Samsara says it has AI on 99% of US roads. Its CEO says no model can catch up.

Samsara (NYSE: IOT) puts sensors on trucks, engines, and industrial equipment the company says touch 99% of US roads every day and ingest 25 trillion data points a year. The interesting part isn't the hardware. In a Podcast Alpha interview, CEO Sanjit Biswas argues that as foundation models from OpenAI, Anthropic, and Google converge on roughly comparable capabilities, the durable AI moat isn't the model. It's the continuous stream of physical-world data those sensors keep collecting.

The proof point is Agents Studio, which Samsara shipped in late June 2026. The product compresses warranty claims from 1-2 hours to under a minute, according to Biswas, by routing sensor data into AI agents built on top of commodity models the company swaps interchangeably between vendors (Sanjit @ 37:18). The mechanism underneath is a data flywheel: more sensors deployed means more events captured, which trains better domain-specific agents, which makes the sensors more valuable to the next customer. Samsara's public framing of that work sits on its Beyond 2026 product blog, which announces new AI features for physical operations.

Samsara is best understood as the back-office nervous system for physical operations: dashcams, engine diagnostics, location tracking, and a software layer that turns that telemetry into safety, compliance, and maintenance workflows for trucking, construction, energy, and food distribution. The customer base runs from small fleets to companies like Home Depot, which Biswas says cut auto-insurance claims 65% after adopting Samsara's dashcams (Sanjit @ 49:29).

That pitch lands differently in 2026 than it would have in 2023. Three years ago, the AI conversation was still about who had the best model. Now OpenAI, Anthropic, and Google ship roughly comparable frontier capabilities, and customers increasingly shop on price, latency, and integration. Biswas says Samsara treats all three as interchangeable inputs: the company's value proposition is the data layer that sits underneath them, not any single vendor's weights. The frame is the same one oil-and-gas data companies have used for a decade. The rig is replaceable, the well log is not.

The financials behind the thesis are in Samsara's Q4 FY2026 SEC press release: the company has crossed $2 billion in annual recurring revenue, is profitable, and is growing around 30% year over year. That scale funds the sensor deployment, which funds the data lead, which is the whole argument. The flywheel is real, but it is a self-reported flywheel. The 380,000 crashes prevented last year (Biswas, Sanjit @ 8:46), the 99% of US roads, and the Home Depot 65% figure all come from Samsara or the CEO, not from independent verification.

The moat claim also has visible kill conditions. Cheaper sensor hardware from new entrants could narrow the data gap. An open data consortium across trucking associations could pool telemetry in ways that undercut a single vendor's lead. Dashcam video carries regulatory and privacy exposure in California, Texas, and the EU that could constrain what Samsara can capture. And the customer base, while broad, is concentrated in a handful of verticals that face their own cyclical and insurance pressures. None of these invalidate the thesis, but each one is a reason it is not yet settled.

The portable question for any AI deployment is the same one Samsara is built around: where does the continuous, regenerable, hard-to-replicate data come from, and is it a by-product of the business or a separate acquisition cost? Model-only companies spend capital on GPUs and on licensing data. Samsara spends capital on hardware that lives in customers' trucks and engines, and the data is a by-product of operations the customer was already running. That distinction is why a fleet-sensor company is now a serious public-market test case for the data-versus-model argument.

The next read is whether Agents Studio's agent-based workflows translate into the kind of switching cost that turns sensor data into durable pricing power. Biswas says Samsara will keep shopping models from OpenAI, Anthropic, and Google interchangeably. The market's question is whether Samsara's data lead keeps compounding, or whether model commoditization eventually equalizes everything it touches.