How a Weekend Webcam Came to See 90 Billion Miles a Year

On Sequoia's Training Data podcast, hosts Sonya Huang and Pat Grady introduce Sanjit Biswas as a rare founder who has scaled AI in the physical world twice: first at Meraki, then at Samsara. The intro frames Samsara as a public company with a market cap in the $20 billion range — Huang later cites $23 billion — with sensors streaming from millions of vehicles and about 90 billion miles of driving captured each year. Huang also teases what she calls the best ticker on the public markets before turning to the through line in Biswas's career.

Biswas describes himself as an engineer by training, with an EE and CS background, a Stanford undergraduate degree, and graduate work at MIT. More than twenty years ago he worked on Roofnet, an attempt to build very large wireless networks when Wi-Fi was still new and internet access was expensive. The idea, he says, was that access should be everywhere, in the air, which required building a big network rather than treating connectivity as a scarce commodity.

Meraki grew out of that work, was funded by Sequoia, and was acquired in 2012 for $1.2 billion. Samsara, by contrast, is aimed at physical operations: energy, utilities, construction, logistics, and other labor-intensive, asset-heavy industries that power infrastructure. Trucking is still a real business — about 20 to 25 percent — but field service and construction are now large, and the company is moving into public sector, student transit, and aviation. It recently signed the largest yellow school bus operator in North America. The promised impact is unglamorous and concrete: less risk, more efficiency, better sustainability, using data and now AI.

Asked what he saw at founding, Biswas rewinds to 2015 and three intuitive bets. Connectivity had gone from rare and expensive to ubiquitous, a change he had watched through the Meraki years. Compute was coming online: the cloud had matured, and NVIDIA's embedded GPUs were already producing striking graphics in a handheld Nintendo Switch. Cameras had gotten extraordinary seven or eight years after the iPhone launch. Put connectivity, compute, and sensors together, he says, and you had the makings of a sea change in processing real-world data.

Grady presses on the fact that nobody had a crystal ball for AI. Biswas's answer is that Samsara was the opposite of Meraki: the founders were domain experts in networking the first time, and complete outsiders the second. They had never driven a commercial truck or worked in a warehouse. What they did have was a feel for the compounding rate of the underlying technologies, plus a sense that physical operations were overlooked. Even if the tools were not ready yet, they figured, in five to ten years — about now — they would be.

That is the lesson he offers current founders. When the ChatGPT moment happened, the models were not perfect; they have gotten much better in two or three years and will get better still. Technical people understand that trajectory in a way consumers and customers often do not. Samsara, he implies, was built by sitting on the right side of those curves long enough for the capability to catch up to the ambition.

Building AI in the physical world is fun, Biswas says, because the world is diverse. Samsara dash cams sit on millions of vehicles and see about 99 percent of U.S. roads: urban, rural, residential, and every kind of weather. That long tail of exceptional cases is both the training set and the inference problem. The hard part is that it is messy and distributed. Streaming every video feed to the cloud would produce an insane bandwidth bill and demand massive data centers. So the company runs inference in the cameras themselves.

That changes the power budget. The devices draw two to ten watts, not kilowatts, with a footprint closer to a phone than a server: Linux, hundreds of megabytes to a few gigabytes of memory, not a microcontroller and not a data center. State-of-the-art models with hundreds of millions or billions of parameters simply will not fit. The workaround is to train in the cloud, distill with teacher models, and ship a small model designed for one job. The dash cam does not need to answer what the capital of France is. It does need to understand the risk profile on the road, so it is trained on data relevant to that task.

Edge and cloud stay connected. Continuous risk while driving means continuous on-device inference; tokens, images, video, and telemetry stream up for heavier work. When an accident happens, Samsara can run a full video-language reasoning model in the cloud to reconstruct whether the driver was cut off, driving defensively, or dealing with bad conditions. Biswas notes they could not have done that five years ago. He is cautiously optimistic about world models but is not training his own, just as the company does not train base foundation models. It is waiting to use them. The dataset waiting in the wings is large: about 90 billion miles a year.

Huang asks how much of that road data they can actually use. Biswas says a lot of it, and that they can train over the entire set. The practical question is what to do with tokens once they reach the cloud, and the honest answer is that this year's uses are far more interesting than those of two or three years ago. The products began with risk. For operations customers, driving to and from the job site is often more dangerous than construction or oil-and-gas work; that is where most fatality and high-severity risk shows up.

A couple of years ago the biggest risk they were seeing was mobile-phone use in big trucks, so they built a detector. Once that was in place, the next unlock was weather. An old-school AlexNet-style convolutional weather detector got a lot wrong and could not read road conditions; more sophisticated models can. Historically the system understood what was happening in the moment — phone, seat belt, following distance. Now it can look over the course of a trip. The two compounding assets, model capability and dataset size, keep working in the company's favor.

The unlock Biswas is most excited about is detecting good behavior, not only risky events. Frontline workers, he says, are doing a great job 80 to 90 percent of the time, and nobody sees it. The system can now give someone a high five or a kudos, which makes people's day — an unexpected silver lining of cameras that were sold on safety. Later in the conversation he puts numbers on coaching: Samsara often reduces customer risk by about 75 percent, roughly half from in-cab alerts in the moment and half from coaching that can break a phone habit. The same loop applies to fuel efficiency, gamifying how drivers operate heavy equipment.

Huang calls the sensor footprint a tech nerd's dream and asks why so few companies have actually put sensors on the physical world at scale. Biswas's answer is blunt: it takes a village. Unlike pure software, the products have to be installed on millions of vehicles, frontline workforces have to be trained, and customers have to get value from day one. At Samsara's scale that now means thousands of people doing installs and change management. He counts the companies with datasets of this magnitude on one hand: Tesla, then probably Samsara, then Waymo — and Waymo is still in the thousands of vehicles, not millions.

That footprint required go-to-market discipline that did not come naturally. When Meraki started, Biswas had never sold anything; as an engineering nerd he even tried to dodge school candy-bar fundraisers by offering to build a website instead. What converted him was impact: if the product is not out there, it is not helping anyone, and sales are what make that sustainable. The 2008 financial crisis shut off risk capital and forced Meraki to operate at break-even, which turned sales execution into an engineering problem — predictable, repeatable, designed. The habit stuck.

At Samsara the company raised capital along the way but reinvested far more from revenue and gross margin. Because it is public, Biswas points listeners to the balance sheet: they have put close to $3 billion into getting the product out, across R&D and customer success. That is only possible with a lot of sales. Why pick a domain they did not know? Curiosity. After Meraki he thought about finishing a PhD; co-founder John Bicket talked him out of the long academic feedback loop. He wandered into energy, the grid, nuclear, satellites, and kept getting pulled toward overlooked infrastructure.

Huang says that two years ago she avoided Waymos and now feels safer in one than not. Biswas is very bullish. He watched Stanford's DARPA Grand Challenge cars as an undergraduate more than twenty years ago, and he now prefers Waymo for its consistency. His core claim is that autonomy increases the operational intensity of the world. There is a third shift between midnight and 8 a.m. that people skip because they are sleeping; logistics could still run. A field technician who needs a part could have it delivered. When automation kicks in, volume usually rises because cost falls. Demand that would not pay $50 for a human drop-off looks obvious at $5 or free. He frames autonomy as an *and*, not an *or*.

Warehouses, he says, already got this wave about a decade ago: lift systems, big arms, industrialized sites where automation is now the default. Workers welcome it because it reduces injury from lifting pallets. Construction sites and job sites are next. On humanoids he is cautiously optimistic and a little scared; headless or handless walkers still sit in the uncanny valley, though friendlier designs such as 1X's NEO exist. It reminds him of self-driving ten years ago: not tomorrow, but inevitable. Samsara's role, zoomed out, is a platform to see operations — workers, vehicles, assets, maintenance — that will still be needed when some of those assets are robots.

Humans and AI are getting closer. In 2015 most customers still ran on pen and paper; now they use apps, and in the last year many have taken quickly to voice bots for freight brokerage. Adoption is highest where task intensity is repetitive. The physical work itself is still done by people because exception handling is everywhere; construction is too diverse to automate the way car manufacturing is. Geopolitics as a US-versus-China story has not shown up in his customer conversations. What he does hear is a push to modernize, especially among energy and construction firms busy with data-center build-outs, asking what should be different in the twenty-first century.

The portfolio did not start with dash cams. In 2015 cameras were not quite viable on cost and backhaul, and the surprise was that most operations still had no real-time idea where field teams were. Uber and DoorDash had live tracking; long-haul logistics was still dropping breadcrumbs every five to fifteen minutes, often on grainy MapQuest printouts that Biswas likens to Minecraft-level graphics. GPS tracking was product one. Customers then asked whether Samsara had a dash cam that worked with the system. Co-founder John Bicket ordered a webcam from Amazon, plugged in USB, and wrote enough code over a weekend to bring a prototype back the next week. Customers loved it. Watching the footage, they could see phones in hand at the moment of accidents, which is how the AI detector was born. The strategy since has been concentric circles: core use case, then adjacent, now about ten products, with the camera still sold alongside the original tracker.

Bandwidth has been a free ride twice. Samsara launched around the 3G-to-4G transition, when YouTube and baseball on phones drove carrier data consumption up and the marginal cost per gigabyte down. Something similar is happening with satellite. Starlink costs many tens of billions to build, but the marginal cost of another device is low, which helps the last 1 percent of coverage — oil-and-gas sites with no roads and little cellular. Cheaper backhaul will not kill the edge. If the workload were only GPS, you would stream it all; eight HD cameras around a truck plus telemetry is another story. Compression still helps, and for controls Biswas expects on-device compute for a long time. Tesla's Full Self-Driving computer is a couple of thousand dollars and many hundreds of watts; Waymo's is probably more. Humans in the cab and on the road make a network outage an unacceptable risk.

By 2030 he expects today's tools to diffuse the way apps did, but he stresses that AI adoption in physical operations is still at the base of the curve at the end of 2025. The new layer he wants is augmented reality and wearables: hands-free frontline work, AI in the ear, visual feedback, a VLM reading the environment — technically on the cusp, possible by 2030. Personally he talks to voice models on the drive to work and likes ChatGPT Pulse for Bay Area events for his three kids, a hint that AI can know you well enough to surface experiences you would have missed. He refuses to split customer value into AI versus everything else. Customers fear data overload from every vehicle, worker, and asset; AI's job is to distill that into something actionable. Last week in Texas, a food distributor, an oil-and-gas company, and Home Depot were already using the data for time-card punches and sharing ETAs — uses that were not on the original roadmap.

His advice to young technical founders is that it is an amazing time to build. Codex, Cursor, and other coding tools let you turn an idea into something real far faster than when Samsara or Meraki started, when the team was buying Dell servers and racking them in a data center. Ten years from now people will say, can you believe we did X, whatever X turns out to be. The best place on an exponential curve, he says, is to be building.