Why we're not easily seduced by the fantastical AI growth story

It seems almost impossible to escape artificial intelligence (AI). Every day, a new application is described for this technological juggernaut, coupled with predictions of major change and disruption. But before AI starts writing our emails, doing our homework and planning our next holiday, it needs to be built!

By built, we mean the infrastructure that trains, powers and distributes AI. This requires massive, specialised warehouses called data centres consuming as much power as small cities, housing thousands upon thousands of specialised computer chips designed specifically for the task.

The scale and cost of the projects announced so far is staggering. The funding needs are beginning to exceed the ability of the major players to fund their growth ambitions internally. Indeed, the tech majors have tapped bond and loan markets for more than US$100bn so far this year, far exceeding any of the last 10 years issuance.

Figure 1 – Tech company borrowing trends

While the tech companies are expected to finance a large proportion of the build themselves, they will in many cases also be the primary tenant of these facilities, creating additional liabilities in the form of lease agreements. Private credit is expected to be involved, with estimates that it could support up to 25% of the potential US$2.9tn needed over the next three years.

Figure 2 – Projected global spending on data centres, 2025-28

Take a recent example that highlights the creativity being harnessed to finance these mega deals. The parties include Facebook-and-Instagram owner Meta (NASDAQ: META), private credit operator Blue Owl, and international asset managers PIMCO and Blackrock.

With the offer of a long-term lease from Meta, Blue Owl was able to commit US$3bn to a huge data centre called Hyperion, while also sponsoring a structure that facilitated a further US$27bn of debt financing provided by major asset managers. Meta will also contribute to the costs of construction, but this structure allows them to preserve capital for growth opportunities stemming from the infrastructure once it is built.

Figure 3 - Hyperion data centre funding structure

The market is finding its voice

While debt will be integral to the growth and spread of AI, excessive or wasteful debt can derail plans even in high growth sectors. Amongst the largest technology companies seeking to stake their claim, enterprise-software giant Oracle is the most indebted.

With an ambitious AI growth strategy and US$110bn of debt already outstanding, the market is already starting to ask how its agenda could be funded. Based on the cost to insure against default of Oracle debt, doubts are beginning to weigh on investors’ minds.

Compared to Oracle’s technology peers, the price of their credit default swaps has roughly tripled this year. Oracle spreads began to materially diverge following a seemingly bullish market update which indicated a sharply higher forward order book, which initially boosted the company’s share price. Upon further consideration, it became clear that massive investment was required to capture this opportunity.

Figure 4 – Technology company credit default swap spreads

Some banks are already exploring ways to hedge their risks in this sector. For example, the Financial Times reported on 5 November that Deutsche Bank was actively considering ways to manage their exposure to the capital-intensive sector:

“The German lender is looking at options including shorting a basket of AI-related stocks that would help mitigate downside risk by betting against companies in the sector. It is also considering buying default protection on some of the debt using derivatives through a transaction known as synthetic risk transfer (SRT).”

We tend to agree that risk profiles are changing rapidly. The capital to bring the AI revolution into reality is needed now, but the path to profitability and earning a return on this investment is still well into the future.

In this context, spare a thought for investors who bought long-dated bonds in the major technology companies a few years ago. Back then, you were lending money to companies with reasonable growth prospects and excellent cash generation. Now, their capital expenditure plans have grown materially which will draw down on available cash and see indebtedness rise further.

Figure 5 – Technology company cash-to-total assets

Australia not being left out

The action is not confined to the United States. The following is a brief checkdown of the latest developments locally:

• Data centre owner and developer NextDC (ASX: NXT) announced a memorandum of understanding with industry giant OpenAI, which will underwrite a new multi-billion-dollar AI campus developed and managed by NextDC.
• Industrial property giant Goodman Group (ASX: GMG) plans to repurpose existing sites that are advantageously positioned close to power infrastructure. One example includes a former Coles distribution facility that will be converted to two data centres whose end value will make them among the largest in the country.
• Hyperscale developer AirTrunk is looking to progress major projects in the western suburbs of Sydney, but is struggling to secure confirmed power and water allocations in an already constrained market.

Amidst the excitement, not everyone is convinced

While it has been easy to get caught up in the huge numbers and even greater potential of AI to revolutionise work and society, some are now taking a more sober approach to the financial implications of all this investment.

One of these clear-eyed sceptics is IBM CEO Arvind Krishna. On a recent podcast, he spent some time laying out his doubts on the sector.

The first concern centres around profitability. Based on an expectation of ultimate total power demand and the existing per gigawatt cost to build, Krishna sees no way that firms such as OpenAI can ever turn a meaningful profit. He uses an assumption of US$50-80bn per gigawatt, and up to 100 gigawatts of construction for a total cost of US$5-8tn. This would require US$500-800bn per year in profit to compensate the investment.

Another doubt lays in the sheer volume of power required. Bottlenecks are already being felt in major markets globally, including the US. One way to think about the shortfall is that the US alone requires the equivalent of 48-50 new grid-scale nuclear power plants just to cater for expected data centre demand.

Figure 6 – US data centre power demand

Finally, he expressed doubt that even if the trillions currently slated does materialise, it might still not be enough to make any meaningful progress toward the goal of artificial general intelligence (AGI). While considerable productivity benefits are expected to come from current or proposed investment activity, AGI proposes an even greater step change in capability, where AGI can complete complex tasks better than humans and with little to no input. In Krishna’s mind, considerable further research is required, and estimates the progress toward AGI is somewhere between zero and one percent.

Where to from here?

This fast-moving and fascinating topic raises several questions for us as credit professionals and potential investors in this space:

1. Can credit markets absorb such a large increase in supply? Yes, we are confident they can, but it might require some adjustment in asset allocation and pricing. The aggregate amount of corporate bonds outstanding grows on average in the low single digits each year. If all proposed AI capital expenditure plans are realised over the next three years, the average growth rate could quadruple through 2028. We expect inflows to remain robust, supported by all-in coupon rates that are attractive on a risk-adjusted basis.
2. Have there been any impacts already? Most likely, but it is difficult to isolate the impact of a single sector within a diverse market such as global credit. To date, the hyperscaler technology companies have been able to fund their growth from cashflow and with the help of equity investors. However, we believe that some of the strong inflows into credit this year have been in response to the expectation that debt will play a growing role in the expansion of the AI ecosystem. While the growing funding needs have been well telegraphed, we can’t discount the possibility of a crowding out effect emerging in 2026.
3. Would Daintree ever invest in data centre or AI debt? Possibly. There are many factors to consider. We would be more attracted to a deal that involves physical security such as the land or building of the data centre itself, and preferably an operating stabilised asset with high quality tenant/s. While still a concentrated asset rather than a diverse range of collateral, the business model is easy enough to understand. When it comes to the providers of equipment inside the data centre or the firms leveraging the equipment to build models or applications, Daintree would be much less likely to invest here because of the sheer capital intensity, frequent replacement cycles and lack of profitability. Daintree is not as easily seduced by fantastical growth stories!

Managed Fund

Daintree High Income Trust

Australian Fixed Income

View

Managed Fund

Daintree Core Income Trust

Australian Fixed Income

View

ETF

Daintree Core Income Active ETF (DCOR)

Australian Fixed Income

View