Open Models Eat the Moat
Miles Beckett@mbeckett
−7.8%
since Oct 1
The idea
I think open source AI models are getting good enough to break the closed labs' pricing power, and the value shifts to what sits on either side of them. So I built exposure to the compute and infrastructure serving the coming flood of cheap inference, plus companies whose AI costs drop as model prices race toward zero, with a smaller stake in the open model champions driving this shift. I spread bets across many names rather than leaning on one, kept a cash reserve to add on dips, and I'm giving this two to three years to play out.
What it owns
- Free models drive far more usage, and every inference query runs on chips that someone must purchase and deploy. · Advanced Micro DevicesIf open models win, developers lose lock-in to proprietary software stacks, and AMD becomes the cost-effective alternative for serving inference., NVIDIAEven as free models weaken pricing power, the volume of cheap inference deployed still requires most of the chips that run it., Taiwan SemiconductorWhoever wins the chip competition, TSMC manufactures the silicon, and an explosion of inference drives demand for wafers across all architectures., BroadcomWhen models are free, large cloud providers design custom chips to serve them cost-effectively, and Broadcom manufactures those specialized chips., Marvell TechnologyIt supplies the custom interconnect chips that power high-volume inference clusters, benefiting from the shift toward custom silicon.28%
- Open models run anywhere, so the servers, networking, and hosting infrastructure that powers them benefits regardless of which model leads. · CoreWeaveIt rents out the compute infrastructure that serves models, and free models increase the variety of workloads customers want to host., OracleIt has no proprietary model to protect, so it can host whichever model is cheapest and most suitable for customer workloads., Dell TechnologiesIt manufactures servers at scale for enterprises deploying open models in their own data centers as inference demand surges., Super Micro ComputerIt builds the racks and systems that inference runs on and moves quickly to integrate new chip architectures as they arrive., Vertiv HoldingsAll that inference generates substantial heat, and Vertiv sells the power management and cooling infrastructure data centers require., Arista NetworksIt connects the chips inside AI clusters and benefits from cloud networking expansion regardless of whether the underlying chips come from Nvidia, AMD, or custom designs.26%
- Software platforms that embed AI gain access to capable intelligence at collapsing cost, which expands their operating margins. · SnowflakeIt lets customers run open models next to their data without paying a closed lab margin, turning raw intelligence into margin for itself., DatadogWhen companies run open models in production, they need observability tools to monitor those systems, which Datadog provides., Palantir TechnologiesIt substitutes the cheapest available models into its platform while charging for the deployment and integration layer above them., ShopifyIt embeds AI across millions of merchant stores, and free models make those features nearly free to operate while customers pay for the platform., ServiceNowIt bakes AI into workflow software and keeps the margin that a closed lab would have extracted, improving its own profitability.24%
- Companies that distribute open models make money when free, capable models become the foundation everyone builds on. · Meta PlatformsIt distributes competitive open models for free, capturing the intelligence benefit in its own products while pricing power shifts away from closed labs., IBMIt sells open models and services to enterprises that choose free, capable models to avoid paying a closed lab's pricing premium.12%
- Cash10%