AI financing web raises concentration risk for UK investors

Sona Asset Management, a London-registered investment firm, has mapped the financial ties running through the artificial intelligence industry, covering transactions over the three and a half years to 31 August 2026. Its researchers, Craig Nicol, Oleg Melentyev and Charlie Callan, traced financing, ownership and commercial links among AI laboratories, cloud providers, chipmakers and the infrastructure firms that build the data centres behind them. The Financial Times published an account of the work on 18 September 2026.

The headline number is large: 176 transactions with an aggregate announced value of $3.6tn (US dollars, as reported), involving 202 entities, to 31 August 2026. But the more useful finding for a UK reader sits underneath the total. Around 120 of those 176 transactions, roughly two in three, carry Sona's own "high-circularity" flag, meaning the entities involved are connected through multiple overlapping relationships, often with capital moving in a loop rather than a straight line from investor to borrower.

That matters in Britain because UK asset managers, banks, insurers and pension schemes hold AI-related exposure across several apparently distinct channels: global equity funds, investment-grade and high-yield bonds, private credit, infrastructure funds and direct counterparty relationships. Sona's mapping, and separate analysis from the Bank of England, both point to the same underlying question: how many of those channels ultimately depend on the spending decisions of the same small group of companies? This is not a forecast that AI financing is about to collapse. It is a description of how concentrated and interconnected the system has become. The aggregate scale of UK institutions' exposure to this network is not publicly quantified, and some transaction-level guarantees, recourse arrangements and private-market exposures remain undisclosed.

What Sona actually measured

Sona's $3.6tn figure is not debt, and it is not money already spent. It is the aggregate announced value of transactions that mix equity investments, loans, and lease or purchase commitments (different financial instruments with different risk profiles) added together. Some of the 176 transactions had no disclosed value at all, so the total is incomplete rather than a complete measure of the network's scale. The published account does not describe how Sona treats conditional, cancellable or multi-year commitments within that figure, and the underlying data dictionary was not publicly available at the time of writing.

The high-circularity flag is Sona's own analytical classification, not a regulatory designation. It identifies entities linked through several relationships at once. For example, a chipmaker that invests in an AI laboratory that then commits to buying compute from a cloud provider that leases capacity from a data-centre operator financed partly by the chipmaker itself. Circularity of this kind is not automatically improper; commercial partners investing in each other is common in fast-growing industries. The point Sona's map makes is structural: it becomes harder to assess any single security in isolation once its issuer's revenue, its counterparty's revenue and its financier's return all trace back to the same small pool of capital.

A separate Sona map, built on Bloomberg supplier data, measured how dependent individual companies are on the spending of one or two larger customers, and overlaid that with total debt and total-debt-to-EBITDA leverage. The published account identifies Oracle and CoreWeave as conspicuous higher-leverage nodes in this map, but it does not publish reproducible company-level figures, so readers should treat this as a directional finding rather than a scored league table.

Why diversified-looking portfolios can share one dependency

A UK institution can hold a chipmaker's shares, an AI laboratory's convertible notes, a cloud provider's investment-grade bonds, a neocloud operator's leveraged loans and an infrastructure fund's stake in a data-centre special-purpose vehicle: five different issuers, five different instruments, apparently well spread. Sona's dependency map and the Bank of England's own analysis both suggest that several of these positions can nonetheless rest on the same underlying event: a change in one hyperscaler's capital-expenditure plans, or a shift in demand from one or two AI laboratories.

The Bank of England's July 2026 Financial Stability Report set out how fast this concentration has grown across UK-relevant markets. The five hyperscalers it defines (Meta, Alphabet, Amazon, Microsoft and Oracle) represented just 3% of outstanding US investment-grade debt at 31 December 2025, but accounted for more than 15% of year-to-date US investment-grade issuance by early May 2026. AI issuers made up 41% of non-refinancing US high-yield issuance in 2026 to the Bank's July report, despite representing only 1% of the JP Morgan US high-yield index at the end of 2025.

The Bank also cited an OECD estimate that private credit's share of AI investment financing rose from 9% in 2024 to 34% in 2025, and noted that AI-related companies made up around half of the S&P 500 by market capitalisation in July 2026, up from around a quarter in 2022, relevant to UK portfolios that track global or US benchmarks even though the S&P 500 is not a UK index. Separately, the Bank's supervisory intelligence recorded roughly a 40% annual rise in global hedge-fund equity prime-brokerage balances by July 2026, alongside greater sector concentration including in semiconductors.

Forward-looking figures cited in the same report illustrate the scale still to come, not what has already happened. Barclays projected $240bn of hyperscaler investment needs financed through investment-grade credit issuance during 2026, a forecast the Bank compared with the $150bn-$170bn a year typically issued in senior debt by the six largest US banks. JP Morgan estimated more than $2tn of aggregate funding could be needed for AI chips over the following five years. Both figures carry substantial uncertainty and should not be read as realised totals.

Where the weaker links may sit

Not every part of the AI financing chain carries the same risk. Large, cash-generative hyperscalers sit at one end of the spectrum; neoclouds, data-centre developers and special-purpose vehicles financed through leveraged loans or off-balance-sheet structures illustrate the other. Sona's separate leverage overlay flagged Oracle and CoreWeave as conspicuous higher-leverage nodes, without publishing reproducible company-level figures; that finding should not be read as placing either company at either end of this spectrum, since Oracle is itself one of the five hyperscalers referenced earlier in this article. The Bank of England said off-balance-sheet financing can raise leverage at the asset level and make it harder to identify where risk ultimately sits, and that debt risk depends heavily on underwriting terms, in particular the quality of leases and guarantees standing behind a creditor's claim.

The Bank also pointed to a possible maturity mismatch: long-term debt financing data centres or chips whose economic lives may turn out to be shorter than the debt itself. It said evidence on the speed of AI-chip depreciation is mixed. Moody's, in a 31 July 2026 analysis, added a related and separate point: a completed data centre does not necessarily generate cash flow the moment it is built. Power connection, commissioning, tenant deployment and utilisation milestones all have to be met first, and delays at any of these stages can strain a project's finances before it earns a return.

Both the Bank and Moody's also record mitigants that cut the other way. Moody's noted that many large developments carry long-term leases with investment-grade hyperscale tenants, and that some recent financing structures allocate delay, cost-overrun, refinancing or technological-obsolescence risk to the stronger party in the contract. Moody's also observed that some current chip financings are structured to fully amortise within the term of a hyperscaler contract, matching debt tenor to the chip's useful life rather than leaving a refinancing gap. Risk, in other words, is structure-specific. A project backed by a strong tenant covenant and amortising debt is not the same as one backed by a thinner counterparty and a bullet repayment.

What UK exposure looks like, and what is not known

No source reviewed for this article quantifies the aggregate amount that UK asset managers, banks, insurers or pension schemes have invested, directly or indirectly, in the entities Sona mapped. That is a genuine information gap, not a finding that UK exposure is small. UK institutions can be exposed through global equity index trackers, investment-grade and high-yield bond funds, private-credit allocations, infrastructure funds, insurer asset portfolios and direct bank lending to AI-related borrowers, channels that the Bank of England's Financial Policy Committee has specifically linked to AI-ecosystem interconnectedness in its assessment of UK financial-stability risk.

The Committee judged that increasing interconnectedness in the AI ecosystem, combined with reliance on external financing, raises the risk that a shock to AI-related assets could spread more widely through the financial system. At the same time, it recorded that the outstanding stock of AI-company debt was still relatively modest at the start of 2026, which it said had contained the immediate financial-stability risk, even as it warned that potential risks were building rapidly. Both statements belong together: a system can be more interconnected and still be carrying a manageable amount of debt today.

For UK risk committees and trustees, the practical questions Sona's and the Bank's work raise are about look-through exposure rather than headline totals: whether apparently separate holdings share a common ultimate customer; how strong the guarantees and lease terms are behind any debt-financed AI infrastructure in a portfolio; whether debt maturities are matched to the expected life of the underlying chips or buildings; and whether a project has moved from construction into stable, contracted cash flow or is still exposed to commissioning and utilisation risk. None of this is a reason to expect a specific outcome for any fund or scheme, and FinTechPulse is not making a recommendation about any product. It is a description of the questions the available evidence suggests are worth asking.

What to watch next

The Financial Times account of Sona's work frames it as an examination of how a single point of failure might ripple through the AI financing network, not a prediction that one will occur. The Bank of England has said it will continue to monitor AI-related debt growth, private-market disclosure and credit quality as part of its financial-stability work; its next Financial Stability Report is the natural place to look for an update on how these exposures have developed. Readers wanting the primary detail behind the figures in this article can consult the Bank of England's Financial Stability Report and Financial Policy Committee record directly, both published on the Bank's website.