Jensen Huang's five-layer AI cake, and which layer a New Zealand investor can actually buy

At Davos in January, Jensen Huang described AI to BlackRock's Larry Fink as "a five-layer cake": energy, chips and computing infrastructure, cloud data centres, AI models, and the application layer on top. He called the buildout underneath it "the largest infrastructure buildout in human history." Whether or not you take Huang's framing at face value, it's a useful map for working out what a satellite position in "AI" would actually be exposed to, because the five layers don't behave the same way, aren't equally investable, and don't sit in the same tax bucket.

The five layers, and what's actually buyable from New Zealand

Energy is the base of the stack - the power generation and grid capacity needed to run data centres, which is turning into one of the physical constraints on how fast the buildout can happen. Kernel's Global Infrastructure fund and Smart's INF fund both give broad exposure to utilities and infrastructure, but neither is built around the specific AI-power story - they're general infrastructure funds that happen to include some of the same names, not a pure play on data centre energy demand. GMO's Power Infrastructure ETF (KWH), listed directly in the US, is built specifically around the power generation, grid and electrification build-out this layer actually needs - a genuine pure play, not an adjacent fund that happens to overlap.

Chips is Nvidia, TSMC, Broadcom, AMD and the rest of the semiconductor supply chain. There's no NZX-listed, PIE-wrapped pure-play semiconductor fund - Smart's thematic shelf covers automation and robotics (BOT), which overlaps with chips without being the same thing. Direct exposure to the chip layer specifically means Hatch, Stake or Interactive Brokers, buying the actual companies or a US-listed semiconductor ETF - VanEck's Semiconductor ETF (SMH) is the standard one - which puts you in FIF territory once your cost basis crosses $50,000.

Infrastructure - the cloud data centres themselves - is largely Microsoft, Amazon, Google and Meta, the hyperscalers actually building and running the AI factories. This is the layer most NZ investors already own more of than they realise: a standard Global 100 or S&P 500 PIE fund is cap-weighted, and these are among the largest companies in any of those indices already. Adding a satellite position here on top of an existing global equity core is often duplicating exposure you already have, the same overlap problem Issue 15 covered with global versus US funds. Where a satellite position does add something is a fund built specifically around data centre real estate and digital infrastructure rather than the hyperscalers themselves - Global X's Data Center & Digital Infrastructure ETF (DTCR) is the direct-listed example - or a fund like Defiance's AI & Power Infrastructure ETF (AIPO), which deliberately blends this layer with the energy layer above it rather than picking one.

Models is where the framing gets interesting for a retail investor anywhere, not just here. OpenAI and Anthropic, the two labs doing the most-discussed work at this layer, are both private companies. There's no way to buy either one on a public market at any size, from New Zealand or from anywhere else. What look like "model layer" plays from the outside - Google DeepMind, Meta AI - are actually just business units inside public companies you already hold if you own a broad global fund, which folds this layer back into the infrastructure point above.

Applications is the layer Huang says is where the actual economic benefit lands - AI built into healthcare, financial services, manufacturing. Some of that sits inside public companies (Microsoft's Copilot work, Salesforce, ServiceNow), already inside a diversified core holding the same way the infrastructure layer is. A meaningful amount of the application layer, though, is still venture-backed startups with no public listing at all.

The bubble question, both sides

At the same Davos session, Fink asked Huang directly whether the world is investing enough in AI rather than too much, and Huang's answer was more, not less. In August, Nvidia announced financing partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, designed to bring over $500 billion of third-party capital into AI infrastructure - and addressed head-on the obvious question of whether Nvidia financing the build-out of its own customers' compute is circular. Nvidia's position is that the capital comes from independent institutional investors underwriting each deal on its own merits, not from Nvidia itself. Sceptics point to the same arrangement and see a chip supplier with an incentive to keep demand for its own product financed, whichever way the underlying AI economics eventually land. Both readings are being made in good faith by people closer to the deals than this newsletter is - it's not a question with a clean answer yet, and anyone telling you it is has picked a side, not found a fact.

What this actually changes about sizing a satellite

Two of the five layers - models and applications - are the ones generating the most headlines and the least public investability. Two more - infrastructure and, to a lesser extent, chips through the mega-cap names - are largely already sitting inside a standard diversified core fund. That leaves energy and the pure-play chip names as the layers where a deliberate satellite position adds something a core holding doesn't already have, and both of those, done properly from New Zealand, mean going direct and accepting the FIF treatment that comes with it. Before sizing anything against the "AI theme" generally, work out which of the five layers you're actually trying to add, and check whether your core fund already owns it.

Where my own build actually sits in this

ARKK is the position I started with - not layer one of Huang's stack, and I'm not pretending it is. ARK's Innovation ETF is a broad, actively-managed bet across disruptive technology generally, not a pure play on any single layer of the AI buildout specifically. I'm treating it as the foundation the rest sits on top of, not because it maps cleanly onto the framework above, but because it doesn't need to.

From there, I'm building satellite exposure across the two layers this piece argues actually need one: energy and chips, the same two the framework above points to as the layers a standard core fund doesn't already cover. Which specific funds and how much of each is still moving as I build the position - I'd rather report back once it's settled than publish a snapshot that's already out of date by the time you read this. I've also added a separate, smaller position in Chinese platform companies, which sits outside this framework entirely - a different bet, for different reasons, not part of the AI buildout at all.

All of it's going in gradually, funded as the money's actually there to invest rather than as one lump sum, the same dollar-cost-averaging discipline Issue 13 makes the case for. It's a live position, not a finished one - what's actually in it will keep changing as I keep adding to it.