Investing in AI: Why the Most Valuable Layer Is the Hardest to Call

Why “Investing in AI” Is Really Five Decisions

The reason the single-stock instinct is so common is that for the past few years one layer of the AI stack — the chips — has done most of the visible work, and one company has dominated it. But a chipmaker, a power utility, a data-centre landlord, and the software firm that eventually turns all of it into something a business will pay for are radically different investments. They sit at different points in the value chain, earn money in different ways, and carry different risks. Lumping them together as “AI” obscures more than it reveals.

Huang’s framework is a tidy way to separate them. He describes AI as a five-layer cake: energy at the base, then chips, then infrastructure, then models, and finally applications at the top. Each layer is useless without the ones below it — there is no clever application without a trained model, no model without compute, no compute without data centres, and none of it without an enormous and growing supply of electricity.

One honest caveat before we climb the cake. This framework comes from the man who sells the shovels. NVIDIA sits squarely in layers two and three — the chips and the infrastructure — which happen to be the layers soaking up the most capital right now. A narrative in which AI is “the largest infrastructure build-out in history” is, conveniently, very good for the company narrating it. That doesn’t make the framework wrong; it’s a genuinely useful map. It just means you should hold the more breathless capital-spending forecasts loosely. Frameworks describe reality. They don’t guarantee it.

The Five Layers — and What Investing in Each One Means

Layer 1 — Energy: the literal foundation

Before there is intelligence, there is electricity. Every AI query traces back to electrons moving and heat being managed, and the scale of demand is the part most investors underestimate. Training and running large models consumes power on an industrial scale, which is why the AI story has quietly become an energy story — from grid operators and independent power producers to the renewed interest in nuclear, including the small modular reactors being pitched to sit next to data centres. South African readers, of all people, will grasp the point intuitively: no electricity, no economy, and increasingly, no AI. Exposure here looks like utilities, power infrastructure, and the fuels and equipment behind them — slower, more regulated, less glamorous, and far less correlated to the AI hype cycle than the layers above it.

Layer 2 — Chips: the compute

This is the layer everyone already knows. The specialised processors that handle the massive parallel computation AI requires are dominated by a short list of names — the designers, the foundries that actually manufacture the silicon, and the equipment makers whose machines make the manufacturing possible. It is the most direct way to own “AI”, and it has been the most rewarding. It is also the most cyclical and the most concentrated: a handful of companies carry an enormous share of the value, valuations are demanding, and any wobble in the spending plans of their biggest customers shows up here first. High reward, high beta, low margin for error.

Layer 3 — Infrastructure: the build-out

Above the chips sit the data centres that house them and the cloud platforms that rent them out — the hyperscalers. This is the capital-expenditure layer: hundreds of billions already committed, with the industry arguing about whether the next trillions are necessary investment or speculative excess. It also reaches into the unglamorous physical economy — networking, cooling, construction, and the property trusts that own the buildings. For most investors this layer is the most accessible, because the hyperscalers are some of the largest companies in the world and already sit, heavily weighted, inside any global equity fund.

Layer 4 — Models: the intelligence

The foundation models themselves — the large language models and their successors — are the layer most people think of as “AI”, and the hardest for ordinary investors to own directly. The leading labs are largely private, and the public-market route is usually indirect: the hyperscalers hold significant stakes in, and supply compute to, the major model developers. There’s a structural tension worth noting here. As models proliferate and open-source alternatives catch up, raw model capability is drifting towards commoditisation — which suggests the durable economic value may not live in the models at all, but in the layer above them.

Layer 5 — Applications: where the value lands

At the top is the layer where AI meets an actual user or business — the software and services built on top of the models. Huang is blunt that this is “where the economic benefit will happen”. It is also, for now, the thinnest part of the investment case: the foundations are being poured at extraordinary cost, but the question of which applications will earn a durable return on all that spending is still largely open. Which is exactly why it deserves its own section.

The Application Layer Is the Biggest Unknown

Here is the most useful thing history can tell us about where AI is headed, and it comes from the last great platform shift: the internet.

The web went mainstream in the mid-1990s. The infrastructure of that era — the cabling, the servers, the browsers, the telecoms build-out — was being laid years before anyone could sensibly tell you what the internet was for, beyond email and a few websites. The companies that captured the most imagination at the time were the picks-and-shovels names. And many of them did fine. But the truly enormous winners — the ones that reshaped daily life — didn’t really arrive until a second event roughly a decade later.

That event was the smartphone. The iPhone landed in 2007, the App Store opened in 2008, and almost overnight the internet went from something you sat down at a desk to use into something in everyone’s pocket all day. Only then did the application layer explode. Uber, Airbnb, Instagram, WhatsApp, mobile banking, the entire on-demand economy — none of these were obvious from the infrastructure build-out of the 1990s. They needed the platform to exist first. The cabling made them possible; it didn’t predict them.

The lesson for AI investors is a humbling one. We are, plausibly, still in the cabling phase. The energy, the chips, and the data centres are being built at staggering cost, but the equivalent of the App Store moment — the shift that unleashes a wave of applications nobody has thought of yet — may still be ahead of us. The biggest winners at the application layer may not be public companies today. Some may not exist at all. That uncertainty is not a reason to avoid the theme; it is a reason to hold the obvious “winner” loosely, and to resist the temptation to bet the portfolio on the version of the future that happens to be visible right now.

The Other Half of the Problem: Who AI Disrupts

The five layers describe where AI creates value. But AI also destroys it, and that is the harder half of the problem for an investor. For every business the technology lifts, others find their model quietly undermined — and spotting those in time matters just as much as picking the winners.

The clearest pressure right now is on knowledge work and software services. IT consulting is the textbook case: Accenture, the world’s largest consultancy, has watched its shares fall by over 50% over the past year even as its revenue kept growing — because the market began pricing in a future where AI automates the junior, billable-hours work that consulting has sold for decades. The irony is almost too neat: the firm sells the very AI its investors fear. Seat-based software faces a related question — if an AI agent can do the work, do you still need as many licences? And content and education businesses have already felt it directly: the online-study company Chegg lost the bulk of its market value once free chatbots could do what students once paid it for.

History rhymes here, and the rhyme is worth sitting with. Think of newspapers in the early 2000s. Their fundamentals held up for a surprisingly long time — people kept reading, the presses kept running, the profits kept coming — even as their share prices began a long, grinding slide. The market had worked out, well before the income statements did, that classified advertising and the print bundle were structurally finished. By the time the numbers confirmed it, the value was long gone. The market is often an earlier and more ruthless judge of structural decline than the financial statements are.

That is what makes this such a treacherous environment to invest in. A disrupted incumbent can look cheap and keep posting respectable results for years while its share price de-rates relentlessly — the classic value trap. So investing in AI isn’t only about finding the winners, which is hard enough given how little we can yet see of the application layer; it’s also about not catching falling knives among the losers.

One practical response is to ask what AI is least likely to touch in the near term. This wave is overwhelmingly digital — it automates what gets done at a keyboard. Businesses anchored in the physical world — infrastructure, utilities and energy, healthcare delivery, real assets, skilled trades — look relatively insulated, at least until robotics and “physical AI” mature. That’s a genuine “for now”, not “forever”: the same roadmap that gives us the five-layer cake points to physical AI as the next wave, so today’s safe harbour may have a five-year shelf life rather than a permanent one. For a diversified portfolio, deliberately owning businesses on neither side of the immediate disruption — neither the priced-for-perfection winners nor the value-trap losers — is sensible ballast.

How South African Investors Actually Get Exposure

The practical news for South African investors is that you almost certainly already own a great deal of AI — whether you meant to or not. The infrastructure layer is dominated by the largest companies on earth, and those companies make up a substantial slice of any global equity index. If your offshore allocation runs through a world equity or S&P 500 fund, a meaningful and growing portion of it is, in effect, an AI infrastructure bet. The first useful exercise is therefore not “how do I add AI?” but “how much AI do I already have, and am I comfortable with it?”

From there, the routes to deliberate exposure are familiar ones. Broad global funds give you the megacaps without a concentrated wager. Thematic and semiconductor ETFs give you a sharper tilt towards specific layers, at the cost of concentration. Direct shares give you the sharpest exposure of all, and the least diversification — a route better suited to the satellite portion of a portfolio than its core, and one we discuss candidly in our piece on DIY investing. Accessing offshore markets in the first place runs through the usual exchange-control allowances, which is its own planning question we cover in the offshore investing guide.

The framework earns its keep here. Deciding which layer you want exposure to, and how much, is a more disciplined question than “which AI stock should I buy?” — and it tends to produce a more resilient portfolio.

A Word on Valuations and Concentration

Two things can be true at once, and with AI they currently are. The technology is genuinely transformative, and the assets that represent it are priced for a great deal of that transformation to go right. Equity markets sit at record highs, with the most AI-exposed names carrying valuations that invite comparisons to the dot-com era. There is a live and reasonable debate about whether we are in a bubble — and a specific risk worth understanding: if models keep getting cheaper and more efficient to run, the “trillions still needed” infrastructure thesis could deflate faster than the build-out crowd expects.

None of this is a market call. It is a reminder that thematic conviction and position sizing are different things. Believing AI will reshape the economy is not the same as believing every company currently labelled “AI” will reward you for owning it at today’s price. This is where the layered view and a clear sense of your own risk tolerance do their best work — a theme we develop in our guide to investment risk and return, and one that sits alongside the broader question of how markets behave over long cycles. The discipline is the same one that applies to any exciting story: own it deliberately, size it sensibly, and don’t confuse a strong narrative with a margin of safety.