The AI supply chain — and where it chokes
Every AI accelerator rides on a stack of upstream suppliers — substrates, lasers, optics, packaging, power — and several of those layers are near-monopolies. His method is to ignore NVIDIA itself and buy the narrowest chokepoint that feeds it.
This universe is about the physical buildout behind the AI boom. Training and running large AI models needs enormous data centers, and the bottleneck is increasingly how fast chips talk to each other. The answer is light: moving data over fiber, using lasers and "silicon photonics" packaged right next to the processor (co-packaged optics). Instead of betting on NVIDIA itself, this approach buys the "picks and shovels" upstream of it, the obscure suppliers of substrates, lasers, foundry capacity, packaging, testing, fiber and power that everyone depends on. The edge is finding supply-chain chokepoints, small companies that are the only or cheapest source of one critical input.
From raw crystal to running GPUs
Thirteen layers, bottom to top. Each is a bottleneck; the names on the right are the ones he actually tracks at that layer — sized by conviction, ringed if he likely owns them.
The demand that pulls the whole chain. He rarely buys this tip — he buys the narrowest chokepoint upstream of it.
Compound-semiconductor substrates (InP / GaAs / SiC)
The raw crystal wafers that lasers and high-frequency chips are grown on. Ordinary silicon can't emit light efficiently, so photonics needs exotic materials like indium phosphide (InP) and gallium arsenide (GaAs); power electronics needs silicon carbide (SiC). Very few firms can make these crystals to spec, so they are a true chokepoint.
Light source: CW DFB / EML lasers
Optical links need a laser to create the light that carries data. Continuous-wave (CW) DFB lasers and EML lasers are the light 'engines.' As the industry shifts to external light sources for co-packaged optics, merchant laser suppliers become a critical, supply-constrained layer.
Optical transceivers (800G / 1.6T)
The modules that convert electrical signals to light and back at the ends of a fiber link, now ramping to 800-gigabit and 1.6-terabit speeds. They package lasers, optics and chips together and are the visible product hyperscalers buy by the millions.
Fiber array units (FAU) & precision optical assembly
A fiber array unit precisely aligns many optical fibers to a photonic chip, a fiddly, high-tolerance step that's hard to do at volume. This alignment and packaging work is a recognized mass-production bottleneck for co-packaged optics.
Co-packaged optics (CPO) integration
Instead of plugging transceivers into the front of a switch, CPO places the optics on the same package as the processor, cutting power and boosting bandwidth. It's the architecture NVIDIA and others are pushing, and it pulls every layer below it together.
Advanced packaging / OSAT
OSAT firms (outsourced assembly and test) stack and bond chips, memory and optics into finished packages, including glass-core substrates and SiC interposers. As transistor shrinks slow, packaging is where performance gains now come from, so capacity here is scarce.
Data-center power & grid
All of this is gated by electricity. AI data centers need enormous, reliable power, which strains the grid and creates demand for transformers, switchgear, power semiconductors and generation. It's the 'boring' bottleneck that can cap the whole buildout.
Glossary
Technology that builds optical components (waveguides, modulators) onto silicon chips so data can move as light instead of electrical signals. It lets chips communicate far faster and with less power than copper wires.
Putting the optical engine on the same chip package as the processor or switch, rather than in a pluggable module at the edge. This shortens the electrical path, saving power and raising bandwidth, and is a central architecture for next-gen AI systems.
A compound semiconductor crystal used to make high-speed lasers and detectors for optical communication. Silicon can't emit light well, so InP is a critical upstream material for photonics.
A compound semiconductor used for radio-frequency chips and certain lasers/LEDs. Like InP, it's made by a small set of substrate and foundry specialists.
A wafer with a thin silicon layer sitting on an insulating oxide layer. It's the preferred base for silicon-photonics chips because it confines light efficiently; supply is dominated by very few makers.
Epitaxy is growing ultra-thin, precise crystal layers on a wafer; the result is an 'epiwafer.' These layers give lasers and transistors their electrical and optical properties.
A precision method of depositing crystal layers one atomic layer at a time, used to make epiwafers for lasers and advanced devices. The machines come from a near-duopoly of equipment makers.
A continuous-wave distributed-feedback laser, a stable single-color light source used as the 'engine' that supplies light for optical links and co-packaged optics.
A module that converts electrical data into light to send down a fiber and converts received light back into electrical signals. Data centers buy these by the millions.
Data rates of an optical link: 800 gigabits and 1.6 terabits per second. Each generation roughly doubles speed, and the industry is ramping from 800G toward 1.6T for AI clusters.
A precision component that aligns and holds multiple optical fibers against a photonic chip so light couples in and out correctly. Doing this accurately at high volume is a known bottleneck.
Outsourced semiconductor assembly and test: contract firms that package finished chips and test them. As packaging becomes the key to performance, OSAT capacity is increasingly scarce.
Techniques for stacking and bonding multiple chips, memory and optics into one dense package (e.g. CoWoS, interposers). It now drives most performance gains as classic transistor shrinking slows.
A next-generation chip packaging base made with a glass core instead of organic material, offering flatter, denser, higher-performance interconnects. Specialized laser tools are needed to drill and pattern the glass.
Memory chips stacked vertically and placed right next to an AI processor to feed it data extremely fast. HBM is in tight supply and a core part of the AI-memory thesis.
The main fast working memory in computers and AI accelerators. AI demand has tightened DRAM supply and lifted prices, a pillar of the 'memory supercycle' idea.
Flash storage memory used in SSDs. AI data pipelines need vast storage, driving a structural demand surge and repeated price increases.
The thesis that AI demand for DRAM, HBM and NAND structurally outruns supply for years, letting memory makers raise prices and expand margins well beyond a normal cycle.
Precise measurement and inspection of chips and packages during manufacturing to catch defects and control yield. New optical and stacked-memory parts create demand for new metrology tools.
A new breed of specialized cloud provider that rents out GPU computing capacity for AI, competing with the big hyperscalers, often built around access to power and chips.
Gallium-nitride and silicon-carbide power semiconductors that handle electricity more efficiently than silicon. They're key to delivering the huge, high-voltage power AI data centers need.
A harmonic drive is a compact, high-precision gear used in robot joints. It's central to humanoid robots like Tesla's Optimus, the anchor of the robotics supply-chain thesis.
A factory that manufactures chips designed by other companies. The leading foundries are the indispensable center of the whole chip supply chain.
A chip company that designs chips but outsources all manufacturing to foundries, keeping costs low and focusing on design and IP.
General technical explanation, not investment advice. See the full universe or the learn guides.