AI toll roads stocks are the picks-and-shovels play Luke Lango built around a five-word framing: “AI just joined the payroll.” When AI was a tool you typed prompts into, the compute cost was a rounding error for most companies. When AI becomes an agent that does work — answers customer emails, processes insurance claims, writes code, executes trades — the compute cost becomes a labor cost. And labor costs show up on every line of the income statement.
Lango published this framework in July 2026 through InvestorPlace. He called the infrastructure stack that handles this traffic “AI toll roads.” The metaphor is direct: every AI agent, regardless of which model powers it, sends its traffic through the same physical infrastructure. The companies that own that infrastructure collect the toll. For context on Lango’s career and track record, his dossier covers the path from Caltech quant to ranked #1 stock picker.
The Demand Driver: Agentic AI and AI Toll Roads Stocks
A chatbot query and an agentic workflow are different animals. A chatbot answers a question and stops. An agent runs a multi-step process: it reads context, makes decisions, calls external tools, writes output, and loops back to verify. A single agentic task can consume 5 to 30 times more compute tokens than a one-shot query, according to Gartner estimates Lango cited.
Goldman Sachs projected monthly token counts for agentic AI could reach 120 quadrillion by 2030. The number is staggering and hard to picture. The practical translation: the data centers being built right now are sized for a workload that is about to multiply by an order of magnitude.
The infrastructure does not care which model wins. OpenAI, Anthropic, Google, Meta — they all run on the same silicon, the same memory, the same networking, the same power. The toll road collects from every traveler.
The Six Toll Booths
Lango broke the infrastructure stack into six categories. Each one is a layer where the agentic AI traffic physically passes through.
Accelerators. The GPUs and custom silicon that run inference. Nvidia owns the training market and most of the inference market. AMD and custom silicon from the hyperscalers are gaining share on the inference side, where efficiency matters more than raw throughput. Every agentic task starts here.
Networking and Custom Silicon. The connections between chips, between servers, between racks. Credo Technology builds the high-speed connectivity gear that lets AI clusters talk to each other in real time. Marvell and Broadcom design custom silicon alongside the hyperscalers. The networking layer is where latency lives, and agentic workflows are latency-sensitive in ways that chatbot queries are not.
Memory. High-bandwidth memory is the bottleneck that gets less attention than chips but is equally critical. SanDisk, Micron, and SK Hynix supply the HBM that keeps the accelerators fed. Edge devices need fast, low-power memory to run inference locally. Memory cycles have their own supply and demand dynamics, and the agentic AI demand wave is hitting a memory market that was already tightening.
Servers and Power. The physical boxes that hold the chips and the power systems that keep them running. Dell Technologies has become the premier AI server play, with AI-optimized server revenue growing 757% year over year. Power management chips from Monolithic Power and Analog Devices determine how efficiently a server uses its electricity. The power constraint is the real ceiling on AI deployment — you can only build a data center as fast as the grid can supply it.
Optical Connectivity. The technology that moves light instead of electricity. Coherent, Lumentum, and Corning build the optical links that connect data centers to each other and to the broader internet. As agentic workflows cross enterprise boundaries — a development Lango explores in his DNSid thesis — the optical backbone becomes the long-haul toll road.
Storage. The layer Lango flagged as overlooked. Agentic AI generates enormous amounts of intermediate data: context windows, decision logs, tool outputs, verification records. Pure Storage and the broader storage stack handle this. Storage gets treated as a commodity in most AI infrastructure analysis. Lango’s argument is that agentic workloads change the storage equation because the data is active rather than archival, constantly referenced, and latency-sensitive.
The JPMorgan Asterisk
In a separate July 2026 article, Lango flagged a development that expands the toll roads thesis. JPMorgan Chase selected SambaNova Systems as its inference-infrastructure partner, deploying on-premises AI hardware inside the bank’s firewall.
The mainstream AI infrastructure thesis assumes inference demand flows through hyperscaler cloud platforms. JPMorgan’s decision points to a segment the cloud-first narrative underweights: enterprises that cannot send their most sensitive data to a third-party server. Banks hold client data and proprietary trading strategies. Hospitals manage patient records. Defense contractors face outright restrictions on running sensitive workloads on commercial cloud.
For these organizations, the toll road runs inside their own buildings. SambaNova went from a rumored $1.6 billion acquisition target to raising $1 billion at an $11 billion valuation in under a year. Private capital decided that secure, on-premises enterprise AI inference is a durable market.
The inference market is splitting in two: hyperscaler cloud captures the majority of demand, while regulated industries form a structurally distinct second market. The toll roads collect from both.
Where Storage Fits
The storage layer deserves its own dimensionalization because it is the part of the thesis most likely to surprise readers.
The standard AI infrastructure narrative is chip-centric. Nvidia makes the GPU, the GPU runs the model, the model generates the output. Storage sits in the background as a filing cabinet.
Agentic AI breaks that framing. An agent reads a document, extracts information, makes a decision based on that information, writes a new document, and logs the entire process for audit. Every step generates active data that the system needs to reference again. The storage functions as working memory, not as a filing cabinet.
Lango’s argument is that storage companies are positioned to benefit from this shift in ways the current valuation models do not fully capture. The memory cycle has its own supply dynamics, and the agentic AI demand wave is hitting a storage market that was already tightening from the cloud AI buildout.
The Toll Road Economics
The economic logic of a toll road is that the upfront cost to build it is enormous, the ongoing maintenance is manageable, and the traffic that flows through it pays for the capital over time. The operator does not care what cargo the trucks are carrying. The toll is the same whether the truck is hauling groceries or electronics.
AI infrastructure works the same way. The data centers cost billions to build. The chips, memory, and networking gear inside them cost millions per rack. Once built, the infrastructure collects revenue from every workload that passes through it, regardless of which AI model generated the workload.
The risk is utilization. A toll road that nobody drives on is a very expensive piece of pavement. If agentic AI adoption is slower than projected, the infrastructure buildout overshoots demand and the toll operators face pricing pressure. The hyperscalers are betting tens of billions that the traffic will come. Lango’s thesis is that the traffic is already arriving — it is just arriving in a form most investors have not priced in yet.
The Layered Thesis
Lango’s toll roads framework sits inside a stack of related theses. The Genesis Mission covers the government capital seeding the cloud and data center layer. Physical AI covers the edge — the chips, sensors, and memory that run inference on devices. The toll roads thesis covers the infrastructure in between: the servers, networking, power, and storage that connect the cloud to the edge.
The three layers describe the same buildout from different angles. Government money seeds the cloud layer, agentic workloads strain the infrastructure layer, and Physical AI pushes compute to the edge. The companies that supply components across all three layers are positioned to benefit from the full cycle.
The toll road is the middle layer, and it is the layer where the traffic is densest. Every agentic task, whether it originates in a hyperscaler data center or runs on a factory robot, passes through the infrastructure stack. The toll collects from every traveler.
Lango’s framing — compute as the new labor cost — is the sticky version of this idea. When AI was a tool, compute was an IT expense; when AI is a worker, compute becomes a payroll line item, and payroll lines are the ones companies optimize, automate, and grow over time rather than cut. The AI data center power bottleneck is the physical constraint that determines how fast this toll road gets built. The toll road operator collects from every one of those transactions.