Most of the money and attention in AI has flowed to a handful of US mega caps. Yet the physical machinery behind the boom, the chips, the memory, the servers and the power that runs them, is built and supplied largely in the emerging world. For investors deciding where the next phase of AI returns will land, that fact deserves more weight than it usually gets.
This is not a speculative play on a theme. It is exposure to the parts of the supply chain that every dollar of AI spending has to pass through.
Follow the scarcity
It helps to think of AI as NVIDIA’s five-layer AI cake, a stack of dependent layers: energy at the bottom, then chips and memory, then the broader systems infrastructure, then the models, and finally the applications. Scarcity at any layer constrains everything above it, and the firms that relieve the binding constraint tend to earn outsized returns for as long as it lasts.
The market still recognizes that GPUs, HBM and advanced packaging remain the binding constraints in AI infrastructure. High-bandwidth memory, the stacked DRAM that sits next to every AI accelerator, is sold out for 2026 across all three suppliers, and 2027 is close to fully booked, with some orders already reaching into 2028. SK Hynix has pre-sold its entire DRAM, NAND and HBM output to customers led by Nvidia, and both SK Hynix and Samsung have warned publicly that AI-driven memory shortages could run past 2027. Analysts covering the group see little price relief before late next year. Advanced packaging tells the same story: TSMC’s CoWoS capacity is oversubscribed through at least the middle of 2027, and each new wafer of capacity gets absorbed almost as soon as it comes online. The chip and memory constraint hasn’t eased. It is the live shortage, and it sits squarely in emerging markets.
The AI memory and packaging bottleneck is still tightening: critical, emerging-market-concentrated supply nodes are booked years ahead of delivery.
Taiwan and Korea hold the nodes that are hardest to replace
Begin with the links nobody can easily substitute. Taiwan’s TSMC makes roughly 90% of the world’s most advanced logic and sets the pace for the whole industry, with capital spending on track toward an estimated $100 billion by 2028, almost all of which feeds back into the equipment makers. TSMC is also the near-sole source of CoWoS packaging, the step that has become the real bottleneck for shipping Blackwell-class systems. Foxconn assembles an estimated 40% of AI servers and has reported AI server revenue up 170% year on year. Delta Electronics leads the world in rack-level power management, which matters more every quarter as power density climbs.
Korea is where the memory gets made, and the supply arithmetic favours the incumbents. SK Hynix holds approximately 60% of the HBM market and has locked up close to 70% of the orders for Nvidia’s next-generation HBM4. Samsung spans HBM, DRAM and foundry, and is ramping HBM4 and HBM4E to close the gap. With capacity sold out and pricing firmly in their favour, both stocks have run hard, Samsung up around 175% and SK Hynix around 220% so far this year, and they now sit inside a memory oligopoly worth roughly $3 trillion between them. As one memory specialist put it, this is not the moment to apply the usual sell discipline. Memory has always been a cyclical business prone to boom and bust, and that risk is real. What sets this cycle apart is how long the shortage looks set to last.
China offers a different angle
China provides a thesis with low correlation to the rest: state-backed substitution. Shut out of EUV lithography and the leading-edge nodes, Chinese firms are building a self-contained domestic ecosystem against real demand. Cambricon is a pure-play AI chip designer. Alibaba’s T-Head unit makes its own inference silicon and RISC-V CPUs. Baidu combines chip design, a frontier model and an application business in one company. For investors who can reach these names, they offer a way to participate that holds up reasonably well whichever direction the US-China technology contest takes.
Don’t ignore the commodities
This is where a style-agnostic approach earns its keep. The big memory and foundry names are quality-growth stocks that screen well on almost any momentum or earnings-revision model, and they belong in the portfolio. But the same AI shortage that bid them up runs straight back into a set of deep-cyclical, value-shaped businesses that a styleboxed investor would never hold in the same portfolio. All Weather Capital is not wedded to a single playbook; we go wherever the industry leader for a given bottleneck happens to sit, whatever label the market has pinned on it.
Watch the raw inputs alongside the stocks. Uranium, copper, germanium, gallium and rare earths decide which shortage can be cleared and how quickly. China refines roughly 91% of the world’s rare earths and produces most of its gallium and germanium, leverage that runs straight back to the companies depending on those materials. The point is to hold the memory compounder and the copper miner and the uranium producer at once, sized to where the constraint is tightest, and to move between them as the binding bottleneck shifts. Tracking the commodity layer gives you an early read, and the agility to rotate before the crowd does.
Physical AI – the next frontier
The same logic points to where AI goes next. The frontier is shifting from software toward physical AI, the robots and humanoids that put intelligence to work in the real world, and that shift leans on a supply chain that looks a lot like the one for chips. Global humanoid shipments were still small in 2025 but grew more than fivefold over the prior ear, and 2026 is shaping up as the step from pilots to early mass production. Each humanoid is a dense bundle of motors and actuators that needs roughly three to four kilograms of rare-earth magnets. China mines close to 70% of the world’s rare earths, processes around 90% of its magnets, and is reusing its mature EV supply chain to build robots at scale through names such as UBTech and Unitree and the lidar and battery makers around them, while Korea supplies much of the precision actuation and componentry. Physical-AI leadership may well be built on emerging-market motors and magnets, and the same style-agnostic discipline is what will surface the leaders as they arrive.
How to position
None of this guarantees outperformance. A shortage is an edge, not a sure thing, and the timing is a matter of judgement. The sensible approach is to map exposure to the constraint most likely to bind next rather than chase a layer that has already re-rated, and to stay willing to own a growth compounder and a cyclical commodity leader side by side when both sit on a real bottleneck. With memory and packaging sold out well into 2027, foundry and HBM stay at the core of the holding, alongside the power and systems names such as grid equipment, rack management, advanced packaging and optical components, the commodity producers feeding all of it, and the Chinese self-reliance plays as a longer-dated option.
A few dates are worth keeping in view: the CPU and systems constraint deepening through 2026, Nvidia’s Vera CPU ramp, HBM4 and HBM4E qualification at the major labs, the November 2026 expiry of China’s export-control suspension on gallium and germanium, and TSMC’s capital spending visibility into 2027. None of these is a cliff edge, but each is a checkpoint worth watching
The point for an allocator is simple enough. This is the largest capital cycle computing has seen, its hardest bottlenecks sit in emerging markets, and they are still tightening rather than loosening. The companies that sit on those bottlenecks, the foundries, the memory makers, the packagers, the commodity producers and the firms that keep them powered, are the ones with real pricing power for as long as the shortage lasts, and most of them are in emerging markets. Finding them takes pragmatism rather than a fixed style: a willingness to cross from the obvious growth names into the cyclical and commodity leaders that sit on the same supply chain, and the agility to move as the scarcity does.
Backing them is a way to own the scarcity the whole boom depends on, rather than paying up for the demand that is chasing it.