Twelve months ago, the AI conversation among clients centred almost entirely on semiconductors and data centres. Today it has matured into something more nuanced and, frankly, more investable. At NCB Capital Markets Cayman, we’ve found it useful to break the AI trade into four distinct layers, each with a different risk and reward profile. Understanding which layer you’re actually buying, rather than treating ‘AI exposure’ as a single trade, has become essential to navigating this market sensibly.
Layer 1: The picks and shovels
The first layer is the infrastructure build-out – the chips, memory, servers and power capacity underpinning the entire AI stack. This is where the bulk of investor attention, and capital, has gone. It’s a familiar pattern: every major infrastructure cycle in history, from the railroads to fibre-optic cable in the dot-com years, has seen the earliest and most visible winners emerge among the builders rather than the eventual users of the technology. This cycle is no different. Bottlenecks have emerged across the supply chain, pushing up component costs – Apple’s recent price increases are a direct result – and we have seen parabolic moves in names like Micron, Dell and AMD.
The bull case is simple: Compute demand will stay structurally strong for years. We’re more sceptical. The cure for high prices is usually high prices. The hyperscalers and large customers don’t sit still when costs spike; they build their own custom silicon, they compress models to use less memory per workload and they find workarounds. With expectations for this layer now priced for perfection, we don’t see it as an attractive entry point today.
Layer 2: The frontier models
The second layer is the large language models themselves. Here the picture is less flattering. It’s genuinely difficult to point to durable differentiation between the leading models such as ChatGPT, Claude or Grok, which means switching costs for customers are low. When your product isn’t meaningfully differentiated, the only lever left to grow market share is price and that sets up a commoditisation dynamic, or a race to the bottom on margins.
The recent wave of capable Chinese open-source models, developed at a fraction of the budget of their US counterparts, only accelerates this. Each new generation of frontier model also appears to require exponentially more compute for smaller gains, while enterprise adoption is not scaling at anywhere near the same pace. For this reason, we plan to avoid the widely anticipated OpenAI and Anthropic IPOs for client portfolios, unless clear evidence emerges that one model has built a genuine, durable differentiating competitive advantage.
Layer 3: The ecosystems
This is where it gets more interesting. A powerful technology doesn’t automatically make a good investment – the real question is what AI does to the competitive position of the businesses deploying it. Much of the financial media debate has focused on the mismatch between today’s enormous capital expenditure and today’s still-modest AI monetization. That’s a fair concern, but we think it’s primarily a timing problem, not a verdict on the technology’s eventual value.
Our lens is simpler. Does AI strengthen an existing ecosystem, or does it erode it over time? Microsoft, Apple and Amazon are examples of the former. None of them has “solved” AI integration yet. Copilot is still evolving, and Apple’s own AI strategy has been notably unhurried, but in each case, AI is being layered onto moats that already exist, such as distribution, devices, cloud infrastructure, and customer lock-in. It reinforces rather than replaces what they’ve built.
The opposite is also true. Per-seat SaaS (software as a service) businesses look vulnerable as AI agents increasingly do the work instead of human users. Names like ServiceNow and Salesforce sit squarely in the crosshairs of that debate. Generative AI poses a direct challenge to the traditional creative software model that has underpinned Adobe’s economics. And financial data aggregators such as S&P Global face longer-term questions as AI-native tools make it easier to assemble and interpret data that once required a paid intermediary.
Layer 4: Proprietary data
The fourth layer is arguably the most durable: companies sitting on data that simply cannot be replicated. Meta’s social graph, Tesla’s real-world driving data, Uber’s mobility data, and Palantir’s deep access to organizational data all fall into this category.
Layering AI onto data of this kind doesn’t just add a feature, it deepens a moat that competitors have no realistic path to copying.
Where we stand
Layers three and four are where we’re actively concentrating our research effort, and where we believe long-term client capital is best positioned. That said, we’re not rushing to deploy aggressively into these layers, because the macro backdrop has turned less accommodating.
Inflation is ticking higher on two fronts at once. Unresolved tensions around the Strait of Hormuz continue to put upward pressure on oil prices, while pricing pressure from the AI infrastructure build-out is spilling into the broader economy through higher software and hardware costs. The Federal Reserve has leaned more hawkish in response, and while it’s still early to characterise Kevin Warsh’s reaction function fully, his historical commentary suggests a central bank less inclined toward a “Fed put” or renewed balance sheet expansion than markets may be hoping for.
The stakes are higher than usual, too. AI-related capital spending accounted for roughly 75% of US GDP growth in the first quarter, which is a striking concentration in a single theme. That means any slowdown in AI-related spending could ripple disproportionately through investment decisions, labour markets and the broader wealth effect. Meanwhile, S&P 500 earnings expectations are already highly elevated, the pace of capex growth looks difficult to sustain, and the bond market, through persistently higher yields, appears to be flagging some of the same concerns.
In short, we have a clear road map of where we want to add exposure as this cycle matures. For now, the macro environment is keeping us disciplined about when.
Related Videos
