The PLA Adopted DeepSeek as Its C4ISR Backbone in 2025. India Should Know What That Means

In January 2025, a Chinese artificial intelligence startup called DeepSeek released a language model that matched the performance of the most advanced American AI systems at roughly six per cent of their training cost. The American AI industry called it the Sputnik moment of the decade.

What most Indian readers do not know is that within weeks of that release, the People's Liberation Army began deploying DeepSeek across its entire command stack. Not as a research curiosity. Not as a pilot. As the backbone of a deliberate, sustained, well-funded military adoption programme. The documented record is now substantial.

Source: Jamestown Foundation, November 2025, RealClearDefense analysis

What DeepSeek Actually Is

For readers who have only encountered DeepSeek as a chatbot, the technical context matters.

DeepSeek is a family of large language models, similar in capability to OpenAI's GPT, Anthropic's Claude, and Google's Gemini. Its breakthrough was not raw intelligence. It was efficiency. DeepSeek's R1 model was trained for approximately six million US dollars, against the one hundred million plus that American competitors typically spend. It runs on Chinese Ascend and Kirin processors that are several generations behind the cutting-edge Nvidia GPUs that train American models. It does roughly the same work, on chips that are not subject to American export controls, at one-twentieth the cost.

That last sentence is the entire reason this article exists. DeepSeek solves the single biggest problem the PLA faced in becoming an AI-driven force: dependence on imported American chips. Beijing calls this property algorithmic sovereignty.The PLA calls it an opportunity.

The Documented Record

Multiple investigations have now mapped the PLA's adoption.

Reuters and the Jamestown Foundation, October 2025. A six-month review of PLA Procurement Network records found that DeepSeek references appeared in a dozen procurement tenders from PLA entities in 2025, against just one for Alibaba's Qwen, its main domestic rival. DeepSeek-related procurement notices accelerated through the year.

Recorded Future's Insikt Group, June 2025. Cybersecurity researchers documented that the PLA likely adopted DeepSeek rapidly after the release of its V3 and R1 models in December 2024 and January 2025. Chinese state-affiliated research institutes, including the Academy of Military Sciences, have proposed training military LLMs using multisource intelligence inputs including open-source, human, signal, geospatial, and technical intelligence.

Indo-Pacific Insights, June 2025. Over 150 separate references to DeepSeek have been catalogued across PLA and Chinese defence-entity procurement records.

Source: Recorded Future Insikt Group, June 2025

What It Is Being Used For

The documented use cases are no longer theoretical.

Autonomous combat vehicles. In February 2025, China's state defence giant Norinco unveiled the P60 military vehicle, capable of autonomously conducting combat-support operations at up to 50 kilometres per hour. Communist Party officials described it as an early showcase of how DeepSeek is being used to close the autonomous-warfare gap with the United States.

Satellite imagery and target identification. Researchers at Landship Information Technology, which integrates AI into Norinco military vehicles, claimed in a February 2025 white paper that their DeepSeek-based system running on Huawei chips can rapidly identify targets from satellite imagery and coordinate with radars and aircraft to execute operations.

Battlefield planning at machine speed. Researchers at Xi'an Technological University demonstrated a DeepSeek-powered system that assessed 10,000 battlefield scenarios in 48 seconds, a task they estimated would take conventional military planners 48 hours.

Drone swarms and unmanned ground systems. The Reuters investigation traced DeepSeek integration into autonomous combat vehicles, drone swarms, and robotic ground systems. The cost-efficiency of running DeepSeek on Chinese chips makes it economically viable to deploy at the platform level, not just at the command level.

Intelligence and OSINT. The PLA is using DeepSeek-class models to process open-source intelligence, analyse satellite imagery, extract events from intelligence streams, and prepare reports. PLA-affiliated researchers have also published warnings about LLM hallucination risks, suggesting deployment is happening with internal eyes open.

Source: The Diplomat, March 2026, DroneXL on the Reuters investigation

Who Is Building It

The most interesting structural detail of the PLA's DeepSeek adoption is who is winning the contracts. Private Chinese companies, not state-owned defence enterprises, have won the majority of contracts to build DeepSeek-integrated tools for the PLA.

Firms like Shanxi 100 Trust have secured PLA contracts worth tens of millions of yuan, building defence applications using Huawei's Kunpeng and Ascend processors.The Diplomat notes that compressed PLA tender timelines structurally favour private IT firms over the slower state-owned enterprises that traditionally dominated defence procurement.

This is the practical face of China's military-civil fusion policy. Beijing has institutionalised a system where a private AI startup releases a model on Monday and PLA research institutes can procure tools built on it by Friday. The lag between civilian breakthrough and military deployment is measured in weeks.

What This Means

Three implications worth holding in mind.

One. The American GPU export controls that were supposed to slow Chinese military AI have not done so. DeepSeek's efficiency means the PLA does not need leading-edge Nvidia chips to deploy useful military AI. The chokepoint moved. Algorithmic efficiency is now the workaround for hardware dependence.

Two. The PLA's adoption pattern is fast, fragmented, and contractor-led. There is no single PLA-DeepSeek programme to point at. There are over 150 contracts across dozens of entities, integrating the same model into hundreds of different systems. This is what intelligentised warfare looks like when it is actually being built, rather than written about.

Three. India's own GPU access cap of roughly 50,000 H100-class chips through 2027, described in an earlier Bidfoil piece, sits inside this larger picture. If algorithmic efficiency lets China deploy useful military AI on Chinese chips, the same efficiency will eventually let other countries do the same. The interesting question is who builds the Indian equivalent of a DeepSeek, and who deploys it across the Indian defence stack first.

The documented record on the PLA side is clear. The model was released on a Monday in January 2025. By Friday of that week, the first PLA pilots were already underway. The model is now the backbone of an ongoing, well-funded, contractor-led, system-wide military adoption programme that has very little institutional drag.

This is the part of the China AI story that Indian defence readers should track most carefully. Not the visible weapons. Not the parade footage. The procurement records. The compressed tender timelines. The private-firm dominance. The 150 references and counting.