Brute Force vs. Efficiency: Two Bets in the AI Data Centre Boom

The four large US hyperscalers, Amazon, Alphabet, Meta and Microsoft, have committed roughly USD 700 billion of capital expenditure for 2026 alone. These are not statements of intent. They are commitments already entered into, and they have triggered GPU orders, land purchases and power purchase agreements.

One figure alongside makes it interesting. Capex for these houses is growing by 60 to 80 per cent year on year, while their group revenues grow in the mid teens. From an infrastructure perspective that gap is the real question: is it build-ahead that demand later validates, or something else? Past capex cycles caution against simply assuming the former.

Where the money goes

A large share of the spending now flows into inference, the operation of fully trained models, as distinct from training, which creates them in the first place. The reason is mundane. Training is getting cheaper, and the real money sits with the end user.

Here a distinction matters for the location question. Part of inference is latency-sensitive: the answer has to arrive with practically no delay for an application to feel fluid. This load wants to sit physically close to the user and drives a distributed build-out near the end consumer. Another part, compute-heavy reasoning inference, tolerates more delay and runs in a few large, central facilities. Where a given capacity sits in this chain shapes its risk profile more than the question of whether AI grows in aggregate.

Two national bets

The most revealing contrast runs between the US and China. The US bases its build-out on brute-force scaling. China proceeds far more selectively: across 2023 to 2025, the combined capex of the Chinese hyperscalers came to roughly a sixth of what their US counterparts spent.

Behind the restraint sits a bet on efficiency. By late 2025, China's leading model was reaching around 90 per cent of the performance of the best US model on standard benchmarks, at a fraction of the investment. If comparable capability becomes attainable with substantially less capital, the assumption that ever-higher compute keeps paying off starts to wobble. A good part of the current capex plans rests on that assumption. China's selectivity is not a reluctance to build. The pull-back is specific: it is aimed at the cost of raw compute, not at the land, power and buildings beneath it, which keep expanding. But the scale of that expansion, and more importantly its manner, are opaque from the outside.

How much is commercially financed, and how much is state-directed or subsidised, is largely a matter of assumption rather than disclosure. That alone makes it hard to tell demand-led build from politically led build.

What this leaves open

None of this answers whether the boom is mispriced. It sharpens the question. From an infrastructure standpoint the question is never whether the sector is attractive, but which asset, at what price, and what could strand it before it has earned out or earned enough.

The usual stranding factors, technical obsolescence, demand risk, financing structure, are well known and more or less priced. The capex-to-revenue gap is by now familiar as well; the market has started to reprice it.

What is priced less well is the commitment embedded in the build-out itself. The current capex wave is a bet on one technology paradigm, that scale keeps winning. A parallel path built on efficiency is developing in plain sight, though, as noted, its foundations are only partly transparent. Neither side is clean: one commits vast, openly financed capital to a single paradigm, the other claims more with less on a basis that is harder to verify.

Which prevails is an open question, and the honest answer is that no one knows yet. What an investor can do is ask what a given asset is worth under either outcome.

I have set out how I think through that question in a separate confidential memo. If it is relevant to how you underwrite digital infrastructure, write to me at claudio@eigenadv.com.

Important Information

Eigen Advisory AG provides advisory services exclusively to qualified investors as defined under Article 10 (3) of the Swiss Collective Investment Schemes Act (CISA) and to professional and institutional clients as defined under article 4 (3) and (4) of the Swiss Financial Services Act (FinSA). Eigen Advisory AG is not licensed under CISA nor as a financial institution under the Swiss Financial Institutions Act (FinIA). Eigen Advisory AG's client advisers are registered in the register of advisers. Services are provided on an advisory basis only. Eigen Advisory does not exercise investment discretion, does not hold client assets, and does not manage collective investment schemes. Investment decisions and capital commitments remain at all times with our clients and their fiduciary and/or any other advisors. The information on this website is provided for general informational purposes only and does not constitute investment advice, an offer, a solicitation, or a recommendation to buy or sell any security or financial instrument. It does not consider the specific investment objectives, financial situation, or particular needs of any individual or institution. Past performance is not indicative of future results.