Samsara CEO：物理世界最大的AI部署Samsara CEO on the Biggest AI Deployment in the Physical World
The Takeaway：物理AI的真正解锁不在于爬取互联网token，而在于把现实世界的messy数据（GPS、摄像头、传感器）数字化，并用agent直接闭环行动。
Samsara联合创始人兼CEO Sanjit Biswas把公司打造成一家服务建筑、能源、物流等物理运营的科技公司，市值约200亿美元，ARR已超20亿美元，利润增长约30%。系统每天覆盖美国99%的道路（通常多次），每年处理约25万亿数据点，服务数百万车辆和一线工人。过去一年，他们帮助避免约38万起交通事故，并减少了数十亿磅二氧化碳排放。
物理AI与数字AI的最大区别在于数据不存在于Reddit或网页上。Biswas说：“这些不是你能在线找到的token。你没法爬Reddit去了解建筑工地发生了什么。”硬件必须扛得住恶劣环境、不可靠网络，并被数百万一线工人真正采用。Samsara从车队GPS和行车记录仪起步，逐步扩展到资产追踪、边缘AI（疲劳/手机检测、实时提醒），再到Agent Studio——能自动处理保修索赔、生成司机简报、调整安全设置的agent。
边缘跑推理与实时告警，云端做视频推理与生成式教练视频。模型策略完全开放：用Frontier Labs、开源可蒸馏模型，也自训小模型。Biswas认为未来5-10年会出现混合车队（人+机器人），长途物流和工地重复作业会先被自动化，但messy的长尾工作仍需要人类判断。数据中心驱动的电网建设正在爆炸——一家公用事业公司计划在未来5年把过去125年的电网容量再翻三倍，90%需求来自数据中心。
对一线司机，摄像头的主要价值其实是“免责”——大多数时候司机做对了，视频能证明。透明沟通+正面反馈是赢得信任的关键。
Samsara联合创始人兼CEO Sanjit Biswas把公司打造成一家服务建筑、能源、物流等物理运营的科技公司，市值约200亿美元，ARR已超20亿美元，利润增长约30%。系统每天覆盖美国99%的道路（通常多次），每年处理约25万亿数据点，服务数百万车辆和一线工人。过去一年，他们帮助避免约38万起交通事故，并减少了数十亿磅二氧化碳排放。
物理AI与数字AI的最大区别在于数据不存在于Reddit或网页上。Biswas说：“这些不是你能在线找到的token。你没法爬Reddit去了解建筑工地发生了什么。”硬件必须扛得住恶劣环境、不可靠网络，并被数百万一线工人真正采用。Samsara从车队GPS和行车记录仪起步，逐步扩展到资产追踪、边缘AI（疲劳/手机检测、实时提醒），再到Agent Studio——能自动处理保修索赔、生成司机简报、调整安全设置的agent。
边缘跑推理与实时告警，云端做视频推理与生成式教练视频。模型策略完全开放：用Frontier Labs、开源可蒸馏模型，也自训小模型。Biswas认为未来5-10年会出现混合车队（人+机器人），长途物流和工地重复作业会先被自动化，但messy的长尾工作仍需要人类判断。数据中心驱动的电网建设正在爆炸——一家公用事业公司计划在未来5年把过去125年的电网容量再翻三倍，90%需求来自数据中心。
对一线司机，摄像头的主要价值其实是“免责”——大多数时候司机做对了，视频能证明。透明沟通+正面反馈是赢得信任的关键。
The Takeaway: The real unlock in physical AI is not scraping internet tokens but digitizing messy real-world data (GPS, cameras, sensors) and closing the loop with agents that take action.
Samsara co-founder and CEO Sanjit Biswas has built a roughly $20B company serving construction, energy, logistics and other physical operations. It has crossed $2B ARR and is profitable while growing ~30%. The system covers 99% of US roads daily (often multiple times), processes about 25 trillion data points a year, and serves millions of vehicles and frontline workers. Last year it helped prevent roughly 380,000 road accidents and avoid billions of pounds of CO2.
Physical AI differs sharply from digital AI because the data simply does not exist online. “These are not the tokens you’re gonna find online. Like, you can’t crawl Reddit and find out about what happened on a construction site.” Hardware must survive harsh environments and unreliable networks, and millions of frontline workers must actually adopt it. Samsara started with fleet GPS and dashcams, expanded into asset trackers and edge AI (fatigue/phone detection, real-time alerts), and now ships Agent Studio—agents that handle warranty claims, generate driver briefings, and adjust safety settings automatically.
Inference and low-latency alerts run at the edge; richer video reasoning and generative coaching videos live in the cloud. The model strategy is deliberately open: Frontier Labs, distillable open-weight models, and small models trained from scratch. Looking five to ten years out, Biswas expects mixed fleets of people and robots. Long-haul logistics and repetitive site work will automate first; the long tail of messy exceptions still needs human judgment. Data-center demand is already forcing extreme infrastructure builds—one utility plans to triple the grid capacity built over the last 125 years in just the next five years, with 90% of that demand coming from data centers.
For drivers, the primary value of cameras is often exoneration—most of the time the driver did everything right and the video proves it. Transparent communication plus positive reinforcement is how you earn trust on the front line.
查看原文 →
Samsara co-founder and CEO Sanjit Biswas has built a roughly $20B company serving construction, energy, logistics and other physical operations. It has crossed $2B ARR and is profitable while growing ~30%. The system covers 99% of US roads daily (often multiple times), processes about 25 trillion data points a year, and serves millions of vehicles and frontline workers. Last year it helped prevent roughly 380,000 road accidents and avoid billions of pounds of CO2.
Physical AI differs sharply from digital AI because the data simply does not exist online. “These are not the tokens you’re gonna find online. Like, you can’t crawl Reddit and find out about what happened on a construction site.” Hardware must survive harsh environments and unreliable networks, and millions of frontline workers must actually adopt it. Samsara started with fleet GPS and dashcams, expanded into asset trackers and edge AI (fatigue/phone detection, real-time alerts), and now ships Agent Studio—agents that handle warranty claims, generate driver briefings, and adjust safety settings automatically.
Inference and low-latency alerts run at the edge; richer video reasoning and generative coaching videos live in the cloud. The model strategy is deliberately open: Frontier Labs, distillable open-weight models, and small models trained from scratch. Looking five to ten years out, Biswas expects mixed fleets of people and robots. Long-haul logistics and repetitive site work will automate first; the long tail of messy exceptions still needs human judgment. Data-center demand is already forcing extreme infrastructure builds—one utility plans to triple the grid capacity built over the last 125 years in just the next five years, with 90% of that demand coming from data centers.
For drivers, the primary value of cameras is often exoneration—most of the time the driver did everything right and the video proves it. Transparent communication plus positive reinforcement is how you earn trust on the front line.