The news came with few details, but the number itself resets industrial priorities: JD.com, one of the world’s largest logistics operators, has announced plans to retrain up to 700,000 workers affected by the spread of warehouse robots. This isn’t mass layoffs, but a radical repositioning of human skills in an ecosystem increasingly governed by autonomous machines.

Automation eats repetitive tasks

The transformation touches every major distribution network: Amazon, Alibaba, Walmart and others have deployed autonomous guided vehicles, sorting arms, and computer-vision-based picking systems. JD.com is no exception. The scale is the telling detail — 700,000 people is a substantial slice of the workforce, signaling that automation is no longer confined to pilot projects but is reaching core processes.

What exactly do these robots do? They move goods, pack, label, and sort parcels at speeds impossible for a human operator. To do so, they rely on machine learning models that process sensor data in real time — cameras, lidar, weight sensors — and make decisions in fractions of a second. Here lies the technical crux for the AI-RADAR community: where does the intelligence guiding these fleets run? Almost always locally, right at the edge.

Why logistics pushes on-premise

Latency constraints are tight: a cart that needs to avoid an obstacle cannot afford the round-trip to a remote cloud. Likewise, data describing warehouse layout, inventory levels, and product flows carries competitive sensitivity and often falls under regulatory scrutiny. The prevailing architecture is therefore self-hosted: industrial servers or embedded GPU modules (such as NVIDIA Jetson or similar) installed directly inside the warehouses, connected to the robots via local network or private 5G.

This scenario fits squarely into the debate we follow at AI-RADAR: the choice between cloud and on-premise is not merely a cost issue, but one of effective control over inference. In logistics, a machine stoppage caused by network latency or a botched remote update can translate into hours of operational paralysis. Keeping AI in-house means being able to orchestrate updates, manage queues, and ensure deterministic performance.

What role do LLMs play in this game

Until a few years ago, the models used in warehouse robotics were convolutional networks for vision or decision trees for path planning. Today, the entry of Large Language Models — and multimodal models — opens more ambitious avenues: a robot could receive natural language instructions (“Pick the third pallet on the left and bring it to bay 12”) and translate them into a sequence of physical actions.

To do that, however, the model must run locally: no facility manager wants to depend on an external API for daily operations. A world is taking shape in which robot fleets integrate quantized LLMs at INT8 or FP16 precision, fine-tuned for specific domains (warehouse layouts, internal nomenclature). Serving frameworks like vLLM or llama.cpp already allow models with 7–13 billion parameters to run on relatively compact hardware, with latencies compatible with task planning — if not yet with real-time control.

What it means for enterprise AI deployment decisions

JD.com’s move is a strong market signal: large-scale physical automation brings with it the need for new skills, but also for distributed, internally manageable computing infrastructure. For organizations building their AI strategy, it becomes critical to analyze the Total Cost of Ownership of a self-hosted deployment versus hybrid alternatives, considering not only hardware but also operating costs related to cooling, maintenance, and continuous model updates.

Unsurprisingly, an increasing number of vendors are pushing dedicated appliances for on-premise inference, from servers equipped with H100 GPUs to multi-GPU workstations for disaggregated environments. Data sovereignty — a non-negotiable principle in many regulated sectors — finds an unexpected ally in logistics: the operational necessity of low latency coincides with the requirement to keep sensitive data within the corporate perimeter.

AI-RADAR explores these trade-offs in its section on on-premise deployments, offering analytical frameworks for those who must decide if and how to bring models into their own facilities. JD.com’s lesson is that the game is not played only on the model or the algorithm, but on the organization’s ability to absorb change — and to equip itself with the right infrastructure to sustain it.