructure fund, about 35% of total wealth rests on one demand driver; halve every assumption and it is still 17%. Part sits in equities, since AI-related firms are about half the S&P 500 (Bank of England, July 2026). The rest sits in sleeves sold as safe.
The profile is common: wealth spread across multiple banks, each reporting its own slice as balanced. The money labelled safe is where the exposure hides, for four reasons.
How much of the S&P 500 is AI-related?
About half, on the Bank of England’s measure: “Within that index, AI companies now account for around half, up from around a quarter in 2022” (Financial Stability Report, 7 July 2026). Other published estimates land above or below that figure, and they disagree because the definitions disagree: some count only chipmakers and hyperscalers, others add power equipment, utilities, data-centre REITs and software companies that sell AI features. The narrower the definition, the lower the number; none of them is zero.
For a global tracker, the same report supplies the second number: the S&P 500 is around half of global equity capitalisation. Half of half means roughly a quarter of world equity value sits in US AI-related companies, before counting TSMC, ASML or Siemens Energy. That quarter is our own arithmetic on two Bank of England figures, not a published statistic. The Pensions Regulator adds the mechanism: in market-cap-weighted funds, “concentration will increase as the market value of related companies rises”. A passive fund buys more of the trade as the trade goes up.
Private credit exposure to data centres and GPU-backed loans
Private credit has become a main lender to the AI build-out. The OECD, cited by the Bank of England in July 2026, estimates private credit’s share of AI investment financing rose from 9% in 2024 to 34% in 2025. A Morgan Stanley forecast that the Bank of England reproduces in the same report (Chart 2.2) has private credit funding about $700 billion of global data-centre capital spending in 2026–28, second only to $1.46 trillion of equity; it is a bank’s projection, not an outcome.
The structures are what make the exposure hard to see. The BIS described in March 2026 (older) how hyperscalers place data centres in joint ventures or special-purpose vehicles that raise private debt, while the hyperscaler holds a minority stake and signs long-term leases or guarantees, “keeping most of the associated debt off the hyperscaler’s balance sheet”. GPU-backed loans such as CoreWeave’s add a second layer: the collateral is chips whose economic life is shorter than a building’s. The Financial Stability Board sized private credit at $1.5–2 trillion in May 2026 and warned it “has not been tested during a severe economic downturn”. If your private-equity allocation includes GP-stake or secondaries funds, ask which of the managers they own run data-centre or digital-infrastructure strategies.
Infrastructure funds sold as the uncorrelated sleeve
Infrastructure and “real assets” funds are usually sold for income that does not move with equities. In an AI build-out, many of their newest assets are data centres, grid connections and generation contracted to the same hyperscalers. The Pensions Regulator names technology, infrastructure, energy, real estate and financial assets as areas that may all tie back to the same AI-driven growth. The income is real; its source is not independent of the equity book.
Discretionary mandates and structured products on the same names
A discretionary mandate is an arrangement in which the bank decides what to buy and sell within agreed limits, and reports against a benchmark. If that benchmark is a global or US index, the mandate inherits the index’s AI weight, and a manager who stays close to it adds no protection. A structured note written on a semiconductor or large technology basket carries the same stack in a different wrapper. Each bank reports its own mandate as diversified, and each is right inside its own walls.
Worked example: three banks, two PE funds, one infrastructure fund
This is a hypothetical family, with round figures and labelled assumptions; it is a method, not a measurement of anyone’s portfolio. Two banks run 60/40 discretionary mandates, a third holds structured notes, and three funds sit outside the banks. The AI-dependent share of each sleeve is an assumption, anchored where possible on the public figures above.
| Sleeve | € allocated | Assumed AI-dependent share | Basis of the assumption | € on one driver |
|---|---|---|---|---|
| Listed equity in two mandates (60% of €27M) | €16.2M | Assumption: 30% | Our arithmetic, not a published figure: AI firms are about half the S&P 500, which is about half of world equity value (BoE), so about 25%; plus non-US stack names | €4.9M |
| Bonds in two mandates (40% of €27M) | €10.8M | Assumption: 10% | Hyperscalers were 3% of US investment-grade debt at end-2025; AI issuers 41% of new non-refinancing US high-yield issuance in 2026 (BoE) | €1.1M |
| Structured notes and single stocks (third bank) | €8.0M | Assumption: 70% | Notes written on semiconductor and large technology baskets | €5.6M |
| PE fund 1: growth buyout, software-heavy | €4.0M | Assumption: 40% | Application-layer companies priced on AI; software borrowers already a fifth of BDC lending (BIS) | €1.6M |
| PE fund 2: private credit | €4.0M | Assumption: 30% | Private credit financed 34% of AI investment in 2025 (OECD, via BoE) | €1.2M |
| Infrastructure fund | €5.0M | Assumption: 60% | Data centres, grid and generation contracted to hyperscalers | €3.0M |
| Property and cash | €2.0M | Assumption: 0% | Residential property and deposits | €0.0M |
| Total | €50.0M | About 35% | Seven sleeves, three custodians, one driver | €17.4M |
No statement in this family’s post shows the €17.4M. The mandates show sector weights, the structured-note statement shows mark-to-market values, and the funds show quarterly letters. Halve every assumption and one demand driver still carries about 17% of total wealth. This is the portfolio overlap problem one level down: not the same stock held twice, but different holdings paid by the same buyers. If your own numbers look like this table, talk to Pedro about AI concentration risk across your banks.
What happens to utilities and private credit if hyperscaler spending slows?
Losses would travel along contracts, not sector labels. Utilities feel it through capacity prices and connection commitments, private credit through vehicle leases and GPU collateral, and suppliers through guarantees and the equity they hold in customers. The IMF warned on 13 October 2026 that weaker AI investment “could lead to a correction” amplified by “rising leverage and concentrated exposures”.
We do not forecast the size or timing of any fall. What can be mapped now is the route a slowdown would take, and who stands behind each step.
| Channel | Who pays whom | What stands behind it | Where a family holds it |
|---|---|---|---|
| Equity repricing | Markets reprice the stack together | Nothing; marked to market daily | Index funds, mandates, structured notes |
| Data-centre leases | Hyperscaler or AI lab pays rent to a vehicle or landlord | Lease terms and guarantees from a few large firms | Infrastructure funds, private credit, REITs |
| GPU-backed loans | AI cloud pays lenders from customer contracts | Chips with a short economic life, plus the contract | Private credit funds, asset-backed securities |
| Power markets | Data centres and all other customers pay utilities and generators | Regulated tariffs and capacity auctions | Utilities in the index, infrastructure funds |
| Supplier guarantees and stakes | Suppliers guarantee or own their customers | The supplier’s own balance sheet | Chipmakers in every equity sleeve |
| Bank and insurer links | Banks fund vehicles and lend to private-credit funds | Fund assets and vehicle collateral | Bank shares and bonds, insurance-linked products |
Each row has a public marker. The BIS warned in March 2026 (older) that banks funding these vehicles create “new shock transmission channels” and that guarantees could be activated, passing losses back to the hyperscaler. The FSB noted in May 2026 that “interconnections between private credit funds and banks, insurers, and private equity firms are deepening”. The Bank of England’s July 2026 scenario for a fall in AI share prices included a 2.2 percentage point fall in UK GDP, a scenario rather than a forecast, and one that reaches far beyond technology holders.
AI bubble vs dot-com bubble: the financing is different
We do not argue that a reversal would cost more than in 2000; we argue that it would reach further. The 2000 telecom build-out ran mainly through listed equity and telecom debt. This build-out also runs through private credit, off-balance-sheet vehicles, regulated power markets, supplier guarantees and, in Intel’s case, a government shareholding. The second-order holders (grid operators, data-centre landlords, the funds and banks lending to them) sit in the sleeves families bought for safety.
The debt side is growing fast. The Bank of England reported in July 2026 that AI hyperscalers’ bond issuance in the first half of 2026 had already exceeded all of 2025, and that AI issuers accounted for 41% of non-refinancing US high-yield issuance so far in 2026. The BIS added on 27 October 2026 that spillovers to US Treasury and investment-grade markets had so far been “small and short-lived”. The plumbing is wider than in 2000; whether it holds is a question for evidence, not for this article.
Three steps to see AI concentration risk across every bank and fund
A prepared family does three things, all operational and none of them regulated advice: it consolidates every holding into one look-through view, classifies holdings by what they depend on, and decides in advance who owns the decision when the evidence changes. The Pensions Regulator asks trustees for the same sequence. Steps a and b are what a written second opinion covers; step c is family governance.
PWA designs and runs the operating layer that makes these steps routine. PWA does not manage money, hold mandates, take commissions or give regulated investment, tax or legal advice. When I map a family’s holdings, I start from the custodian statements and fund documents, not the product names, because the names are where the correlation hides.
a. Consolidate every holding across every custodian into one look-through view
The first step is one statement of everything the family owns, valued on the same date. Consolidated reporting is that statement: every bank account, mandate, structured note, fund and entity on one balance sheet, with funds looked through to what they hold. A custodian is the bank that holds the assets and reports them; with three custodians you have three partial views and no total.
The look-through matters most for the sleeves outside the banks. A private-credit fund’s quarterly letter names its sectors; its loan tape names the borrowers. An infrastructure fund’s report lists assets; its offtake schedule lists who pays for them. The mechanics of building this across several banks are in our guide to consolidated reporting across multiple banks. PWA builds it through its consolidated reporting and administration service. If the family’s holdings span several banks and funds, send Pedro the list of custodians and funds.
b. Classify holdings by what they depend on, not by asset class
The second step replaces the asset-class label with three questions: what demand pays this holding, which counterparty pays it, and what financing chain sits between them. Asset class tells you how a holding is valued; dependency tells you when it fails. Classified this way, the worked example’s seven sleeves collapse into fewer drivers than the statements suggest.
| Holding | Statement label | Demand driver | Counterparty | Financing chain |
|---|---|---|---|---|
| Chipmaker shares | Equity, technology | Hyperscaler capex | A few direct customers | Vendor stakes and guarantees |
| Utility note | Structured product, utilities | Power demand, partly data centres | Ratepayers and data centres | Capacity auctions, regulated tariffs |
| Data-centre loan | Private credit, senior secured | Hyperscaler or AI lab leases | Tenant and guarantor | Off-balance-sheet vehicle |
| Infrastructure fund asset | Real assets | Contracted capacity | Hyperscaler offtaker | Fund leverage plus lease |
This is the core of investment and wealth architecture: PWA designs the classification, maps each holding to it, and keeps it current as funds report. It does not tell the family what to sell. To see your own holdings classified this way, talk to Pedro about classifying holdings by demand driver.
c. Decide in advance who owns the decision, and what evidence triggers it
The third step is a written rule, agreed before it is needed. An investment policy statement, the document that sets a family’s objectives, limits and decision rights for its investments, should state the maximum share of total wealth allowed on one demand driver, who reviews it, and which evidence prompts a review: a change in hyperscaler capex guidance, a covenant event in a data-centre vehicle, a guarantee drawn. A family council, the body through which family members govern shared wealth, decides who holds that authority when the principal is unavailable.
The same rule protects the family from the next pitch. The fund manager offering an “uncorrelated” AI-adjacent vehicle should be asked what pays its assets, and the answer belongs in the same classification; our note on how to evaluate an investment opportunity sets out the questions, and our piece on what is investment governance covers the wider frame. PWA writes these rules through its family governance work. To set the threshold and the owner before the evidence changes, talk to Pedro about an AI exposure decision rule.
Checklist: AI concentration risk across every bank and fund
- List every custodian, fund, vehicle and direct holding the family owns, with the latest statement date for each.
- Look through every fund to its underlying holdings, borrowers or contracted counterparties.
- Tag each holding with its demand driver, its paying counterparty and its financing chain.
- Total the share of wealth on AI spending, and repeat it with every assumption halved.
- Record which holdings carry guarantees, GPU collateral or off-balance-sheet vehicles.
- Write the threshold, the reviewer and the evidence triggers into the investment policy statement.
- Ask every manager, including new ones pitching, how they measure this concentration.
Quarterly statements that each say “diversified”?
Tell Pedro how many banks, mandates and funds are involved. The first conversation sets out what a look-through would measure and who in the family would own the answer.
Talk to Pedro about a look-through viewQuestions we hear
What is AI concentration risk?
AI concentration risk is the share of your total wealth whose value or income depends on a few companies continuing to spend on AI. It sits in chips and software, and also in utilities, data-centre landlords, private credit and infrastructure funds, because each is paid, directly or through a chain, by the same buyers.
Is the S&P 500 still diversified?
It holds about 500 companies, but the Bank of England estimated in July 2026 that AI-related firms make up around half of its value, up from a quarter in 2022. A tracker still spreads money across names; it no longer spreads it across drivers. Concentration rises automatically as those companies’ market value grows.
What is circular financing in AI?
Circular financing is when a supplier invests in, lends to or guarantees a customer that buys its products. NVIDIA invested in OpenAI and guaranteed up to $105 billion of lease and power payments for an OpenAI affiliate’s Ohio data centre; AMD gave OpenAI and Meta warrants that vest on GPU purchases. The same money appears as investment, revenue and collateral.
How exposed is private credit to data centres?
More than most investors assume. The OECD, cited by the Bank of England, estimates private credit’s share of AI investment financing rose from 9% in 2024 to 34% in 2025. Much of it sits in off-balance-sheet vehicles backed by hyperscaler leases and guarantees, held by funds that families own through private-markets allocations.
How is the AI bubble different from the dot-com bubble?
We do not call it a bubble; we compare the plumbing. The 2000 build-out ran mainly through listed equity and telecom debt. This one also runs through private credit, off-balance-sheet vehicles, power markets and guarantees between suppliers and customers, so a slowdown reaches sleeves that families bought as the uncorrelated part of the portfolio.
What happens to my portfolio if the AI bubble bursts?
Nobody can tell you the size of a fall, and we do not try. What can be measured is where it would land: the share of your wealth whose income depends on AI spending, which lenders and guarantees stand behind it, and which holdings would fall together. That map is the useful answer.
How do I protect my portfolio from an AI bubble?
Start operationally: consolidate every holding across every bank and fund, classify each by what it depends on, and write down who decides what when the evidence changes. What to buy or sell is a decision for the family and its regulated advisors; the measurement and the decision rule come first.
What is hyperscaler capex?
Hyperscaler capex is the capital spending of the largest cloud and platform companies on data centres, chips, networking and power. For 2026, Amazon guided to about $220 billion, Alphabet to $195–205 billion and Meta to $135–145 billion. That spending is the revenue of the layers below them in the AI stack.
Sources
All pages opened and checked on 3 November 2026. Sources dated before 3 May 2026, more than six months old, are marked “older” here and in the text.
- Bank of England, “Financial Stability Report – July 2026”, 7 July 2026. Supports: AI companies around half of the S&P 500, up from around a quarter in 2022; S&P 500 around half of global equity capitalisation; Philadelphia Semiconductor Index up 97% year to date; 2028 hyperscaler capex consensus from under $600bn to over $1tn; OECD estimate of private credit’s share of AI financing (9% in 2024, 34% in 2025); Morgan Stanley forecast of 2026–28 data-centre funding sources, reproduced in Chart 2.2 ($1.46tn equity, $700bn private credit); hyperscalers 3% of US investment-grade debt at end-2025; first-half 2026 hyperscaler issuance above all of 2025; AI issuers 41% of non-refinancing US high-yield issuance; 2.2 percentage point UK GDP scenario. bankofengland.co.uk
- The Pensions Regulator, “AI concentration risk: Action for trustees”, 22 October 2026. Supports: hidden concentration in diversified portfolios, look-through across asset classes, simultaneous losses, passive-fund concentration, questions for managers. thepensionsregulator.gov.uk
- NVIDIA (Jensen Huang), “AI Is a 5-Layer Cake”, 10 March 2026 (older). Supports: the five-layer stack and its dependency. blogs.nvidia.com
- NVIDIA, Form 8-K, second-quarter fiscal 2027 results, 26 August 2026. Supports: revenue $96.2bn; Data Center $89.0bn. sec.gov
- NVIDIA, Form 10-Q for the quarter ended 26 July 2026. Supports: five direct customers at 22%, 14%, 13%, 11% and 10% of receivables; guarantees capped at $105bn for an OpenAI affiliate’s Ohio campus; $36bn AI cloud service agreements; $25bn equity investment commitments. sec.gov
- Broadcom, Form 8-K exhibit, third-quarter fiscal 2026 results, 2 September 2026. Supports: revenue $29.6bn; AI semiconductor revenue $16.7bn; next-quarter AI guidance $21.7bn. sec.gov
- TSMC, Form 6-K, third-quarter 2026 results, 15 October 2026. Supports: revenue $46.72bn, up 41.2%; 79% of wafer revenue at 7nm and below. sec.gov
- AMD, Form 10-Q for the quarter ended 27 June 2026. Supports: OpenAI and Meta warrants for up to 160 million shares each; none vested. sec.gov
- Reuters (via Free Malaysia Today), “Mistral AI raises €1.7bil as ASML becomes its top shareholder”, 9 September 2025 (older). Supports: ASML €1.3bn, about 11%. freemalaysiatoday.com
- Intel, Form 8-K exhibit 99.1, 22 August 2025 (older). Supports: US government 9.9% stake, $8.9bn. sec.gov
- GE Vernova, third-quarter 2026 results, Form 8-K, 28 October 2026. Supports: backlog $197bn; gas backlog and slot reservations 122 GW. sec.gov
- Siemens Energy, “Earnings release Q3 FY 2026”, 5 August 2026. Supports: order backlog €162bn. siemens-energy.com
- PJM, “PJM Capacity Auction Procures 138,318 MW of Generation Resources”, 14 July 2026. Supports: $325/MW-day cap; 6,831 MW short. pjm.com
- Reuters (via The Star), “Amazon lifts investment plans after strong cloud sales”, 30 July 2026. Supports: capex forecast raised to $220bn. thestar.com.my
- Bloomberg (via Bloomberg Law), “Alphabet Falls as $205 Billion Spending Plan Fuels AI Cost Fear”, 23 July 2026. Supports: 2026 capex range $195–205bn. bloomberglaw.com
- Reuters (via The Star), “Meta eyes record $97.5 billion equity raise to fund AI expansion”, 27 October 2026. Supports: offering raised to $97.5bn; capex $135–145bn; NVIDIA $3bn and AMD $1bn; first negative free cash flow in over a decade. thestar.com.my
- Alphabet, Form 8-K exhibit 99.1, third-quarter 2026 results, 28 October 2026. Supports: purchases of property and equipment $51.4bn against $24.0bn. sec.gov
- Microsoft, Form 10-Q for the quarter ended 30 September 2026. Supports: additions to property and equipment $39.4bn against $19.4bn. sec.gov
- Oracle, Form 10-Q for the quarter ended 31 August 2026. Supports: remaining performance obligations $664bn; capex $28.5bn against $8.5bn; about $125bn of senior notes and borrowings. sec.gov
- CoreWeave, Form 8-K, 30 March 2026 (older). Supports: $8.5bn DDTL 4.0 facility, A3 rating, secured by GPU servers and a customer contract. sec.gov
- Digital Realty, Form 8-K exhibit, third-quarter 2026 results, 29 October 2026. Supports: record backlog of $1.8bn at its share. sec.gov
- Financial Times (via The Irish Times), “OpenAI secures up to €93bn in record funding deal”, 27 February 2026 (older). Supports: Amazon $15bn upfront plus $35bn conditional; OpenAI’s $100bn Amazon compute commitment on top of $38bn; NVIDIA $100bn plan abandoned. irishtimes.com
- Bloomberg (via Bloomberg Government), “OpenAI Valued at $852 Billion After Completing $122 Billion Round”, 31 March 2026 (older). Supports: $122bn at $852bn; Amazon $50bn, NVIDIA $30bn, SoftBank $30bn. bgov.com
- Bloomberg (via Bloomberg Law), “Anthropic Raises at $965 Billion Valuation, Eclipsing OpenAI”, 28 May 2026. Supports: $65bn at $965bn; Google contribution. bloomberglaw.com
- Microsoft, “Microsoft, NVIDIA and Anthropic announce strategic partnerships”, 18 November 2025 (older). Supports: Anthropic $30bn Azure commitment; NVIDIA up to $10bn, Microsoft up to $5bn. blogs.microsoft.com
- Reuters (via Dawn), “SpaceX acquires xAI in record-setting deal”, 3 February 2026 (older). Supports: xAI valued at $250bn. dawn.com
- Reuters (via The Star), “OpenAI, Oracle sign $300 billion computing deal, WSJ reports”, 11 September 2025 (older). Supports: about $300bn over roughly five years. thestar.com.my
- Reuters (via The Star), “OpenAI issued warrants worth $5.5 billion in SB Energy, WSJ reports”, 31 August 2026. Supports: NVIDIA guarantee of up to $105bn for OpenAI’s Ohio data centre. thestar.com.my
- BIS, Bulletin 137, “Circular relationships among AI firms”, 1 October 2026. Supports: 55.2% of incoming AI investment from other AI firms; 46.4% of AI-to-AI deal value with a supply relationship; opacity. bis.org
- BIS, Bulletin 140, “Building AI on credit: structures and spillovers”, 27 October 2026. Supports: off-balance-sheet commitments transform credit ratings; small, short-lived spillovers so far. bis.org
- BIS, Quarterly Review, “Financing the AI infrastructure boom: on- and off-balance sheet borrowing”, 16 March 2026 (older). Supports: vehicle structures, guarantees, bank funding as new shock transmission channels. bis.org
- BIS, Bulletin 128, “AI disruption in private credit: exposure to software firms in BDCs”, 14 July 2026. Supports: about $115bn lent to software firms, a fifth of BDC lending, over 80% of technology portfolios. bis.org
- Financial Stability Board, “Report on Vulnerabilities in Private Credit”, 6 May 2026. Supports: $1.5–2tn market; untested in a severe downturn; deepening links with banks and insurers. fsb.org
- IMF, “Global Financial Stability Report”, October 2026, 13 October 2026. Supports: AI-related correction risk amplified by leverage and concentrated exposures. imf.org
Pedro Souto is the founder of PWA — Private Wealth Advisory. He works with UHNW founders, families, and single-family offices to build an independent operating layer around complex wealth — combining markets, analytics, technology, and governance to turn fragmented advice into one coherent picture, and to make someone, finally, responsible for the whole of it.