AI data centres: are the
returns stacking up?

Demand keeps accelerating

The three large cloud businesses all grew faster in the June quarter than in the March quarter. Google Cloud revenue rose 82% year on year, Microsoft Azure 43% and Amazon Web Services 37% – the latter its fastest growth rate in more than four years. Simultaneous acceleration across three competing businesses is difficult to square with the idea that companies are building capacity nobody wants. Between them, they also disclosed more than US$1.5 trillion in contracted future revenue.

Growth alone doesn’t tell you whether the capital is earning a return. For that, the structure of the spending matters. Importantly, the facility and the computing equipment within it have very different economic lives. The building, power connection and cooling infrastructure can last around 30 years and are paid for once, while the computing equipment typically lasts five or six years and is replaced repeatedly. Amazon management has indicated that this equipment pays for itself in under three years.

Because the same facility can support several generations of computing equipment, utilisation across successive refresh cycles can materially improve its lifetime economics. On our own calculations, the blended return over the full life of a facility sits comfortably above the cost of capital – provided the data centres remain occupied.

A note of caution

There is, however, an important caveat in the reported numbers. The assumed useful life of these assets is a management judgement, and those assumptions have consistently been revised in ways that flatter near-term profits. During the quarter, for example, Microsoft extended the projected life of its buildings from 15 to 25 years. Longer assumed lives reduce annual depreciation charges and increase reported margins. We don’t believe this undermines the returns we’re seeing, but it’s something worth watching.

For now, the evidence supports continued capacity expansion beyond 2027, with demand still exceeding available supply. The next phase, however, will be more demanding – and the risks are rising.

The funding gap widens

Amazon’s free cash flow over the past 12 months has turned negative. Alphabet’s fell below zero during the quarter for the first time in its history as a listed company, while Microsoft’s fell by roughly a quarter.

None of these companies is in financial difficulty, but they’re no longer funding the build-out entirely from their own cash flow. Alphabet raised equity in June, while Meta has begun placing individual data centres into partnerships with outside investors to keep them off its balance sheet. This is the funding gap we wrote about a year ago, and a sign that the cycle has moved beyond its more comfortable phase.

As we wrote in October last year, chip suppliers have continued to invest in, and in some cases underwrite, the customers buying their products. There is nothing improper about these arrangements, and the underlying orders are real. However, they do mean that the same dollar can be counted as revenue at more than one point in the chain – and that a disappointment at one participant could travel further than it otherwise would.

Concentration becomes a risk

One of the quarter’s most interesting disclosures was almost buried. Microsoft’s contracted future revenue grew 84%, but only 25% excluding OpenAI. One customer therefore accounts for roughly a third of the backlog that investors have been treating as evidence that the spending is safe.

Microsoft’s exposure is unusually concentrated, while demand across the other hyperscalers is broader. Meta’s spending is largely for its own advertising business, while Amazon’s and Google’s backlogs are spread across thousands of corporate customers. However, the fastest-growing portion of demand can be traced to a small number of AI laboratories, principally OpenAI and Anthropic, and their economics deserve closer attention.

OpenAI’s annualised revenue run rate has reportedly crossed US$40 billion, but it is still expected to report losses in the tens of billions this year. Positive cash flow is also unlikely before the end of the decade, given infrastructure spending commitments running into the hundreds of billions.

Anthropic is the stronger business today. Its annualised revenue run rate reached around US$65 billion in July, having increased roughly fortyfold in 18 months. Most of its revenue comes from corporate rather than consumer customers, which generates better margins and enabled the company to report its first operating profit in the second quarter. Anthropic is reportedly targeting a listing in the coming months to raise further capital, while OpenAI is leaning towards 2027.

The China challenge

The competitive threat to both comes from an unexpected direction. Chinese laboratories – including Alibaba, DeepSeek, Moonshot and Zhipu – now publish models that can be downloaded and run without licence fees, with capabilities not far behind the American frontier. On the widely followed developer platform OpenRouter, Chinese models have risen from less than 2% of usage in early 2025 to roughly half today, reflecting their lower cost for similar tasks.

In our view, Chinese models pose a greater threat to the economics of model developers than to demand for infrastructure. Western companies can run these models in their own environments or through Western cloud providers. Lower model costs should also encourage greater usage, supporting demand for the chips, networks and data centres needed to run them.

Investment implications

We’re past the point where returns on this spending were unmeasurable, but well short of the point where they’re proven. Demand, pricing and margins across the three large cloud businesses are holding up better than sceptics expected. While we remain cognisant of the risks that are building, the evidence currently points to the cycle extending for longer, with demand still comfortably outstripping supply.

We continue to have exposure to various parts of the semiconductor supply chain, as well as to the US hyperscalers. Within semiconductors, we have rotated some of our holdings in memory and semiconductor capital equipment, which have performed very strongly over the past year, towards chip designers Nvidia and Broadcom.

Nvidia remains the leading supplier of accelerated computing systems, while Broadcom benefits from the shift towards custom accelerators and the networking infrastructure needed to connect large AI clusters. We believe the risks to their longer-term earnings power are now more fully reflected in their share prices than they were a year ago.