Builders Keep Winning
Erwin Blom@erwblo
−5.3%
since Sep 30
The idea
I believe the companies already building and shipping frontier AI models keep extending their lead over everyone still learning to use them, since they control the models, the compute, and the customer relationships. So this basket owns three layers: the public owners of frontier models, the chipmakers and fabs supplying the compute those labs burn through, and the platforms already turning models into products people pay for, plus cloud and data-center infrastructure underneath it all.
What it owns
- These companies either own frontier models or large stakes in the labs building them, positioning them to capture value as AI capabilities advance. · MicrosoftOperates both as a model owner through its stake in frontier AI development and as a distribution channel through Copilot and Azure, positioning itself to capture value across the AI stack., AlphabetIt builds Gemini, runs it on its own chips, and distributes it through Search, Android, and Cloud, creating an integrated full stack if its models stay at the frontier., Meta PlatformsIt runs its own frontier lab and puts those models in front of billions of users across its apps, betting that AI spend improves ads and new products rather than becoming a cost center., AmazonIt holds a large stake in Anthropic and rents compute through AWS, positioning itself to benefit if Anthropic stays a top lab and AI workloads keep landing on its cloud.32%
- Frontier labs spend billions on chips and infrastructure, and these companies collect that spending across the entire race regardless of which lab leads. · NvidiaIt supplies the chips every frontier lab burns cash on to train and run models, so it collects across the entire race regardless of which lab wins as long as compute budgets keep growing., Taiwan SemiconductorIt manufactures AI chips for Nvidia, AMD, Broadcom, and the big platforms' custom silicon, collecting across multiple designers if chip demand stays strong., BroadcomIt builds custom AI chips for platforms designing their own silicon and the networking that ties data centers together, benefiting when builders want alternatives to Nvidia., Advanced Micro DevicesIt serves as a second source for labs and clouds seeking alternatives to Nvidia, capturing meaningful budget if its chips stay competitive and compute spending keeps accelerating., Micron TechnologyIt makes high-bandwidth memory that AI chips are starved for, and pricing holds as long as labs keep scaling models and memory remains a bottleneck.31%
- Software companies already selling AI products and agents to customers who need help deploying models, turning frontier capabilities into revenue. · Palantir TechnologiesIt sells the layer that deploys models inside enterprises and governments, capturing value from customers who lack the expertise to implement AI on their own., ServiceNowIt ships AI agents inside workflow software that large companies already run, turning frontier models into products through existing customer relationships., SalesforceIt pushes Agentforce through the world's largest sales and service software base, distributing models through channels that model builders cannot reach on their own., AppLovinIt operates an AI ad engine that already generates revenue today, collecting across mobile and beyond if its targeting continues to beat competitors.15%
- Models require somewhere to run, and the cloud providers and infrastructure makers renting that capacity are operating at or near full capacity. · OracleIt has become the landlord for frontier labs, signing large contracts to host their training compute, and collects if those deals convert to sustained revenue., Vertiv HoldingsIt supplies the power and cooling gear every AI data center needs, riding the buildout without betting on a specific chip winner as long as capacity expansion continues., CoreWeaveIt rents Nvidia compute to AI labs exclusively, making it the most direct bet on labs burning cash on capacity while also swinging harder than most positions.13%
- Cash9%