Joining the dots between big AI

Financial connections between tech firms are substantial. And — if we put aside challenges this presents to antitrust regulators/the foundations of functioning capitalism — they are not particularly worrying. At least, they are not particularly worrying in a world where firms don’t fail.

We live in a world where firms do sometimes fail. As such, the financial links connecting what look like nice safe firms in the AI ecosystem to potentially systemically important weaker links have the potential to matter quite a lot.

A new paper by Craig Nicol, Oleg Melentyev, and Charlie Callan of Sona Asset Management, best known for their massive London-based credit hedge fund, has a go at mapping the AI universe and cataloguing the interconnections between major players. Spoiler: they are many.

The authors kindly shared the data they’d mined with Alphaville. And we threw them into our dataviz systems to make this hot mess of a cat’s cradle radial network chart:

We appreciate that the chart isn’t the most mobile-friendly of things, but it gives a sense as to the extent of financial connections between different entities. Financial connections include equity, debt, lease commitments or whatever. Nodes are sized by the amount of commitments to which each is reportedly attached.

So the green OpenAI blob is big (lots of commitments pouring both in and out). But one of the tiny grey blobs representing Citadel is small (because we don’t actually have a specific number for their commitment, just reports that they were involved in a $2bn Series C equity capital raise for neocloud Nscale). Hover your pointer on a node to see where financing is coming from and where it’s going.

The data covers 176 deals totalling $3.6tn, transacted between 202 entities, in the three and a half years to August 2026. It looks like the true value of deals could be somewhat higher just because of the number of deals whose value is not disclosed.

OpenAI and Anthropic are both central to the ecosystem, and have each accumulated vast commitments. We’re not saying that they’re the weak links in this web, but their lack of positive (unadjusted) earnings may mean that they warrant a little more attention. And Sona’s data enables us to zoom in on them:

The two Sankey diagrams show various commitments made to the two monoline AI labs, and the commitments that they have, in turn, made to others. Hover your pointer over a leg of the chart and you’ll see details of each deal.

In most cases, money committed to a lab looks like it comes from a different source to which the money is then committed. But not always. As the authors note:

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.

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.

Moreover:

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.

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.

Zoom out and the AI financing web “increasingly reflects the characteristics of a closed-loop system”, they argue.

In fact, of the 176 transactions analysed, around 120 carry what the analysts call a ‘high circularity flag’, meaning that nodes are connected by multiple relationships, often with capital flowing in a loop.

And, like the risk that open models eat frontier labs’ lunch, the presence of what looks a lot like circular financing featured in S&P Global Ratings’ ‘known unknown’ risks to hyperscalers’ credit ratings only two weeks ago.

Are our charts really the best way to visualise the interconnectivity of the AI ecosystem? Sona has a banger of its own, but sadly only in static picture format. Behold:

Sona splits nodes into four separate orbits. The central orbit has the key AI players, memory and foundry majors, etc; the second ring shows other companies; the third investor syndicates; and the outer ring investors and financiers.

It’s not just the tangled web of financing that binds the fate of tech firms together. Using Bloomberg supplier data, Sona mapped the concentration and dependency of the AI supply chain by plotting revenue dependence of firms to one another.

Sure, they found that chip firms sell a lot of chips to chip buyers. But they also found that some firms’ revenue is almost existentially tied to the capex decisions of one or two players. Here’s a chart showing the top five companies’ revenue dependence on the ecoverse:

Having mapped revenue dependency and used this to configure the positioning of each company on a network chart so that it reflects each company’s connections to the key AI players (who they have fixed on the periphery), Sona overlays leverage — framed here as total debt to ebitda — to produce this:

Basically, the bigger the bubble the bigger the total debt, and the darker the red the greater the leverage. Most of the bubbles are fairly greenish, but it’s no surprise to see Oracle and CoreWeave are bright red.

Circular financing isn’t actually cheating. And it would be weird if chipmakers and neocloud companies didn’t find their revenues dependent on the business of hyperscalers and AI labs.

But the incestuous network of revenue-capex-financing connections does mean that if you want to understand the creditworthiness of one element of the system, you’re going to have to look at the system as a whole.

And that’s really the point of the work: to examine how single points of failure might impact the AI universe. Because we live in a world where firms do sometimes fail.

Further reading:
— Are credit rating agencies getting fed up with hyperscalers? (FTAV)
— How big is the open-model threat to AI hyperscalers? (FTAV)