Beyond Data Centers, the Next Wave of AI Winners

Anyone who invested through the 1990s does not need a lesson in GPUs to understand the setup. The internet age followed a simple sequence: a breakthrough arrived, capital rushed into the infrastructure, and then the biggest opportunities broadened to the businesses that made that infrastructure useful.

Between 1996 and 2001, telecom companies spent roughly US$444 billion laying fibre and building network capacity. NASDAQ: CSCO Cisco, Intel and the big telecom carriers became the obvious ways to invest in the internet. By 2002, however, only a small percentage of the fibre already in the ground was carrying traffic. The network had been built well ahead of demand, leverage was high and investors had paid extraordinary prices for the companies supplying it.

That did not mean the internet failed. It meant the infrastructure was becoming the floor underneath a much larger economy. Search, online commerce, digital payments, mobile computing and cloud services all grew on top of capacity that somebody else had already paid to build.

Cisco is the useful caution. At the March 2000 peak it briefly became the world's most valuable company. The stock then fell roughly 85-90% and took more than 25 years to reclaim that nominal peak. The business survived. The network remained essential. The entry price was the problem.

The investment question is not whether Nvidia stops being important. It is which businesses become more valuable because Nvidia succeeded.

The Second Phase Created a Different Set of Winners

Several of the businesses that turned abundant connectivity into actual products went on to dramatically outperform the companies that had dominated the build-out. Not all of them. But enough to shift the centre of gravity entirely.

Amazon had already fallen more than 90% in the bust. Yet from 2002 through 2010 its share price rose roughly 14.7 times. NASDAQ: AAPL Apple rose roughly 25 times over the same period as the iPod and then the iPhone moved the internet from desktops into pockets. NASDAQ: GOOGL Google listed in 2004 and still delivered a multi-bagger as it organised and monetised the information already flowing across the network.

None of them replaced the infrastructure companies. NASDAQ: AMZN Amazon didn’t displace Cisco, Google didn’t displace Intel, Apple didn’t make fibre unnecessary. They built enormous businesses on top of the system that was already there. That distinction matters.

Map the Internet Age Onto the AI Age

We think this is the more useful lens for the current AI cycle. The technology is different and analogies have limits, but the investment sequence is familiar: build the computing base first, then figure out the new problems that base creates.

The first phase has been dominated by AI’s equivalent of processors, fibre and networking gear. Amazon, Alphabet, Meta and Microsoft lifted combined capex from roughly US$140 billion in 2023 to more than US$416 billion in 2025. The International Energy Agency expects global data-centre electricity consumption to almost double, from about 485 TWh in 2025 to around 950 TWh by 2030. That is a staggering amount of concrete and copper.

Nvidia was the clearest beneficiary because accelerators were the first scarce resource. Then the constraint moved to data-centre space and power. Those issues remain, but chips and megawatts aren’t the whole problem anymore. Capacity has to be financed and actually built. Thousands of accelerators have to talk to each other efficiently. AI has to generate measurable productivity outside the data centre. And autonomous agents have to operate inside sensible guardrails.

Here is where the internet comparison gets uncomfortable. The technology is moving much faster this time. Fibre laid during the dot-com boom could remain useful for decades. Today’s AI hardware can become economically dated within a few years, while new chips and architectures arrive at a pace that would have been unthinkable in the telecoms era. The internet took decades to progress from early web pages to mature platforms. AI is compressing comparable stages into years. That makes the investment consequences more severe. Capital pours into what looks like a scarce bottleneck, only for technology to improve or a new architecture to shift the constraint elsewhere before investors have earned an adequate return. Being right about AI demand is not enough. You also have to be right about which bottleneck stays valuable long enough to pay for itself.

The build-out is creating its own second-order opportunities. We group them into four areas: capacity, connectivity, physical execution and control.

1. Switch It On – Capacity

Buying accelerators is not the same as delivering usable compute. You need powered sites, cooling, networking, functioning clusters. And customers increasingly want that capacity now, not when a conventional cloud construction program eventually gets around to it.

That's the neocloud opportunity. CoreWeave and Nebius aren't simply selling GPUs; they're selling time-to-capacity. CoreWeave reported Q2 2026 revenue of about US$2.58 billion, up 112% year on year, with backlog near US$104 billion before more than US$25 billion of subsequent commitments. Yet only around 1.5 GW of power was active at quarter-end against a much larger contracted pipeline, and 2026 capex guidance sits at US$35–39 billion.

Then there's the power-first route. NASDAQ: IREN  IREN came out of Bitcoin mining with something most AI startups don't have: operational data centres with significant owned power already connected to the grid. Its Childress, Texas site runs at roughly 3.3¢/kWh, with the broader portfolio around 5¢ — well below the 6–11¢ band most AI data-centre models assume and roughly half the US industrial average. Cheap power that's already switched on is a structural edge when the biggest bottleneck in the chain is getting megawatts live. ASX: MP1 recently offered GPU as a service and ASX: NXT NextDC also focus on Datacenter-as-a-Service (DaaS)

The order book across all these models shows the demand is real. The investment question is execution: can contracted power, land and hardware become live revenue before financing costs and construction risk catch up? Nebius offers yet another model with customer prepayments helping fund reserved capacity. A signed contract is not a working cluster, and a power asset is not a functioning AI data centre. Investors should pay attention to that gap.

2. Connect It – Connectivity

The internet needed routers and switches because more connected computers created more traffic. Large AI clusters create an analogous problem inside the data centre. One accelerator works fine on its own. Put tens of thousands together and the network becomes part of the computer.

Nvidia's own data-centre networking revenue reached US$14.8 billion in Q1 FY2027, up 199% year on year. But Nvidia isn't the only company the clusters depend on. Astera Labs supplies much of the less visible connective tissue (retimers, CXL products, Scorpio fabric switches) and reported Q2 2026 revenue of US$392.4 million, up 104%. Marvell sits across both custom silicon and connectivity — it designs custom AI accelerators for hyperscalers who want alternatives to off-the-shelf GPUs, and supplies the electro-optics, PAX retimers and Ethernet controllers that stitch large clusters together. Its data-centre revenue has been the fastest-growing segment of the business [verify latest quarter figures before publishing]. Both Marvell and Astera Labs accept the basic design — lots of separate processors talking to each other over a network — and focus on making it work better at scale: faster links, lower latency, cleaner signal integrity as clusters grow from thousands of GPUs to tens of thousands.

Cerebras is asking a different question entirely. Instead of speeding up the connections between conventional chips, it puts an entire wafer of 900,000 AI-optimised cores onto a single processor to eliminate the inter-chip communication problem rather than optimise around it. If giant clusters keep scaling, Marvell and Astera benefit. If the architecture itself hits a wall, Cerebras has the edge. All three are genuine bets on connectivity being a persistent bottleneck, just with very different answers.

3. Make It Useful – Physical Execution

The internet became much more valuable when it stopped being something people looked at on a desktop and started changing how commerce, payments and everyday business actually worked. AI has to make the same jump. And unlike a chatbot demo, the examples that matter here are the ones where somebody can point at a line item and say: that paid for itself.

On the factory floor, 542,000 industrial robots were installed globally in 2024, more than twice the number a decade earlier, and the International Federation of Robotics expects annual installations to exceed 700,000 by 2028. But factory robots are only one slice. On the farm, Deere & Co NASDAQ: DE is using computer vision to distinguish crops from weeds in real time, its See & Spray technology can cut herbicide use by up to two-thirds on certain passes, which is the kind of measurable cost saving that gets a purchase order signed without a board presentation. On the road, NASDAQ: TSLA Tesla launched its Cybercab robotaxi service in Austin in June 2025 [verify: widely reported at launch, but confirm current status], while Waymo has been running commercial driverless rides across multiple US cities for over a year. None of these need a general-purpose humanoid that understands the entire world. They need AI that can do one specific thing well enough to justify the spend.

The common thread is that a CFO can measure the outcome. Fewer chemicals per hectare, higher throughput per shift, lower cost per mile. That is what will drive adoption faster than any demo reel, and it's why we think physical execution is where AI spending converts into real economic output first. For investors who'd rather not pick a single platform (and we'd include ourselves in that group), vehicles like the Global X Robotics & Artificial Intelligence ETF NYSE: BOTZ offer diversified exposure across industrial automation, autonomous systems and adjacent plays.

4. Control It - Governance

This is the layer most investors are sleeping on. Less visible than the others, but in our view the most important as AI moves from answering questions to taking actions. An autonomous agent can access a database, write code, modify a workflow, send an email, interact with a payment system. At that point the question shifts. Can the AI do something? Sure. Should it be allowed to?

NASDAQ: IBM  IBM found 77% of technology leaders believe AI adoption is outpacing governance; only 11% feel fully prepared for AI-agent deployment at scale. NASDAQ: GTLB GitLab found 92% of DevSecOps professionals face AI-code governance challenges. Those aren’t abstract numbers. They describe a new enterprise problem where permission and auditability have to keep pace with capability, and right now they aren’t.

NASDAQ: PLTR Palantir increasingly operates as a runtime control layer, helping organisations decide what data AI systems can access and what actions they can take. GitLab sits further upstream, controlling how software gets created, reviewed, secured and audited. The easier AI makes it to produce code and take autonomous action, the more valuable those controls become. The risk, as always, is paying for that thesis before the revenue catches up.

What the Internet Comparison Does (and Doesn’t) Tell Us

The analogy explains how investment leadership can shift as scarcity moves. It doesn’t mean AI infrastructure must repeat the fibre bust, or that NASDAQ: NVDA Nvidia must follow Cisco’s trajectory. Today’s hyperscalers have stronger balance sheets, much of the demand is backed by large contracts, and AI hardware depreciates far faster than fibre ever did.

But the narrower lesson holds: an essential technology can be a poor investment at the wrong price, while companies one layer out can become increasingly valuable as the original infrastructure succeeds. We’ve seen this before. The first-wave winners can remain excellent businesses even as the opportunity set gets wider around them.

The Next Phase Is About What the Infrastructure Enables

Twenty-five years ago investors were fixated on processors, fibre and networking equipment because those were the scarce assets. The infrastructure mattered. But the opportunity didn’t stop there. Once the network existed, an entirely new generation of companies became the market leaders.

AI won’t repeat that history line for line. But the sequence is worth paying attention to. The first phase was about building enormous computing power. The next is about turning that power into something useful, connected, productive and controlled.

That’s why we’re looking beyond the data centre. Not because the data centre stops mattering, but because success at the infrastructure layer is what creates the demand for everything around it. And that’s where we think the next set of opportunities will be most underpriced. We have open recommendations on the DaaS sector and a thematic investment via Structured Investments over 2 years 

We are holding a free webinar next week, Register here to watch live or On-Demand