Post by VistaShares

Everyone in AI is hunting for the next bottleneck. The smarter move is to stop guessing which component binds next, and follow the capital instead. Two years ago the bottleneck was advanced packaging. TSMC's CoWoS capacity was sold out, and the entire buildout waited on it. Then it was memory, as high bandwidth memory became the tightest component in the stack and sold out through 2026. Now it is power. You can take delivery of an AI accelerator faster than you can source a grid transformer, which runs three to four years out. The bottleneck keeps moving. The capital behind it keeps compounding. Hyperscaler capex is on track to clear 600 billion dollars in 2026, up from roughly 400 billion in 2025. McKinsey models 5.2 trillion dollars of AI data center investment by 2030. That money flows through the same layers every cycle: compute and memory, foundry and packaging, lithography, then networking, power, and cooling. The names that bind change. The map does not. Here is what makes this a strategy and not a guess. At each chokepoint, value capture is extreme. TSMC runs near 60 percent gross margins. ASML holds 100 percent of the market for EUV lithography. SK Hynix recently reported a 72 percent operating margin on memory that is sold out through 2026. Nvidia sits near 75 percent gross margins. These are not commodity layers competing on price. They are toll booths, and the toll gets collected no matter which lane is jammed. The lesson runs across the whole stack. The chokepoint binding today was not binding two years ago, and it will not be in two more. The layers that capture the economics keep capturing them, cycle after cycle. Owning the spend beats owning your guess about the spend. Don't bet on where you think the next bottleneck is. Follow the capex. This is the thesis we have been building at VistaShares, across the layers instead of betting on any one of them. What layer do you think binds next? #AIinfrastructure #semiconductors #investing DVx Ventures Adam Patti