The Compute Behind AI Coding
Noel@noel
−10.7%
since Oct 10
The idea
I believe AI coding tools are becoming must-haves for developers, but the winning tools themselves are hard to own directly, so instead I own the compute underneath all of them. The basket leans toward custom chips and the clouds renting that capacity, since that's where I think the biggest weight sits, with networking and memory filling out the build-out and Nvidia as an anchor rather than the whole bet. I'm keeping a cash reserve to add if prices fall, running this over the next couple of years, and I'd rethink it if AI companies start pulling back on data center spending.
What it owns
- As demand for AI coding tools grows, the cloud platforms that host and serve these tools see more customers renting computing power. · Alphabet Inc.Operates a cloud platform renting custom-designed processors to AI companies, while running its own coding tools on the same chips., Amazon.com, Inc.Hosts a leading AI coding assistant on its cloud and manufactures custom processors for it, capturing both hosting fees and chip revenue., Microsoft CorporationRents AI computing capacity on its cloud platform to model makers and operates a coding assistant on that platform., Oracle CorporationRents AI computing capacity on its cloud platform under long-term contracts, betting that AI customers will sustain high cloud spending., CoreWeave, Inc.Rents graphics processors to AI companies that need more computing power than they build internally, directly capturing demand for AI infrastructure expansion.30%
- The major cloud providers and AI companies design specialized chips for coding workloads, and these companies either design them or manufacture them at scale. · Broadcom Inc.Co-designs the processors that a major AI coding platform runs on, earning revenue each time developers use that platform., Taiwan Semiconductor Manufacturing Company LimitedManufactures nearly all advanced AI chips in the take, from graphics processors to custom cloud chips, winning regardless of which processor architecture dominates., Marvell Technology, Inc.Designs custom chip components and networking links for cloud providers building their own AI infrastructure.25%
- Thousands of AI processors must communicate simultaneously across data centers, driving demand for specialized switches, cables and optical components. · Arista Networks, Inc.Sells networking switches that connect thousands of AI servers within each data center, growing as clouds add capacity., Credo Technology Group Holding LtdManufactures high-speed cables that link AI servers within a data center rack, scaling with each new computing cluster., Astera Labs, Inc.Manufactures connector chips that enable communication between processors and memory inside AI servers, appearing in every new system., Lumentum Holdings Inc.Manufactures lasers and optical components that transmit data between AI servers, including in custom chip systems.15%
- Graphics processors remain the workhorse for most AI training and inference work, so these chipmakers capture a large share of the underlying compute demand. · NVIDIA CorporationSupplies graphics processors that power a large share of AI coding models and infrastructure, serving as a foundational bet across multiple data centers., Advanced Micro Devices, Inc.Supplies graphics processors as an alternative to the market leader, with growing adoption among cloud providers and AI companies seeking a second source.12%
- Modern AI chips require massive amounts of high-speed memory installed directly on the processor, creating a supply-constrained bottleneck. · Micron Technology, Inc.Manufactures the high-speed memory that sits on every advanced AI chip, with demand outpacing supply across the industry.8%
- Cash10%