02The strategic landscape

What the headlines look like

Pulled from the public record over roughly eighteen months, from the mouths of the vendors building the technology and the research desks of the banks and firms watching them. Teal cards push towards investment. Gold cards push the other way. Both sets are published and attributed. Both sets are addressed to the same executive readers, often in the same week.

Reading these headlines in sequence, a reasonable executive reaches two conclusions at once: that AI is the defining investment opportunity of the decade, and that it may also be the defining capital misallocation of the decade. Both conclusions have support in the public record. The question is what to do with that fact.

The decision no one wants to make

Ask a chief executive in 2026 whether their company should spend money on generative AI, and the answer is almost never a clean yes or no. It is usually a sentence that begins with "we are looking at it," followed by a list of concerns that do not quite fit together. In this environment, allocating serious capital to a multi-year AI programme feels less like a strategic choice and more like a bet in a casino where the rules of the game change between hands.

The common diagnosis of this hesitation is that executives do not understand the technology, or do not care, or are afraid of being disrupted. In my experience of advising senior leaders through several technology cycles, none of these is correct. Executives are interested. Technology functions understand the mechanics better than they are given credit for. What is actually happening is simpler and more interesting: decision paralysis under contradictory signals. When every input to a decision contradicts another input, the rational short-term response is to postpone. It is why airline passengers freeze at a departure board full of contradictory gate changes, and it is why boards defer AI programmes while asking for yet another briefing.

Breaking that paralysis is not a question of more information. It is a question of a sharper framework for reading the information one already has. The five dynamics that follow are, taken individually, well known. Taken together they describe a strategic environment with specific, actionable implications. The chapter ends with a framework for turning those implications into a defensible allocation decision.

Vendor signals are contradictory, by design

The same voices telling executives to adopt generative AI immediately are telling them, in the same breath, that the technology might cause the extinction of the human species. Sam Altman has said, on the record, that in the worst case AI could mean "lights out for all of us," and has also reportedly sought up to $7 trillion to expand chip production for the technology that could cause it. Dario Amodei has written about AI as potentially the greatest revolution in human history, and about why his own company's work might not be controllable. Elon Musk signed the 2023 open letter calling for a six-month pause on frontier AI development, and launched his own frontier AI company four months later. The pattern is not confined to three men. It characterises most of the public communication from the companies building the technology.

The temptation is to assume one of the two messages is sincere and the other is a tactic. Something more interesting is going on. Both messages serve the vendor. The adoption message creates urgency and drives revenue. The extinction message creates the case for regulation written in a way that favours incumbents, since only large well-capitalised firms can meet onerous compliance regimes. The two messages are not in conflict from the vendor's point of view. They are complementary tools pointed at different audiences.

For an executive trying to make a sincere capital-allocation decision, the implication is uncomfortable but useful. Vendor signals are a reliable guide to vendor strategy, not to the state of the technology. The appropriate posture is to decouple the two: take vendor statements as information about vendor positioning, and look elsewhere for information about the technology itself. The best "elsewhere" is almost always one's own small pilot programme, built narrowly and measured honestly, which is the subject of Chapter 04.

The capital flows are largely circular

Nvidia has taken equity positions in a long list of AI companies whose largest cost of operation is buying Nvidia GPUs. In each case, the money invested appears, within a few quarters, as revenue when the investee buys hardware. A similar pattern operates between Microsoft and OpenAI: Microsoft's commitment comes with the condition that the money be spent largely on Microsoft's Azure cloud. From Microsoft's accounting perspective, that is an investment going out and compute revenue coming in.

Zooming out, Sequoia's David Cahn has published an analysis showing that for the compute capital expenditure already committed, end-user AI revenue would need to reach roughly $600 billion annually to justify the investment. Current end-user AI revenue is a small fraction of that number. The implication is specific. Growth signals in the AI sector should be read with a filter. Nvidia's quarterly revenue, the hyperscalers' reported AI revenue, the valuations of model vendors, are legitimate numbers, but they contain a component of recycled capital that will not survive a correction. Any scenario-planning exercise should run the numbers with the recycled component zeroed out and see what remains.

The industry metrics measure the wrong thing

The headlines quoted at the top of this chapter include the most repeated figure in boardrooms right now: enterprise AI spending projected into the trillions over the next few years. The source is usually Gartner, IDC, or a bank's research desk. The number is presented as a market signal, with the implicit suggestion that peers are investing, so you should too. Rarely does the same report put equal visual weight on the measurement that matters: of all that spending, how much produced measurable return?

Spending is not value. Deployment is not ROI. Intent to adopt is not realised benefit. The analysts know this. But the reports that circulate most widely lead with the figure most likely to drive the next consulting engagement or software renewal, and the measurement that would protect the buyer is usually one page deeper, or absent. A CEO who reads "the AI market will reach $X trillion" and acts on it is not responding to a value signal. They are responding to a spending signal dressed as a value signal. The two are different, and the difference is where most of the ninety-five per cent comes from.

The practical response is to read sector reports with one question in mind: what fraction of the quoted spending has produced measurable benefit? If the report does not answer, assume the answer is unflattering, and look to the independent studies that do (MIT NANDA, academic survey work, honest post-mortems from the few organisations that publish them). This is work. It is also the only reliable way to translate "the market is growing" into useful input for a capital-allocation decision.

Today's prices are subsidised

OpenAI reported projected losses of roughly $5 billion in 2024, against roughly $4 billion in revenue. The losses are not a one-off from a bad quarter. They are structural: the variable cost of serving a query from a frontier model exceeds the per-query subscription revenue the vendor collects. Sam Altman has publicly confirmed that even the $200-per-month ChatGPT Pro tier loses money, which is to say the vendor loses more the more a power user uses the product. Anthropic's economics are similar. The entire frontier-model category operates below cost, in the expectation that scale will eventually make it profitable.

The implication for executive capital allocation has two parts. First, every ROI calculation currently being done inside the organisation is priced against a cost the vendor is not charging. If a Copilot deployment pays back at thirty dollars per user per month, the responsible question is whether it still pays back at sixty or ninety. Any model that does not stress-test for a price doubling is incomplete. Second, the vendor-selection decision is also a pricing-power decision. Choosing to build a critical workflow on a vendor's API is implicitly accepting whatever that vendor charges in three or five years. The answer is not necessarily to avoid frontier APIs. It is to build with a credible exit option, which usually means a smaller specialised model running on infrastructure the organisation controls as the fallback. That is the "frugal AI" option introduced in Chapter 01, and it is strategic insurance, not a compromise.

US growth now leans on AI spending

In 2025, the money US tech companies spent building AI data centres became one of the main engines of American growth. By one estimate, from the Harvard economist Jason Furman, it accounted for almost all US growth in the first half of the year. Economists disagree on the exact share, because most of the chips and servers are imported, but none disputes the direction. The same concentration shows in the stock market, where a handful of large tech companies, whose profits depend on this build-out, account for much of the recent gains. Apollo's chief economist, Torsten Sløk, adds the warning that matters most: this spending is growing nearly twice as fast as the housing boom before 2008, and a cycle built that fast can unwind just as fast.

For a European executive, this has an unpleasant consequence. European companies have invested little in this build-out, so they have not shared in the growth. But if the cycle turns, the shock will still reach them, through share prices, interest rates and exchange rates. Europe imports the risk without having captured the gain. The point is not to panic. It is to recognise that a decision to invest in AI is now also a bet on the economic climate, and that any multi-year AI plan should hold up in a scenario where the AI boom slows.

How to decide anyway

The five dynamics are real. The 95% is real. What a C-level reader needs is not another layer of caveats but a decision framework robust enough to produce an answer. The following four principles, applied together, are enough to cut through the current environment and produce a defensible AI capital-allocation posture.

1. Principle 01 Decide conditionally, not categorically The right answer is not yes or no. It is "yes, under conditions we specify in advance": named owner, quantified pain, bounded scope, measurable baseline, kill-criterion, learning-sized budget.
2. Principle 02 Assume the vendor will fail Several frontier AI vendors will not exist in five years. Build with a thin abstraction layer, a small-model fallback, and no proprietary formats where open ones work. Familiar discipline from the cloud era.
3. Principle 03 Model ROI against unsubsidised prices Run every business case twice: at today's prices, and at two to three times those prices. If the case fails the second test, redesign the deployment before funding it.
4. Principle 04 Plan for the AI boom to slow An AI investment is now also a bet on the economic climate. Every multi-year plan should include a scenario in which the AI spending boom slows, and the programme should still pay off under it.

Applied together, these four principles do what the strategic environment is designed to prevent: they produce a clear, defensible decision. The answer in most cases is to engage with generative AI, deliberately, at bounded cost, with governance from day one, and with an exit path if the economics or the vendor change. That is a familiar operating posture. It is the one that worked for the previous three technology cycles. The current moment is noisier than those were, but the discipline required to navigate it is not new. Generative AI is as useful as the organisation deploying it makes it, and the guardrails are the price of making it useful.