Foreword
AI is a once-in-a-generation technological super-cycle that is transforming the way humans live, companies operate and economies evolve.
The winners and losers from revolutionary technology cycles can take many years to emerge. The growth of ecommerce in the late 1990s was expected to spell the end for bricks and mortar retailers. Today, Walmart is a near trillion-dollar company. After the dot-com bubble burst, many e-commerce businesses went under. Today, Amazon is a near $3tn company. The AI cycle will most likely be no different.
Where it will be different is its financing needs, which will dwarf virtually all prior cycles. Over the period from 2026 through 2030, we estimate AI-related financing needs to be around $7tn. If this forecast holds true, in relative terms it will mark the largest investment cycle since the construction of the railroads in the 1800s.
GPU financing, neoclouds and other parts of the AI ecosystem are evolving fast and require significant capital to keep pace. Power demand is huge and requires massive investment in infrastructure. These financing requirements fascinate us, in no small part because much of it will be debt.
The opportunities arising from the AI buildout are as great as the risks posed by lax underwriting. While we are hesitant to compare the debt funding of the AI super-cycle with prior super-cycles, we believe it is important nonetheless to be aware of financing and other risks emerging today.
This series of three papers focuses on three topics. The first topic, explored in this paper, examines the complexities of the AI universe and the many linkages between its constituent operators. The second paper explores the impact that massive hyperscaler debt funding has had on the structure of the US investment grade market, and our argument that this is the most significant structural change to modern-day credit markets. The third paper posits the theory that Europe’s AI moment is still to come, and why it rests in adopting AI rather than developing it.
We hope you enjoy reading.
Craig, Oleg and Charlie.
Executive Summary
The AI industry has grown increasingly interconnected. In this paper, we map it two ways: first, as a supply chain based on the real-economy flow of goods and services between the public companies that are central to the AI buildout. Second, as a capital and commitment web, based on the flow of equity, debt and purchase commitments that fund the leading public and private companies in the AI buildout.
Neither map is exhaustive and neither is a valuation judgement. Instead, they are designed to illustrate the burgeoning complexity of the AI universe and three features we think matter for investors:
- The connections are highly concentrated, with genuine single points of failure.
- Funding has become increasingly reflective of a closed-loop system.
- While the largest players in the universe are financially robust, they are inherently connected to the capacity-builders that are characterised by leverage and negative cash flows.
Should credit investors be concerned? Yes and no. There is, for example, some mirroring of the mortgage market of 2008: opaque and concentrated exposures, counterparties linked in complex ways that few have mapped, and demand partly underwritten by the same balance sheets that depend on it.
The difference is that the AI buildout features largely cash-generative businesses that are funding real, physical capacity, not synthetic leverage stacked on a deteriorating asset.
That said, whenever there is a flood of money chasing specific sectoral exposure, it is time to be more judicious with your underwriting principles. The central question for a credit investor now is the same one that mattered in 2008: how much of this demand is being underwritten by the same handful of balance sheets, and where are the weak links if the cycle turns?
Key Takeaways
- This is a large and fast-growing universe. Our supply-chain map spans 255 listed companies partaking in the AI buildout worth a combined $50tn in market capitalisation. Five years ago, this market cap figure was less than half, at $23.5tn.
- The 255 mapped companies have $5.7tn of gross debt. Some 76 of these companies have large debt stacks in excess of $10bn. Some 28% of the 255 companies mapped are index-eligible bond issuers, with $2.1tn of notional outstanding in the Global Agg or High Yield indices.
- The 100 most connected public companies – the “core” of our AI universe – together carry $2.3tn of gross debt against $1.8tn of EBITDA. Some $1.3tn of this debt is in index-eligible bond form, across 45 issuers.
- Genuine chokepoints. We estimate that trade within the core of our AI universe of public companies ran at ~$230bn in 1Q26, which, on an annualised basis, equates to ~2.9% of US nominal GDP. So much spending is retained within a small group that single points of failure and tail risk events could carry systemic-like consequences. One potential point of failure is the few near-irreplaceable suppliers in the EUV lithography chain, all of which are Europe-based.
- Equity book diplomacy. Strategic minority stakes have become an instrument of influence. NVIDIA has assembled by far the widest book: roughly two dozen positions spanning its customers (OpenAI, Anthropic, CoreWeave), its suppliers (SK Hynix, Marvell) and even a direct competitor (Intel). The hyperscalers deploy the same tool more narrowly, taking stakes in the labs to tie AI demand to their clouds.
- Related-party complexity. Capital, product and demand chase each other around the same handful of names. To trace one thread: Microsoft invests in OpenAI (~$135bn ownership stake via a recapitalisation); OpenAI commits ~$250bn to purchase Microsoft compute; much of that compute runs on NVIDIA silicon; NVIDIA, in turn, invests in the neoclouds (i.e. CoreWeave and IREN) which lease capacity back to Microsoft and OpenAI; NVIDIA also takes a progressive equity stake of up to ~$100bn directly in OpenAI and was in talks to guarantee a $250bn financing for an OpenAI data centre, plus another $100bn financing for OpenAI to purchase NVIDIA chips (by August 2026, it had entered residual-value guarantees of roughly $105bn on leases for about 4.25GW of capacity where OpenAI is the tenant, and had signed MOUs with six large asset managers to mobilise more than $500bn of third-party compute financing); and OpenAI itself takes equity and warrant positions, in CoreWeave and AMD, to lock in chip and compute supply.
The AI financing and infrastructure demand loop
Figure 1: Illustrative AI financing and demand relationships
Source: Sona Asset Management
- The model layer attracts the capital but the balance sheet risk settles into the infrastructure layer. For a credit investor, this means the default risk sits one ring out from the core. The hyperscalers and semis/memory manufacturers generate enormous operating cash flows and have deep financial flexibility. The periphery is where strain is most noticeable. Neocloud and data centre platforms carry the highest leverage, the thinnest margins and the weakest cash flows. They are also susceptible to hyperscaler vertical integration and the possibility of advances in lower-cost orbital compute. If the cycle turns, vulnerability could emerge here given they have fewer financial levers available to them. A frenzy for exposure is making the space ripe for adverse credit selection.
2 Source: https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/
5 Source: https://thenextweb.com/news/nvidia-openai-ports-pike-105bn-residual-value-guarantee
7 Source: https://openai.com/index/openai-amd-strategic-partnership/
Mapping the Universe
We map the companies in the AI universe in two ways: first, as a supply chain highlighting flows of goods and services; second, as a transaction chain showing capital and commitment flows. We do not claim this to be an exhaustive universe, more an illustration of the complexity and interconnectedness of the burgeoning industry.
Figure 2 is our zoomed-out view of the supply-chain universe. In plain terms:
- It maps 255 listed companies: 15 key players we have selected (the large bubbles) plus their top 20 customers and top 20 suppliers.
- The web is naturally chaotic. Its purpose is to show the density of connections, particularly at the core.
- The arrowheads show the direction of product flow, from supplier to customer.
- The larger the company bubble, the larger its log-adjusted market capitalisation.
- The sparse outer ring consists of companies with a single link to the core.
- The dense core is made up of the ‘many-to-many’ companies: those tied to several key players at once. This is where we focus next (Figure 3).
8 Methodology: We use Bloomberg data, taking the top 20 customers and top 20 suppliers for 15 key AI players selected by us. The 255 nodes are well below the mathematical maximum of 615 because (i) some key players have fewer than 20 customers or suppliers in the data, and (ii) many of those customers and suppliers connect to more than one key player. This map covers listed companies only, so some key private names (e.g. Anthropic and OpenAI) are absent; the transaction-chain map that follows does include private companies.
Supply Chain View: The Core
Figure 3 remains dense, with 85 interconnected companies sitting inside the ring of the 15 key players. Note the spacing: we have given the key players fixed spots, but the other companies are positioned based on relationships.
For example, TD SYNNEX sits in the middle of the core because it buys from multiple key AI players (which makes sense given it is an IT distributor). Meanwhile, Entegris sits in the outer ring of the core, between TSMC and ASML, because it only sells to these two key players (which is again logical given it provides niche contamination control materials to semiconductor manufacturers).
How Relationships Affect Positioning in the Core
Figure 4: LHS: TD SYNNEX is pushed to the centre given it buys from 11 of the 15 key AI players.
Figure 4: RHS: Entegris sits between TSMC and ASML given it only sells to these key AI players
Source: Data from Bloomberg; mapping by Sona Asset Management.
We can assess the concentration and dependency of the AI supply chain by using the core diagram shown in Figure 3, which features only the ‘many-to-many’ names: companies that have at least two connections to our 15 key players. The total USD value of trade between these 85+15 companies is worth $230bn per quarter (1Q26) – which is equivalent to ~2.9% of US nominal GDP when annualised.
Figure 5 breaks down this trade value into revenue dependence. A bigger bubble represents a higher reliance on revenue from the AI supply chain universe as a proportion of overall revenue.
Revenue Dependence in the AI Supply Chain Universe
Figure 5: Company-Specific Customer Reliance as % of Total Revenue. Largest bubble is Applied Digital at 86%.
Source: Data from Bloomberg; mapping by Sona Asset Management
Many of these bubbles are large but they conceal two different profiles. For the chip manufacturers (i.e. semis and memory), revenue reliance is high but not alarmingly so given a long tail of customers. For example, SK Hynix’s largest customer (NVIDIA) accounts for ~27% of its revenue; NVIDIA’s largest customer (Microsoft) accounts for ~22% of its revenue, and TSMC’s largest customer (NVIDIA) accounts for ~19% of its revenue.
The concern is with the smaller capacity providers, whose reliance is not only high but concentrated in one or two names. For one, CoreWeave draws ~67% of its revenue from Microsoft alone. Applied Digital makes ~56% of its revenue from Oracle, plus a further ~30% from CoreWeave – itself a Microsoft-dependent business (see Table 1 below). These are the names for which a single counterparty’s capex decision may be existential. Further, as the credit metrics show, they are also the most indebted and most cash-flow negative. Revenue concentration and balance-sheet fragility therefore compound in the same handful of names (see further below for our transaction chain, which draws this point out).
We interpret these figures cautiously. Nonetheless, they show that a large proportion of spending takes place inside the core of the supply chain universe. This raises the risk of systemic consequences from any single point of failure, such as EUV lithography, or a tail risk event, such as disruption to Taiwanese trade.
9 Source: Sona Analysis via Bloomberg
10 Note: these data are based on the most recent disclosures via Bloomberg, hence these figures are influenced by the timing of trade.
Top 5 Demand-Side Reliance – Share of a Company’s Revenue Dependent on our AI Universe
| Company | Universe as % of Revenue | Type | Sells to11: |
|---|---|---|---|
| Applied Digital | 86% | Shared | ORCL (56%), CRWV (30%) |
| SK Hynix | 84% | Hub | NVDA (27%), MSFT (9%), AAPL (8%), GOOGL (7%), AMZN (7%) |
| CoreWeave | 82% | Hub | MSFT (67%), META (8%), GOOGL (5%) |
| NVIDIA | 73% | Hub | MSFT (22%), AMZN (14%), SMCI (10%), GOOGL (5%) |
| TSMC | 66% | Hub | NVDA (19%), AAPL (17%), AMD (5%), QCOM (5%) |
Source: Bloomberg; via the most recent quarterly estimate or disclosure.
Figure 6 examines the leverage embedded within our core universe. Whilst there are some flashes of red, most nodes are green, representing low leverage levels. The 100 core companies have approximately $2.3tn in total debt and $1.8tn in total EBITDA, resulting in a universe leverage multiple of 1.3x, as at Aug-2026.
Stripping telecoms out of the figures, the leverage multiple falls to 1.0x. Using net debt drives this down to a mere 0.1x, mainly because the 15 key players in our outer ring have a combined net cash position. In other words, the core universe isn’t highly leveraged taken as a whole. However, this masks the fact that individual companies like ORCL, CRWV, INTC and HPE have large debt stacks and high leverage levels – and are important links in the AI supply chain.
Where the Leverage Lies (bigger bubble = more total debt; more red = more leverage)
Figure 6: Many-to-Many Connections by Total Debt (Logarithmic Bubble) and Leverage (Inverse Colour)
Source: Data from Bloomberg; mapping by Sona Asset Management
11 Percentage given as a respective proportion of Company revenue.
12 Source: Sona Analysis via Bloomberg
When considering the representation of our core universe in the Global Aggregate Corporate Index and the Global High Yield Index, there is $1.3tn of par value outstanding as of August 2026. Some 99% of these bonds are situated within the investment grade universe, and in total are worth a material 8.9% of the index.
It is worth noting that the 15 key players are, in aggregate, well-capitalised. They hold a combined net-debt position of roughly -$90bn (i.e. net cash) and generated ~$450bn of free cash flow (LTM; based on most recent data as of August 2026) between them. So, the systemic leverage risk within this universe does not sit with the 15 key players. It sits with the critical capacity providers building at the start of the value chain and will most severely impact the smaller leveraged suppliers.
The starkest divergence in the core universe is not in leverage but in capital expenditure and free cash flow. The 100 companies within the core registered ~$840bn of capex in the last reported year, of which the 15 key players collectively accounted for ~$620bn, or 21% of their combined revenue. The spending is heavily top-loaded: Amazon (~$132bn), Microsoft (~$116bn), Alphabet (~$91bn), Meta (~$70bn), and Oracle (~$56bn) make up the bulk of it, and the AI buildout is now consuming an outsized share of sales: 35% at Meta, 33% at TSMC and 83% at Oracle.
For most of the mega-caps, this capex remains comfortably self-funded: Apple, NVIDIA, Alphabet and Microsoft each generate $65–100bn of annual free cash flow. One layer down, this inverts: Oracle is now spending well ahead of its cash generation (free cash flow of roughly -$24bn). Another layer down, among the neocloud and data-centre names, capex is far outstripping cash flow. For example, CoreWeave’s capex ran at roughly twice its revenue last year, on net leverage above 8x. It is these names, not the key players, that are funding the physical buildout externally, which is where our capital and commitment map picks up the story.
Capital & Commitment Flows
Whereas our supply chain universe maps the flow of goods and services between companies, our transaction chain maps the flow of capital and commitment. It shows us which companies are financing each other in what increasingly reflects the characteristics of a closed-loop system.
Capital & Commitments: Inflows across the AI Universe
Figure 7: AI Ownership and Financing Web – Logarithmic Bubble Represents Capital / Commitments Inflow
Source: Company announcements, Bloomberg, Reuters, FT, WSJ, Forbes, SEC, ACCC, TechCrunch
We scraped the web for company announcements and other reputable source material specifying the transactions that have occurred over the three and a half years to 31 August 2026 in the AI sphere. This exercise was not exhaustive, but it was substantial: we found 176 unique transactions across 202 entities totalling $3.6tn in announced quantum.
We used the data to create four orbits. The inner ring represents 21 key players: our original key AI players, the memory and foundry majors (Samsung, SK Hynix, Micron, TSMC and Broadcom) and several key private companies. The second ring represents all other companies. The third ring shows investor syndicates. The fourth ring is investors and financiers. Arrowheads point the way capital is committed, from investor to target, and from customer to provider.
First, we highlight the density and concentration. Figures 7-8 show that capital and commitments are pooling into a small number of companies: the AI model/app companies (top right) and the hyperscalers (bottom right). Our mapping suggests that more than $250bn of announced non-conditional equity capital is directed towards the AI model/app layer alone. A parallel concentration appears on the commitment side, where a handful of multi-year compute agreements dwarf everything else. The largest of these includes OpenAI’s reported commitments to three hyperscalers: $300bn to Oracle, $250bn to Microsoft and $100bn to Amazon.
The second highlight is circularity. Of the mapped transactions, over 120 carry a high circularity flag, meaning the counterparties are connected by more than one relationship, often with capital flowing in a loop. One such loop from the AI model layer: NVIDIA has conditionally committed to invest up to ~$100bn in OpenAI (2025) with a further ~$30bn indicated in 2026; OpenAI in turn commits hundreds of billions of dollars to Oracle, Microsoft and Amazon for cloud and compute; these companies are among the largest buyers of NVIDIA silicon: the circle is complete.
14 Source: Sona Analysis
Capital & Commitments: Outflows across the AI Universe
Figure 8: AI Ownership and Financing Web – Logarithmic Bubble Represents Capital / Commitments Outflow
Source: Company announcements, Bloomberg, Reuters, FT, WSJ, Forbes, SEC, ACCC, TechCrunch
Capital that begins as a chip vendor’s equity investment can therefore reappear as its customer’s revenue and as its own future demand. The pattern repeats itself: Amazon has invested in OpenAI (~$50bn) while also securing a ~$100bn compute commitment from it; Microsoft’s ~$135bn ownership recapitalisation of OpenAI sits alongside OpenAI’s ~$250bn purchase commitment back to Microsoft Azure. NVIDIA holds a web of smaller strategic stakes across its own supply and customer bases, with CoreWeave, Intel, Marvell, Nokia and IREN among them.
15 Source: https://openai.com/index/five-new-stargate-sites/
16 Source: https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/
17 Source: https://openai.com/index/amazon-partnership/
18 Source: https://openai.com/index/scaling-ai-for-everyone/
19 Source: https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/
Vendor and strategic financing of this kind can flatter the apparent strength of end-demand: a single committed dollar may be recognised as investment by one party, pipeline or revenue by another, and capital expenditure by a third as it travels the loop. This is not evidence of anything improper – strategic investment alongside a supply relationship is nothing new or extraordinary – but the density of these loops in our AI universe suggests that fortunes are concentrated in a small number of outcomes.
Much as the supply chain section highlighted single points of failure (e.g. EUV lithography) and tail risks (e.g. Taiwan disruption), the transaction chain highlights another fragility: should one key AI model/app company reset its spending plans, the effects would be felt deeply by the compute providers and the semiconductor and memory manufacturers.
For the credit investor, a consequential shift is in how the AI buildout is being financed: the early phase of the cycle was predominantly equity; our mapping shows the balance moving towards debt and contractual commitment. Broadly speaking, the companies absorbing the most capital and the largest forward commitments – the neoclouds and data-centre platforms – also carry (per Figure 6) the highest leverage and revenue concentrations, thinnest margins and weakest free cash flow metrics. The equity story and the credit story therefore diverge: the model layer attracts the capital and the headlines, while the balance-sheet risk settles into the infrastructure layer.
Our two mapping analyses tell a consistent story. The supply chain map shows an industry whose real-economy connections are concentrated and seemingly mutually dependent; the transaction map shows that its financing has become concentrated and, increasingly, self-referential. Neither map is a valuation judgement, and both are illustrative rather than complete. But they frame the central question we are asking: not whether AI demand is real, but how much of the demand is being underwritten by the same set of balance sheets, and where, if the cycle turns, the true vulnerability sits.
Conclusion
Our analysis of interconnections in the AI universe reminds us of an exercise conducted in the wake of the global financial crisis that sought to map the shadow banking system. It showed that exposures were concentrated and opaque, few had mapped counterparty links, and demand was partly underwritten by the same balance sheets depending on it.
Where the analogy differs is the collateral. AI is an intangible product, but the buildout is physical and the financing sits against real capacity rather than synthetic leverage stacked on a deteriorating asset. Further, credit metrics remain strong for the key players in our AI universe. This gives us some comfort that failure to meet return on invested capital targets is a residual value issue for equity, not an impairment problem for credit.
However, with ever more capital deployed in the AI universe, the margin for error is getting thinner. Indeed, with increasing amounts of off-balance sheet financing and hard-to-track contingent liabilities, the true extent of leverage and circularity becomes harder to observe.
The impairment risk for credit lies in (1) single points of failure; (2) geopolitical tail risk; and (3) as our analysis highlights, leverage amassing in the infrastructure layer of AI, where the quality of underwriting is now increasingly varied and a flood of capital is incrementally pushing the bar lower.
In our view, it is time to be more cautious on credit selection and appreciate the asymmetry of the asset class.
In our next report, we examine the market structure beneath the AI buildout, how $7tn in projected AI capex gets funded, whether the US credit market can absorb it, and who gets rationed if it cannot.
Disclaimer
Sona Asset Management (UK) LLP is a limited liability partnership formed in England on 17 June 2016 with company number OC412361. It is authorised and regulated by the Financial Conduct Authority in the United Kingdom (FRN: 814191). Sona Asset Management (US) LLC is a limited liability company formed in the State of Delaware on 9 December 2016 with file number 6244242. References to “we”, “us”, and “Sona” refer to Sona Asset Management (UK) LLP and/or Sona Asset Management (US) LLC.
The Information in this document (the “Information”) is confidential, proprietary, and intended only for the recipient. It has been developed internally and/or obtained from sources believed to be reliable, but Sona does not guarantee the accuracy, adequacy, or completeness of such sources. The Information, including statements concerning financial market trends, projections or other forward-looking statements regarding future events, targets, forecasts or expectations regarding the strategies described herein, is based on current market conditions, which will fluctuate and may be superseded by subsequent market events or for other reasons without notice. There is no assurance that any events will occur or targets will be achieved, and actual outcomes may be significantly different from those shown here.
The Information does not represent valuation judgments with respect to any financial instrument, issuer, security, or sector that may be described or referenced herein. It is not intended to, and does not, relate specifically to any investment strategy or product that Sona offers. It is being provided merely as a framework to assist the recipient’s own analysis and/or views on the topic discussed herein. It has been prepared solely for informational purposes. Any references to a target portfolio or allocations of such a portfolio refer to a hypothetical allocation of assets and not an actual portfolio. Any allocations contained herein are subject to change. There is no assurance that they will be achieved, and actual allocations may be significantly different.
Charts and graphs provided herein are for illustrative purposes only. Nothing contained herein constitutes investment, legal, tax, or other advice, nor is it to be relied on in making an investment or other decision.
The Information is not research and should not be treated as research.
The views expressed in this document:
- reflect the current views of Sona as of the date hereof, and Sona does not undertake to advise you of any changes;
- may not be reflected in the strategies and products that Sona offers or invests in, including strategies and products for which Sona provides investment advice;
- may be inconsistent with any future investment recommendations, techniques, methods, analysis, positions (long or short) or securities transactions Sona makes when managing client or proprietary accounts; and
- do not necessarily reflect the views of any individual investment professional at Sona.
There can be no assurance that an investment strategy will be successful. Historic market trends are not reliable indicators of actual future market behaviour or future performance of any particular investment which may differ materially and should not be relied upon as such.
The Information should not be viewed as a current or past recommendation or a solicitation of an offer to buy or sell any securities or to adopt any investment strategy.
The investment strategy and/or themes discussed herein may be unsuitable for investors depending on their specific investment objectives and financial situation. Please note that changes in the rate of exchange of a currency may affect the value, price or income of an investment adversely. Sona may not and does not assume any duty to update forward-looking statements. No representation or warranty, express or implied, is made or given by or on behalf of Sona or any other person as to the accuracy and completeness or fairness of the Information contained in this publication and no responsibility or liability is accepted for any such Information. By accepting this document, the recipient acknowledges its understanding.