True

2026年3月，海淀的一家物理AI仿真基础设施公司光轮智能，完成了10亿元的A++及A+++轮融资，就此成了全球具身数据领域的第一家独角兽。这家公司有一个反直觉的定位，它不造机器人，它卖给机器人公司的，是让它们少用真机、也能快速学会本事的仿真数据和评测。

In March 2026, Guanglun Intelligence, a physical-AI simulation-infrastructure company in Haidian, completed a 1-billion-yuan A++/A+++ round, becoming the first embodied-data unicorn in the world. It has a counterintuitive positioning: it builds no robots. What it sells to robot companies is the simulation data and evaluation that let them learn fast with far fewer physical machines.

要讲清它凭什么拿10亿，得先说具身智能绕不开的数据瓶颈。真机遥操作数据质量最高，但成本堆起来根本没法规模化；仿真数据便宜，却隔着Sim-to-Real这道gap，模拟器里的策略搬到真机上常常水土不服。光轮押的就是仿真加人类视频数据这条路，行话叫卖铲子，谁挖金矿都得先买它的铲子。

To see why it is worth a billion, start with the data bottleneck embodied AI cannot dodge. Real-robot teleoperation data is the highest quality, but its cost is impossible to scale. Simulation data is cheap, yet it sits across the Sim-to-Real gap, and policies trained in the simulator often misbehave on a real machine. Guanglun bets on the simulation-plus-human-video path, the industry calls it selling shovels; whoever digs for gold must first buy its shovels.

光轮把这套能力做成三层。仿真层，它是全球唯一自研物理求解器并搭起完整基础设施的企业，多物理场高精度仿真自主可控。数据层，它建了全球最大规模的非本体数据引擎，覆盖仿真合成数据和人类视频数据，给全球超过80%的头部具身智能团队提供仿真与数据服务。评测层，它有全球首个工业级仿真评测平台RoboFinals，在给行业定评测标准。

Guanglun stacks the capability in three layers. The simulation layer makes it the only company in the world with a self-developed physics solver and a complete infrastructure, achieving autonomous control over high-precision multi-physics simulation. The data layer holds the world's largest non-embodied data engine, spanning synthetic and human video data, and it serves more than 80% of the world's top embodied teams with simulation and data. The evaluation layer carries RoboFinals, the first industrial-grade simulation evaluation platform, which is setting the industry's benchmark.

这套东西的效果是能算出来的。在它搭的数字平行工厂里，机器人可以做成千上万次的并行训练，模拟各种极端工况，把机器人开发周期从传统的3到6个月，压到2到3周，综合训练成本能降十倍以上。

The payoff is quantifiable. In its digital parallel factory, a robot can run tens of thousands of parallel training passes and simulate all kinds of extreme conditions, compressing the development cycle from the usual three to six months down to two or three weeks, and cutting overall training cost by more than ten times.

技术很快换成了市场认可。光轮在仿真合成数据、仿真评测、人类视频数据三个维度，都拿过全球交付冠军，是唯一覆盖三项的规模化交付企业，合作方名单里排着字节跳动、阿里巴巴、智元机器人、英伟达、谷歌。营收上，2025年它实现了10倍增长，2026年第一季度单季，预计收入就要超过2025年全年总和。

Technology quickly turned into market recognition. Guanglun took the global delivery crown in all three dimensions of synthetic data, simulation evaluation, and human video data, the only firm delivering at scale across all three. Its partner list reads ByteDance, Alibaba, Zhiyuan Robotics, NVIDIA, and Google. On revenue, it posted ten-fold growth in 2025, and the single first quarter of 2026 is projected to exceed the entire prior year.

把这些串起来，就能回答那个标题里的问题。在具身智能还没真正大规模商业化之前，数据作为基础设施，会比终端应用更早形成商业回报。光轮不碰最烧钱的整机，它把整个行业都需要的仿真与评测做成生意，所以0真机数据，也能拿到10亿。

Threaded together, this answers the question in the headline. Before embodied AI reaches real-scale commercialization, data, as infrastructure, will form commercial returns earlier than terminal applications. Guanglun stays off the cash-burning whole-machine business and turns the simulation and evaluation every player needs into a business, which is why zero real-robot data still earns it a billion.