GPU Shortage Outlasts the Pullback
Israel Derdik@iderdik
−13.6%
since Oct 9
The idea
I believe AI adoption keeps growing even as the hype-driven waste fades, and demand for AI chips still outpaces supply, especially in high-bandwidth memory and advanced packaging, both sold out into 2027. I built this around the companies that keep winning once supply catches up, chip designers, foundry, and toolmakers, with smaller stakes in memory makers and GPU renters who benefit most while things stay tight, plus a cash buffer. I would rethink this if new memory and packaging capacity comes online faster than demand grows.
What it owns
- Companies designing AI chips sell everything they can make while supply stays short. · NVIDIA CorporationIts GPUs remain the most sought-after chips for AI applications, with design advantages and software integration that typically sustain market share as supply-demand dynamics normalize., Broadcom Inc.It designs custom AI chips for big cloud providers, and those chips need the same scarce memory and packaging, so it profits no matter whose design wins., Advanced Micro DevicesWhen Nvidia's chips are hard to get, large buyers turn to it as the second source, so a long shortage feeds its GPU sales.23%
- Whoever builds the new memory and packaging capacity has to buy these machines, so they win whether the shortage lasts or eases. · Applied MaterialsMemory makers need more equipment to build high-bandwidth memory capacity, so it wins either way a long shortage plays out., Lam ResearchIts etch and deposition tools are used to stack and connect the layers in high-bandwidth memory, so more capacity means more orders., TeradyneIt tests high-bandwidth memory and AI chips before they ship, and more complex chips need more testing time., Onto InnovationIt makes the inspection and measurement tools that catch defects in advanced packaging, which gets harder as more memory is stacked onto each chip., Kulicke and Soffa IndustriesIt makes the bonding machines that attach chips together in advanced packaging, so every new packaging line is a potential order.19%
- Bonding memory onto AI chips is the tightest step in the supply chain, and only a few companies can do it at scale. · Taiwan Semiconductor ManufacturingIt runs CoWoS, the packaging step that bonds high-bandwidth memory onto GPUs, which is fully booked and the tightest part of the AI chip supply chain., Amkor TechnologyIt picks up advanced packaging work that others cannot fit in, so it earns most while packaging stays the bottleneck., ASE Technology HoldingIt is the largest outside chip packager and tester, and gets more advanced packaging work when others' lines are full.18%
- Cloud providers that rent out GPUs by the hour can charge more when nobody can buy chips fast enough. · Oracle CorporationIt rents out GPU capacity to AI companies and earns most while GPUs are scarce., CoreWeaveIt does nothing but rent out GPUs, so it gains the most when they are scarce and feels it first if rental prices drop., Nebius GroupIt is a fast-growing GPU rental cloud that benefits as AI use spreads past the biggest labs., IREN LimitedIt rents out GPUs from sites where it already controls the power, and power is another thing new AI data centers are waiting on.17%
- Every AI chip needs stacks of special memory that is sold out, so memory prices and profits jump while the shortage lasts. · Micron TechnologyIt is the only US maker of high-bandwidth memory, the part that is now the real shortage, with supply sold out for years., Sandisk CorporationIt sells the flash storage AI data centers fill up with, and it profits most while memory supply stays tight.13%
- Cash10%