The Agent Toll Road
Ethan Walkrman@ethanwalkerman
−7.2%
since Oct 6
The idea
I believe personal taste agents are already here, routed through convenient aggregators and billed pay-as-you-go rather than built DIY. Since those aggregators stay private, this basket owns the toll road underneath: inference chips and metered clouds, edge and API delivery, usage-billed software plumbing, and the payment rails agents spend through, weighted toward the biggest choke points. I keep a modest cash reserve to add on dips and avoid single-name bets, planning to hold for years unless the usage-based toll model itself breaks down.
What it owns
- Agents consume computing by the call and the hour, which is exactly how these cloud providers bill their customers. · MicrosoftAzure meters a huge share of AI API calls agents make and bills by consumption, the exact model this take relies on., AlphabetIt sells metered cloud capacity and its own Gemini API by the call, so it profits both as the infrastructure and the tollbooth., AmazonAWS pioneered pay-per-use computing, the consumption model agents create, and its usage-based revenue grows as agent traffic routes through aggregators., OracleIt leases raw inference capacity to AI labs whose models agents call, and its backlog must convert into sustained metered revenue., CoreWeaveIt meters GPU time for AI workloads and exists only for that purpose, making it the purest bet on sustained agent-driven inference demand.26%
- Every API call an agent makes runs on inference chips somewhere in the stack, and these companies either build or manufacture them. · NVIDIANearly every agent API call runs on its chips somewhere in the inference stack, and its data center revenue grows as usage scales., Taiwan SemiconductorEvery inference chip serving agent traffic, whether from NVIDIA, AMD, or custom designs, gets manufactured at its foundries., BroadcomIt designs custom inference chips for the clouds and provides the networking gear that ties their data centers together., Advanced Micro DevicesAs agent workloads compound, clouds need a second chip supplier to break single-vendor lock, and it supplies that competition in real deployments.25%
- Agents generate billions of small API requests that travel over networks, and these providers route, secure and charge for that traffic. · CloudflareIt routes and secures agent API traffic at the edge while increasingly charging for machine-to-machine requests., AkamaiIt delivers and secures the flood of API calls agents generate, charging for exactly that in its API security and compute segments., TwilioIt pioneered the pay-per-API-call business model, and agents that message, call, or verify users spend through it by the request.13%
- When everything runs pay-as-you-go, every transaction must be metered and logged, and these platforms charge by consumption to do it. · DatadogPay-as-you-go systems require constant measurement, and it is the platform companies pay to meter their API usage., SnowflakeIt bills by consumption for data storage and serving, the same pay-as-you-go model this take depends on, as agent workloads drive usage up., MongoDBIts Atlas database is usage-billed and consumption-grows with AI applications, attracting builders who want managed infrastructure instead of self-hosting.13%
- Agents that pay per API call need infrastructure to move money per transaction, and these networks are building that capability. · VisaAgents that spend per API call need payment rails, and it is building agentic payment products to route that transaction volume., CoinbaseIt provides a free, decentralized payment option with stablecoin and per-transaction infrastructure for agents to pay without card networks., MastercardIt is a rail for machine-to-machine spending, taking a small cut when agents pay for services on behalf of their owners.13%
- Cash10%