AI's Smartphone Decade
Josh Sookman@jsookman
−7.9%
since Oct 6
The idea
I think AI today looks like smartphones did back in 2011, with a long runway of strong compounding ahead, so this take skips the still-private application layer and sticks to infrastructure and platforms. It leans into the chokepoints that tend to avoid commoditization, like foundry and chip design, alongside the physical buildout of power, cooling and networking, then balances that with the hyperscale platforms that already look obvious today. Roughly 60% sits in infrastructure, 40% in platforms, spread across many names rather than single bets, with a small cash buffer held back.
What it owns
- The companies that own the software stack and distribution channels where AI runs at scale, the way Apple and Google owned smartphones. · MicrosoftIt owns the default productivity platform where enterprises deploy AI, through OpenAI's partnership, Azure, and Copilot embedded in office tools., AlphabetIt controls the full stack: its own chips, models, and distribution through search and Android, so it can monetize AI across the entire user experience., Meta PlatformsIt deploys AI in front of billions of people daily and converts engagement into advertising revenue, proving that consumer-scale AI monetization works., AmazonAWS is where most enterprises will run AI workloads in production, and the Anthropic stake plus its own chips give it multiple ways to win the cycle., OracleIt won a central role in OpenAI's buildout and holds major compute contracts with hyperscalers, positioning it as a platform winner if those backlogs convert to revenue.36%
- The layers in chip design and manufacturing where a single company or a few control the supply of what AI workloads need most. · NVIDIAIts AI chips remain the default way clouds and labs buy compute because its software ecosystem creates lock-in that keeps hardware from becoming a commodity., Taiwan SemiconductorNearly every advanced AI chip gets manufactured in its fabs, giving it pricing power as long as leading-edge production stays concentrated there., BroadcomIt designs custom AI chips for hyperscalers who want alternatives to NVIDIA, so it wins when companies choose to build their own silicon., ASML HoldingIt is the only manufacturer of EUV lithography machines required to print leading-edge chips, making it the most concentrated chokepoint in the entire supply chain., Micron TechnologyIt is one of only three suppliers of the high-bandwidth memory that AI models require, giving it pricing power as long as AI demand stays tight., KLA CorporationIts inspection tools catch defects at the leading edge of chip manufacturing, and every new fab that gets built purchases its gear.36%
- The physical infrastructure that has to be built regardless of which AI model wins: power, cooling, networking, and the real estate to hold it all. · Eaton CorporationIt makes the electrical equipment that connects and powers AI datacenters, a layer of demand that runs regardless of which model or chip wins., Constellation EnergyIt owns nuclear power plants that datacenters are signing long-term deals to tap, so it benefits directly as AI's primary constraint shifts to electricity supply., Arista NetworksTraining large models requires wiring tens of thousands of chips together, and it supplies the Ethernet networking that enables those clusters., Vertiv HoldingsEvery AI datacenter needs its cooling and power distribution equipment, and the density of AI workloads makes that gear more critical than ever., CoreWeaveIt is a cloud that exists purely to rent AI compute, so its heavy buildout and debt payoff only if GPU demand remains ahead of supply for years.19%
- Cash9%