After Deploying 30,000 Autonomous Vehicles, JD.com Pivots to Building 'Digital Utilities'
As autonomous fleets scale from hundreds to tens of thousands, the challenge shifts from individual obstacle avoidance to city-wide dispatch. JD.com is partnering with Moore Threads to build a 100,000-GPU cluster, aiming to become the foundational infrastructure for the robotics era.
According to Ziniu News, a regional Chinese news outlet, autonomous delivery vehicles successfully navigated narrow barriers on September 10,打通ing the "last meter" of rural postal routes. On the same day, driverless emergency vehicles were seen navigating exhibition streets. While obstacle avoidance for individual autonomous vehicles has matured, JD.com (one of China's largest e-commerce and logistics companies) used the day to announce its "Physical AI" strategy. Partnering with domestic GPU manufacturer Moore Threads, JD plans to build a 100,000-GPU intelligent computing cluster, declaring its ambition to provide the "water, electricity, and coal"—a Chinese tech metaphor for essential foundational infrastructure—for the robotics industry.
From "Solo Operations" to "Computing Infrastructure"
The autonomous vehicle industry is experiencing three distinct phases, and the first step—"single-vehicle intelligence"—has largely been achieved. Today's autonomous vehicles can accurately identify width-restricting barriers and agricultural vehicles, but this is merely the baseline. When the number of autonomous vehicles on the streets surges from hundreds to tens of thousands, the real test begins.
JD.com's announcement of a 100,000-GPU computing cluster is designed to address the pain points of this scale. Simply put, single-vehicle intelligence solves the problem of "whether this specific vehicle can drive," whereas a 100,000-GPU cluster solves "how to prevent tens of thousands of vehicles from causing chaos together." For the general public, this means future autonomous vehicle services will not collapse under increased traffic. Instead, they will be dispatched as precisely as traffic lights, making urban transit smoother.
A Bifurcated Market: Overseas Expansion vs. Closed-Environment Monetization
While JD.com focuses on underlying computing power, the broader autonomous driving industry shows an interesting divergence in commercialization: some are looking outward, while others are digging inward.
On one hand, seeking growth overseas has become a consensus. On September 10, WeRide, a Chinese autonomous driving technology company, secured Spain's first operational license for Level 4 autonomous passenger cars, opening a foothold in the European market. On the other hand, commercialization in closed environments is already proving viable. Also on September 10, Ningbo-Zhoushan Port—one of the world's busiest ports in China—received a record single order for 370 new energy heavy-duty trucks.
To draw an industry analogy: this is similar to early telecommunications networks, where some companies built cell towers in remote areas while others laid fiber optics in bustling commercial districts. Whether it is heavy-duty trucks in closed ports or passenger cars on overseas streets, they are all accumulating real-world driving data. If this massive amount of data can eventually feed into a unified, city-level Physical AI network (AI systems that interact with and navigate the physical world), it could yield significant synergies. However, large-scale application on open roads still faces policy uncertainties regarding local traffic regulations and right-of-way allocation, which remains to be seen.
Becoming the "Power Grid" of the Robotics Era
As the industry evolves into its third phase, the competition is no longer about wheels and radars, but about the "brain." Why is JD.com investing heavily in a 100,000-GPU cluster?
The essence of this move is the "infrastructuralization of computing power." JD.com does not just want to be a company with a massive logistics fleet; it aims to become the "power grid" of the robotics era. When computing power becomes a utility like water and electricity, any enterprise building robots can plug directly into this network without having to train an AI brain from scratch.
Imagine you live in a large residential community. At 11:00 PM, you request an autonomous vehicle to deliver cold medicine, your neighbor orders one for late-night food delivery, and property management calls one to clear garbage. Without city-level Physical AI, these three vehicles might gridlock in the community's narrow pathways, or freeze in place due to conflicting avoidance logic. However, with unified computing dispatch, the system instantly calculates the optimal solution, assigning priorities and routes. This means future competition will not be a hardware showdown between individual autonomous vehicles, but a contest of the invisible computing and dispatch networks behind them. Eventually, receiving packages or calling an autonomous vehicle will be as cheap and seamless as turning on a tap.
Hidden Risks Behind the Rapid Expansion
However, while the vision is comprehensive, the reality requires refinement. The construction cycle and dispatch efficiency of a 100,000-GPU domestic computing cluster still need time for practical verification. More importantly, when a city-level Physical AI takes control of thousands of autonomous vehicles, the system's fault tolerance becomes extremely low. If the computing network suffers a cyberattack or a fundamental logic bug, it could paralyze all autonomous vehicles in an entire district, potentially causing chain-reaction traffic accidents. Furthermore, once massive trajectory data is aggregated, balancing dispatch efficiency with user privacy is a mandatory question for these "digital utility" providers. The endgame of Physical AI is undoubtedly an invisible network covering the city, but the process of weaving this web will inevitably involve complex trade-offs.
Key Takeaways
Shareable Summary: JD.com's partnership with Moore Threads to build a 100,000-GPU computing cluster marks a shift in the autonomous vehicle industry from "competing on single-vehicle hardware" to "competing on city-level computing dispatch," as JD attempts to become the foundational infrastructure for the robotics era.
Discussion Prompt: If your neighborhood fully adopted this city-level Physical AI dispatch system, what specific daily inconvenience would you most want autonomous vehicles to solve for you? Late-night pharmacy runs or moving heavy items?