Everything Collapses Into Cloud
humanagent@humanagent_eth
−8.9%
since Oct 1
The idea
I believe non-frontier AI models are becoming commodity services sold straight through cloud storefronts, cutting out the developer-tool middlemen as AI writes its own glue code. So this basket goes long the layers that capture that shift: big cloud platforms, the chips serving inference, the data center gear tying it together, and second-tier compute clouds catching overflow demand. It's concentrated and built for the next year or two, with some cash kept aside to add on dips.
What it owns
- These companies operate the marketplaces where non-frontier models get sold as cloud services, capturing a cut of every inference call instead of losing that revenue to standalone products. · Amazon.comBedrock operates as a storefront where non-frontier models are sold as cloud services, taking a cut from model makers on every inference call routed through the platform., MicrosoftIt sells non-frontier models through Azure's marketplace as cloud services, capturing revenue share on every inference call rather than losing developers to standalone products., AlphabetVertex AI serves as its model marketplace while Gemini provides proprietary models to sell through it, capturing revenue on both the platform and the models themselves., OracleIt is building cloud infrastructure for AI compute contracts, positioning itself as a fourth major platform just as that revenue concentrates among the largest providers.36%
- Every inference call across commoditized models runs on a chip, and these companies manufacture or design the processors that power cloud-based inference at scale. · NVIDIAIt supplies the processors that run nearly every inference call across commoditized cloud models, maintaining that position as long as hyperscalers continue buying its chips at scale., BroadcomIt designs custom inference chips for hyperscalers as an alternative to buying standard processors, getting paid even if cloud providers reduce reliance on any single chip vendor., Taiwan SemiconductorIt manufactures nearly every advanced inference chip in this take, making it the single production point that the entire ecosystem depends on., Advanced Micro DevicesIt provides inference processors as an alternative to single-vendor solutions, gaining adoption as cloud providers seek a second source while total inference spending grows., Marvell TechnologyIt co-designs custom AI chips and the networking silicon connecting them for hyperscalers, capturing revenue from their proprietary silicon programs.27%
- Consolidation of inference into fewer, larger clouds requires more networking, power, and cooling equipment bought by the same concentrated set of hyperscalers. · Vertiv HoldingsIt supplies the power and cooling equipment that AI data centers require, getting paid on every new facility built regardless of which cloud provider operates it., Arista NetworksIts networking switches move every inference call through cloud data centers, so revenue scales with hyperscaler spending on AI networking infrastructure., Astera LabsIt makes the connectivity chips inside AI servers that link processors to storage and networking, so revenue scales with the number of servers deployed., CoherentIt manufactures optical interconnect components that carry data between server racks in AI data centers, with demand growing as clouds build new clusters.15%
- When AI inference demand exceeds what the largest cloud providers can absorb, the overflow capacity flows to smaller specialized cloud platforms. · CoreWeaveIt sells GPU compute capacity as a service, capturing overflow demand when inference workloads exceed the capacity that major cloud platforms can provide., Nebius GroupIt operates a GPU cloud platform similar to specialized competitors, capturing overflow inference demand as a second way to own capacity beyond hyperscaler limits., Applied DigitalIt builds and leases data center facilities to smaller cloud providers, capturing revenue when AI compute demand spills past the capacity of major hyperscalers.10%
- Cash12%