−3.5%
since Sep 29
The idea
I don't think one AI lab ends up owning the whole stack, so instead of betting on a winner I own the toll roads in between. That means neutral rails like Cloudflare, MongoDB, and Datadog that get paid regardless of whose model wins, the payment and identity layer agents need to transact, a spread of chipmakers instead of one bet, and application companies rebuilding around agents. No single position dominates, and I'm holding cash on the side. I'd rethink this if one lab actually captured the entire stack top to bottom.
What it owns
- Infrastructure that AI applications need regardless of which model provider they use sits in the middle of the stack and collects fees on traffic flowing through it. · CloudflareIt sits between users and model providers, collecting fees on AI traffic flow regardless of which lab wins the underlying model race., MongoDBIts document database and vector search capabilities make it a natural foundation for any AI application regardless of which model it calls., DatadogIt monitors AI workloads in production for companies that need visibility across their models and infrastructure, earning fees as AI becomes a larger share of their computing., ElasticIt provides search and vector retrieval that most useful AI applications need before answering questions, serving that step for any model., TwilioIt owns messaging and voice channels plus customer data that agents need to reach people, earning fees on that infrastructure regardless of which model powers the interaction.30%
- When autonomous agents transact with each other or on behalf of people, they need payment rails and identity verification that work at machine speed without human friction. · CoinbaseCrypto rails can process machine-to-machine payments without requiring humans or bank accounts, and Coinbase captures that infrastructure if agents transact on blockchain., OktaIt manages identity and permissions, and agents acting on behalf of people need verification and authorization just as humans do., VisaIt processes payments over existing card networks and is building agent-specific credentials, making it the hedge if agent commerce grows on traditional rails instead of crypto., CircleIts regulated stablecoin settles instantly without intermediaries, making it the most plausible money for machine-to-machine payments.24%
- Companies building their own AI chips or buying from multiple vendors need chip designs, semiconductors, and networking that let them avoid depending on a single supplier. · Advanced Micro DevicesIt produces chips that companies adopt as Nvidia alternatives when they refuse to depend on a single semiconductor vendor for AI compute., BroadcomIt builds custom AI chips and the networking for clusters, serving hyperscalers that want their own silicon instead of standardized alternatives., Arm HoldingsIt licenses chip designs to every vendor building custom AI semiconductors, collecting royalties as the chip market diversifies away from single suppliers., QualcommIt designs chips that run inference locally on phones and laptops where no lab controls the hardware, benefiting if on-device AI becomes a reason for upgrades.22%
- Platforms that let agents interact with data, code, products, or customers need to stay relevant as the nature of work shifts from humans writing instructions to machines executing them. · ShopifyIt provides structured product data and checkout infrastructure that shopping agents need to transact with independent merchants at scale., SnowflakeIt lets companies run models against their own data without shipping it elsewhere, keeping the data warehouse as the center of gravity for AI workloads., GitLabIt manages the pipeline around code review, security, and deployment, and that infrastructure matters more when agents write and modify code at scale.14%
- Cash10%