Edge Intelligence Compounders
Uncle Vibecoder
@leo_guinan
+1.9%
since Sep 29
The idea
I believe as intelligence gets cheaper, the value shifts away from the biggest model builders and toward the edge, where models are tuned, cached, and used, with the solved-work byproducts becoming reusable assets over time. So the basket leans heaviest into caching and delivery, backed by edge silicon, high-volume software companies piling up that solved-work archive, and a small sleeve cutting the cost of running models locally, with cash held on the side. No single position dominates, and I'd rethink this if the economics stopped favoring proximity to the user over raw model scale.
What it owns
- These networks sit between models and users, winning as tuning and serving intelligence moves out of central data centers. · CloudflareIt runs inference and stores model state in hundreds of cities, positioned where intelligence gets used rather than in central clouds., Akamai TechnologiesIt has the oldest and widest delivery network and is repurposing it for edge compute and inference serving., EquinixIt owns the buildings where edge networks, local clouds, and enterprises physically meet, collecting rent as inference moves outward., DigitalOceanIt lets ordinary builders tune and run models without a hyperscaler, positioned as the trusted small cloud close to the edge., FastlyIt offers programmable compute at the cache layer, exactly where the cost of updating and serving models gets decided.34%
- Cheap on-device and near-device chips make running and updating models at the edge economical in the first place. · QualcommIt makes the chips that let phones and laptops run tuned models locally, the hardware foundation for intelligence moving to where it is used., Arm HoldingsIt licenses chip designs to nearly every edge device manufacturer, so it collects royalties on volume regardless of which chipmaker wins in the market., NXP SemiconductorsIt puts processors in cars, factories, and payment terminals where applied intelligence meets people who cannot build it themselves., AmbarellaIts chips run vision models inside cameras and cars with no cloud round trip, betting that seeing and deciding happens on device., Lattice SemiconductorIt sells tiny reprogrammable chips that do inference on almost no power, making tuning and updating models at the edge cheap.26%
- Companies closing millions of tickets and merges accumulate solved-problem archives that become reusable assets. · ServiceNowIt sits on years of solved workflows and is turning that archive into AI agents that resolve work automatically., AtlassianIt holds millions of resolved tickets and project histories that become reusable search paths for how teams fixed problems., DatadogIt collects telemetry of everything its customers run and sells intelligence back as monitoring and AI-driven fixes built on that data., GitLabIt owns the full history of how code got written, broken, and fixed inside its customers, which its AI assistant can reuse.21%
- Memory and power-efficient components quietly drive down the cost curve that keeps making local intelligence cheaper. · Micron TechnologyRunning models locally requires substantial memory in every phone, laptop, and edge server that it supplies., Monolithic Power SystemsIt makes the power management chips that let dense compute run efficiently in small spaces, keeping the cost of running models locally low.11%
- Cash8%