North Korea Reportedly Sought 20,000 Nvidia Chips for State AI Push

A major chip request meets a restricted supply

North Korean trading officials reportedly sought help from Koreans living overseas to purchase 20,000 advanced Nvidia artificial intelligence chips in China, a request that points to computing ambitions far beyond the older equipment identified in published North Korean research.

The people approached declined because the chips were subject to US export controls. The reported request comes as Pyongyang pursues a domestic AI system and sets out plans to apply the technology across civilian industries and military operations over five years.

The precise Nvidia models, the timing of the approach and any subsequent deliveries have not been publicly established. North Korea is also reported to have sought advanced chips through other routes, including China or Russia, while institutional research describes alleged acquisitions through Chinese intermediaries.

A graphics processing unit, or GPU, performs many calculations at once. Originally developed to produce computer graphics, GPUs now provide much of the computing power used to train AI systems. Training means processing examples so that a model can learn patterns, such as how words fit together or what objects look like.

A purchase of 20,000 advanced chips would represent a substantial expansion of North Korea's potential computing resources. Putting those chips to work would also require servers, fast connections between machines, cooling, dependable electricity and people able to operate the equipment.

The reported approach sits alongside allegations that Chinese companies obtained about 240,000 Nvidia AI chips through a network involving businesswoman Alice Huang, in transactions estimated at $7.5 billion. That alleged network concerns purchases for Chinese companies; the North Korean request concerns a different prospective buyer. The quantities illustrate the scale of demand surrounding restricted AI hardware.

What Pyongyang wants to build

North Korean authorities reportedly instructed institutions to stop using foreign AI services, citing concerns about information gathering and outside cultural influence. Inspections reportedly began in March, alongside instructions for major bodies to develop a domestic system called the National Intelligence Information Network.

The project has been described as Juche AI, drawing on the state's doctrine of self reliance. Its intended role is to process North Korean language and cultural data within an isolated domestic network, rather than depend on services operated abroad.

Research groups at the State Academy of Sciences, Kim Chaek University of Technology and the Ministry of State Security are reportedly developing language models using local data. A language model learns patterns in text and uses them to generate or process language, including answering questions, translating documents and recognizing speech.

For institutions following the reported instructions, the change would mean replacing access to foreign services with tools developed under domestic control. Keeping processing inside the country would give authorities greater control over the information entered into those systems and the material used to train them.

The International Institute for Strategic Studies, in its assessment of AI ambitions, describes a wider agenda. The final report of North Korea's 9th Party Congress identifies AI, energy and space technologies as priorities for industrial development. It also sets a military objective of developing attack systems that use AI and operate without crews.

Older equipment reveals an existing research base

North Korean researchers have already published work using hardware substantially older than the advanced chips reportedly sought in China. An assessment based on papers in international academic journals identified Nvidia Tesla P100 and GeForce RTX 2070 GPUs, along with Qualcomm Snapdragon 820 processors, in studies involving neural networks and speech technologies.

The devices were released between 2016 and 2018. Kim Il Sung University was among the institutions identified as using this equipment. The academic papers provide specific evidence of the tools available to at least some researchers, rather than a complete inventory of the country's computing resources.

Kim Min-jung, head of the Advanced Technology Strategy Center at the Institute for National Security Strategy, assessed that this hardware gives researchers enough computing power for experiments involving surveillance, speech processing and video tracking. Her assessment described experimental capacity rather than an extensive AI computing infrastructure.

Older processors can still be useful for focused tasks. A system trained to recognize a limited set of spoken commands, for example, has a narrower job than a general language model expected to answer questions on many subjects. Researchers can also adapt models that have already been trained elsewhere, reducing the work required to build a tool for a particular purpose.

The assessment suggested that used equipment may have entered North Korea through other countries, including China. The devices themselves are subject to restrictions prohibiting their transfer to North Korea, showing that access problems extend beyond the newest generation of AI chips.

How the AI push has developed

North Korea's ambitions bring together university research, the use of foreign AI tools and plans for military development. Several dated developments help explain how these activities fit together:

• Late 2023: North Korea launched a military reconnaissance satellite, which the IISS assessment says is widely believed to have limited image resolution.
• 2024: North Korea generated about 25 billion kilowatt-hours of electricity, roughly 4% of South Korea's output, according to figures cited by IISS.
• Third quarter of 2025: Research published by 38 North describes increasing use of deepfake images and synthesized voices by North Korean IT workers.
• February 2026: A 38 North research paper examined AI use across North Korean employment schemes, cyber operations and military experimentation.
• March 2026: IISS assessed the AI ambitions set out in the 9th Party Congress report and the constraints facing their implementation.

The developments show two activities advancing together: using existing AI services for immediate tasks, and trying to build domestic capabilities for longer term state objectives. The first can draw on tools already available abroad; the second demands more equipment, infrastructure and specialized data inside North Korea.

Electricity and data are as important as chips

North Korea's electricity supply is a central constraint. The approximately 25 billion kilowatt-hours generated in 2024 amounted to about one twenty-fifth of South Korea's output. The country relies mainly on hydropower and coal, with wind and solar projects supplementing a supply that already struggles to meet household needs.

AI training requires sustained computing, not simply occasional access to a powerful machine. A large installation needs power for processors and for cooling equipment that removes the heat they produce. Interruptions can make lengthy training runs harder to manage and increase the burden on operators.

IISS also identifies limited access to suitable training data as a major obstacle, particularly for military applications. A model intended to detect military equipment or support reconnaissance needs relevant examples from sensors, radar or satellite imagery. Its usefulness depends on the quality and variety of those examples.

North Korea's satellite launched in late 2023 is widely believed to produce images with limited resolution, according to the institute. The country also lacks advanced radar and sensor systems. These limitations affect the detail available to train and operate tools intended to identify targets or interpret military activity.

Renting computing power abroad presents another challenge. IISS says large training projects commonly require tens or hundreds of GPUs, contracts lasting a year and companies whose identities have been verified. Such arrangements create financial and identity checks that are harder to navigate than brief access to an online AI service.

AI is already being used in employment schemes

The Stimson Center's Korea Program, through a 38 North research paper, describes North Korean IT workers using AI to obtain foreign employment, conceal their identities and generate income. The paper places these activities alongside efforts to acquire technical knowledge and intellectual property.

Researchers describe workers using language models to tailor resumes and assist with interviews. In observed cases, operators fed interview subtitles into a model to generate answers while a conversation was taking place. Other techniques included synthesized voices, manipulated images and noise cancellation intended to conceal background sounds.

One case described in the paper involved a North Korean worker who secured an interview with an organization focused on AI through a recruiter in Florida. During the interview, the worker checked company information and used a language model to prepare responses. The same day, roughly 20 other companies expressed interest in the worker for roles connected to AI development.

The paper describes a workforce with different responsibilities. Many workers focus on earning money, while some have permission to steal intellectual property or pursue particular technologies. Employment at an AI company can therefore provide income, practical training and access to proprietary work.

Research Center 227, associated with the Reconnaissance General Bureau, is reportedly being established near universities that train technical personnel. The paper presents such facilities as possible hubs for connecting skills acquired abroad with domestic research programs.

Cyber operations offer a more immediate use

North Korean cyber groups are reported to use AI to generate malicious code, improve phishing messages and debug software used in attacks. Phishing involves sending deceptive messages designed to persuade someone to reveal information, open a harmful file or visit a fraudulent website.

Language tools can make those messages more fluent and better suited to a particular target. Code generation tools can help an operator produce or revise parts of a malicious program. These uses fit into existing operations and can draw on models developed outside North Korea.

IISS distinguishes this activity from the more demanding task of building AI for military missions. Publicly available models can produce text, images or code immediately. A system intended to guide a weapon or interpret a particular battlefield needs further training for that environment.

The distinction helps explain why restrictions on advanced chips can constrain domestic infrastructure while North Korean operators continue using AI abroad. Access to a finished service and ownership of the computing equipment needed to train a specialized system involve different resources.

Military claims meet demanding technical requirements

North Korea has claimed that it developed autonomous reconnaissance and attack drones equipped with AI, and that its guided multiple rocket launchers incorporate AI guidance. The Party Congress agenda places such systems alongside electronic warfare and capabilities intended to disrupt operations in space.

IISS Korea Chair Lami Kim assesses that effective military AI could allow North Korea to gather and process information faster, supporting quicker operational decisions. For South Korea, the concern includes more capable targeting of military facilities, critical infrastructure and leadership locations.

The technical demands remain substantial. A drone that recognizes objects must work with the information available from its sensors. A reconnaissance tool must distinguish useful details from incomplete or poor quality images. Training for those jobs requires data connected to the conditions in which the system will operate.

Civilian research and state surveillance provide other applications. The institutional assessments describe work on translation, facial and voice recognition, and vehicle licence plate identification. An AI school has reportedly been established at Kim Il Sung University to train specialists.

The reported request for 20,000 chips brings the scale of Pyongyang's ambitions into focus. Its present activities range from experiments on older processors to the use of foreign AI in cyber and employment schemes. Building a much larger domestic system would require North Korea to secure the chips while also expanding the electricity, facilities, data and expertise needed to use them.

Key Points

• North Korean trading officials reportedly sought help buying 20,000 advanced Nvidia chips in China, but the people approached declined because of US export controls.
• Authorities reportedly ordered institutions to replace foreign AI services with a domestic system using North Korean language and cultural data.
• Published research identifies older Nvidia and Qualcomm hardware in North Korean AI experiments.
• IISS identifies electricity, computing infrastructure and suitable training data as major constraints on larger projects.
• Institutional research describes existing AI use in fraudulent employment schemes, phishing and malware development.
• North Korea's Party Congress set out plans for AI development across civilian and military sectors over five years.