Custom Silicon Bottleneck
Ben Furlong@0xbendo
−1.3%
since Oct 16
The idea
I think the real bottleneck in AI isn't power or data centers, it's chip production, especially custom silicon built for each hyperscaler. So this basket is anchored in the designers behind those custom chips, plus the toll collectors who get paid on every design regardless of who wins: design software, licensed blueprints, the foundry and packaging step. I also hold memory and storage, since HBM and NAND supply looks sold out years ahead, along with CPU designers riding the same wave.
What it owns
- The hyperscalers designing custom AI chips need specialized partners to co-design them, and the companies in this group sit at the center of that work. · BroadcomIt co-designs nearly all hyperscaler custom AI chips, placing it at the center of the custom silicon buildout., Marvell TechnologyIt co-designs custom chips for major hyperscalers and benefits directly from the shift toward custom silicon over off-the-shelf parts., Astera LabsIt builds the connectivity that links custom accelerators to each other, so every new cluster design creates demand for its chips., Credo TechnologyIt makes the cabling that links chips inside AI clusters, a chokepoint as cluster buildouts demand faster connections between components.25%
- High-bandwidth memory and data center storage are already sold out years in advance, making them genuine supply constraints as AI workloads scale. · Micron TechnologyIt makes the high-bandwidth memory that every AI accelerator needs, and supply is already sold out years ahead., Western DigitalIt supplies data center storage where AI output has to land, and demand is outrunning storage maker supply., SandiskIt makes NAND flash storage, the layer that data centers need to absorb the output generated by all this custom silicon., Seagate TechnologyIt makes data center storage riding the same squeeze as AI workloads generate more data than storage supply can absorb., RambusIt makes the interface chips between processors and memory, collecting fees as memory speeds and volumes climb with agentic workloads.21%
- Every new chip design pays for design software and licensed blueprints, so more custom chips created means more royalties collected regardless of who designs them. · SynopsysIt sells the design software that custom chip teams pay for before they build anything, making it a toll on growing design volume., Cadence Design SystemsIt sells design software that every chip team pays for, collecting a toll on each new custom chip project regardless of the designer., Arm HoldingsIt licenses chip blueprints that most custom CPUs and many accelerators start from, collecting royalties on every new design.18%
- AI accelerators need capable host processors beside them, and the CPU designers gaining data center share will ride the same buildout wave. · Advanced Micro DevicesIt designs CPUs that data centers pair with AI accelerators, and gaining share in that socket matters as AI clusters expand., QualcommIt brings chip design expertise into data center CPUs and edge devices that will run autonomous agents., IntelIt sells CPUs for the agent era and operates advanced fabs on US soil, adding manufacturing capacity to a supply-constrained chain.14%
- Advanced chip manufacturing and the specialized packaging that follows are the narrowest points in the supply chain, controlled by a handful of companies. · Taiwan SemiconductorIt physically manufactures nearly every advanced custom chip, making it the purest expression of production as the bottleneck., Amkor TechnologyIt does advanced packaging, a step that has become one of the tightest chokepoints in getting custom chips out the door.12%
- Cash10%