Full-text search. Typo-tolerant.

Black Lens Intelligence Desk ·

PLA Adoption of DeepSeek: Procurement Confirmed, Operational Fielding Constrained by Compute

PLA Adoption of DeepSeek: Procurement Confirmed, Operational Fielding Constrained by Compute

PLA Adoption of DeepSeek: Procurement Confirmed, Operational Fielding Constrained by Compute

Bottom Line

KJ-1: PLA and PLA-affiliated entities — including PLA units, People's Armed Police formations, military hospitals, and defense-industrial-base contractors — have executed dozens of documented procurement actions integrating DeepSeek models across cyber, OSINT, autonomous systems, medical, and staff-support functions in 2025. "PLA-affiliated" is used throughout this product to encompass this full scope; where the procuring entity is a PAP unit, military hospital, or DIB contractor rather than a PLA combat unit, that distinction is noted. (High confidence — primary procurement records, plap.mil.cn; corroborated by CSET/ETO and Jamestown Foundation)

KJ-2: Verified deployments are concentrated in rear-area, staff-support, training, and simulation roles; no public evidence confirms DeepSeek in live combat targeting or kinetic operations. The confidence rating of High applies specifically to the absence of public evidence for combat use and to the confirmed existence of non-combat deployments; it does not assert that combat use is impossible or that classified deployments do not exist. (High confidence — Reuters, Jamestown, PLA Central Theater Command Hospital announcement)

KJ-3: DeepSeek's MoE architecture (671B total / ~37B active per token) and distilled 32B/70B variants enable deployment on degraded compute, partially offsetting U.S. export controls — but full-frontier training remains compute-bound. (High confidence — DeepSeek V3 Technical Report; CSIS)

KJ-4: Semiconductor constraints — HBM scarcity, SMIC 7nm yield, Huawei Ascend packaging bottlenecks, and CUDA-equivalent software gaps — limit PLA scaling from staff-support to force-wide C4ISR integration. (High confidence — RAND, CSIS, BIS rulemaking)

KJ-5: A claim that DeepSeek appears in 150+ PLA procurement records originates from a single unnamed U.S. State Department official cited by Reuters and remains unverified. (single source — Reuters, 23 June 2025; underlying dataset not disclosed to press) (Moderate confidence)

Assessment Area Finding Confidence
PLA procurement of DeepSeek Dozens of dated tenders on plap.mil.cn explicitly name DeepSeek; 2 fully URL-anchored in this evidence base High
Non-combat operational use (hospitals, training, staff) Confirmed deployments at PLA Central Theater Command Hospital and PAP units High
Combat-support / C4ISR integration Research-stage; NUDT battalion-level prototype; Norinco P60 demonstrator Moderate
Live combat or kinetic targeting use No public evidence Unverified
Operational-scale compute sufficiency Constrained by HBM, packaging, yield; partially mitigated by MoE efficiency High

Key open question: Whether Huawei Ascend 910C production and HBM stockpiles can support force-wide DeepSeek inference at C4ISR scale before 2027 — resolution would shift KJ-4 toward either fielding viability or sustained bottleneck.

Key Judgments

KJ-1: Dozens of PLA procurement notices published on the PLA Procurement Network (plap.mil.cn) in 2025 explicitly require DeepSeek model integration across intelligence, cyber, autonomous-systems, and staff-support functions. Two notices are fully URL-anchored in this evidence base; additional notices are documented by Jamestown Foundation and CSET/ETO through primary-source review. The gap between "dozens documented by Jamestown" and "2 verifiable through direct URLs in this product" reflects a search-coverage difference, not an evidentiary absence — but the confidence rating is calibrated to the independently verifiable subset. (High confidence on the existence of a substantial procurement pattern; Moderate confidence on the precise count)

KJ-2: The most reliably documented PLA use cases are non-combat: military hospital decision support, staff document generation, OSINT crawling, cyber range training, battle-scenario simulation, and drone-swarm research — not verified live targeting. This judgment is distinct from the combat-support/C4ISR row in the summary table, which reflects a separate, Moderate-confidence assessment of prototype and demonstration activity. (High confidence on non-combat use; Moderate confidence on combat-support prototype status)

KJ-3: Norinco's P60 autonomous combat-support vehicle, unveiled in early 2025, is the highest-profile public showcase of a DeepSeek-powered military platform, but the technical basis of "powered by DeepSeek" — whether onboard inference, cloud-linked processing, or a fine-tuned variant — has not been publicly specified, and operational fielding status remains opaque. (Moderate confidence)

KJ-4: DeepSeek's mixture-of-experts architecture and distilled smaller variants (1.5B–70B) materially lower the compute floor for military deployment, enabling local on-premises inference on Huawei Ascend hardware at rear-area command nodes. (High confidence)

KJ-5: U.S. export controls on advanced GPUs, HBM, and semiconductor manufacturing equipment have created a 12–24 month adoption lag but have not halted PLA access, due to stockpiles, smuggling networks, and architectural workarounds. (Moderate confidence)

KJ-6: The PLA's adoption tempo — driven by open-weight model availability and decentralized contractor procurement — structurally outpaces U.S. JCIDS/ATO acquisition cycles, conferring a doctrinal-experimentation advantage independent of raw compute parity. This judgment rests on the qualitative contrast between weeks-to-deployment observed in plap.mil.cn award notices and the multi-year JCIDS/ATO process documented in DoD acquisition policy; no quantitative cycle-time comparison is available in the open-source evidence base. (High confidence on the directional assessment; the magnitude of the gap is not precisely quantified)

KJ-7: Open-weight release of DeepSeek-R1 in January 2025 is an irreversible proliferation event: model weights are downloadable from public repositories and cannot be recalled, enabling adversary fine-tuning beyond Chinese state control. The compressed timeline — January 2025 release to dozens of documented PLA procurement actions by October 2025 — is itself a significant indicator of adoption velocity and should be foregrounded in any policy assessment. (High confidence)

The People's Liberation Army is procuring DeepSeek models at scale and integrating them across staff, intelligence, cyber, training, and unmanned-systems functions — a pattern documented in the PLA's own public procurement records, corroborated by independent analysis from Georgetown's Center for Security and Emerging Technology (CSET) and the Emerging Technology Observatory (ETO), the Jamestown Foundation, Reuters investigations, and Recorded Future's Insikt Group. The evidence base supports high confidence in procurement and rear-area deployment. It does not support claims of verified combat fielding. The binding constraint on scaling from staff-support to operational C4ISR is not policy or doctrine but semiconductor supply — specifically high-bandwidth memory, advanced packaging yield at SMIC, and the immaturity of Huawei's CANN software ecosystem relative to CUDA.

The timeline matters: DeepSeek-R1 was publicly released in January 2025. The procurement notices, hospital deployments, academic publications, and contractor awards documented in this product span January through October 2025 — a nine-month window. That compression is analytically significant: it indicates that open-weight availability, not procurement lead time, is the primary driver of PLA adoption velocity.

Methodology and the Central Analytical Contradiction

This product applies a five-tier evidence hierarchy. Tier 1: primary PLA procurement documents on plap.mil.cn and ccgp.gov.cn, with notice IDs and dates. Tier 2: peer-reviewed Chinese-language academic publications (CNKI, conference proceedings from the National University of Defense Technology and AVIC institutes) and patent filings. Tier 3: independent OSINT analytical institutions with declared methodology (CSET/ETO, Jamestown Foundation, SemiAnalysis, CSIS, RAND). Tier 4: investigative reporting by major wire services (Reuters, SCMP) where reporters describe their source review process. Tier 5: secondary aggregation and U.S. government statements where the underlying dataset is not disclosed. Chinese state-media outlets (PLA Daily, Global Times, Xinhua, Guangming Daily) are treated as Tier 4 when reporting verifiable facts but downgraded to a separate "deliberate signaling" tier when promoting capability claims, and are flagged accordingly throughout.

The central contradiction the council research surfaced: one model (perplexity/sonar-pro) asserted that no primary PLA procurement notices directly naming DeepSeek were identified, while three others (openai/gpt-5.5, anthropic/claude-sonnet-4-6, openai/o3) cited specific plap.mil.cn notices with IDs, units, and budgets. The contradiction is resolved by Tier 1 evidence: primary procurement notices on plap.mil.cn naming DeepSeek do exist, are publicly accessible, and were documented with direct URLs. The perplexity/sonar-pro finding reflects a search-coverage gap, not an evidentiary absence. The resolution is presented in the procurement section below.

A second contradiction concerns operational use: perplexity/sonar-pro claimed no verified PLA exercise or training use exists, while claude-sonnet-4-6 and o3 cited the PLA Central Theater Command General Hospital's public announcement of DeepSeek-R1-70B deployment and the Xi'an Technological University battlefield-simulation claim. Resolution: confirmed non-combat deployments (medical, training, simulation) exist and are publicly announced by the PLA itself; no source in the evidence base documents live combat use. Both positions are partially correct depending on how "operational" is defined.

The capability staging taxonomy used throughout: (1) intent (doctrine, RFPs); (2) acquisition (awarded contracts); (3) prototype (research papers, patents, demonstrators); (4) demonstration (showcases, exercises); (5) operational fielding (force-wide integration in real missions). DeepSeek-PLA evidence clusters at stages 1–4; stage 5 is unverified.

The Procurement Record: What plap.mil.cn Actually Shows

The PLA Procurement Network (plap.mil.cn) carries dated, signed notices that name DeepSeek explicitly. Two records anchor the evidence base with full URL verification.

The first is a procurement notice from a unit in Lhasa, Tibet Autonomous Region, for a project described as "exploring the use of the DeepSeek large model for data-pattern processing, construction, and verification." The notice required a local DeepSeek-R1 server capable of running at least the 32B-parameter variant, paired with a crawler server for open-source data ingestion into a local database. Budget: RMB 300,000. (The primary notice states this figure; "approximately" is not used — the figure is drawn directly from the notice text.) Delivery expected April 25, 2025. The Lhasa record is the clearest publicly available demonstration of a PLA workflow requiring local on-premises DeepSeek deployment with OSINT ingestion. (High confidence — plap.mil.cn primary notice, Reference 1)

The second is "DeepSeek大模型+教学工具集采购" (project number 2025-JK22-F1015), a result notice published on plap.mil.cn with bids opened May 27, 2025. Shaanxi Big Data Group Co., Ltd. was ranked first at RMB 1,820,158; Xi'an Huizhi Information Technology Co., Ltd. second at RMB 1,850,000; Beijing Qunzhihé Information Technology Services Co., Ltd. third at RMB 1,795,000. The requirement was a DeepSeek-based teaching toolset covering summarization, translation, retrieval, and data analysis. (A June 9, 2025 cancellation reported by zcwzc.com procurement mirror is not independently confirmed on the PLA site and is not treated as verified.) (High confidence on the award notice itself — plap.mil.cn primary notice, Reference 2)

Beyond these two records, the Emerging Technology Observatory (ETO), citing its review of Chinese-language procurement documents, identified additional named requirements:

A July 2024 RFP documented by CSET requested 10 million core hours of physical supercomputer cluster resources with explicit NVIDIA V100/A100 requirements. This notice predates DeepSeek-R1's January 2025 release and does not reference DeepSeek by name; it is included here as evidence of PLA compute-procurement patterns and AI infrastructure investment, not as evidence of DeepSeek adoption specifically.

CSET separately documented a May 2025 PLA research laboratory request for sixteen NVIDIA H100 GPUs — chips explicitly under U.S. export control — to support electromagnetic-characterization and infrared-characterization modeling using DeepSeek-70B and Alibaba's Qwen-32B. A July 2025 RFP requested a high-performance server with NVIDIA H20s for a "demonstration and verification system."

Jamestown Foundation analysts Sunny Cheung and Kai-shing Lau, writing on October 27, 2025, reported that PLA procurement documents from the prior six months contained "dozens of distinct procurement documents explicitly calling for tools based on DeepSeek models." They identified Shanxi 100 Trust Information Technology (山西百信信息技术) as a leading private contractor, marketing infrastructure built on Huawei Kunpeng CPUs and Ascend AI chips, with defense contracts worth tens of millions of renminbi from PLA entities and the China Aerospace Science and Technology Corporation (CASC). Reuters reporters Eduardo Baptista and Fanny Potkin reported on October 26–27, 2025 that twelve PLA tenders in 2025 referenced DeepSeek while only one referenced Alibaba's Qwen. (Reuters described reviewing approximately two dozen tenders and patents; the methodology for arriving at the count of twelve is not detailed in the published report — this figure should be treated as an approximation rather than a precise audit result.)

A Reuters report by Michael Martina and Stephen Nellis on June 23, 2025 cited an unnamed senior U.S. State Department official asserting DeepSeek was referenced "more than 150 times" in PLA and defense-industrial-base procurement records. Reuters explicitly stated it could not independently verify the dataset. (single source — Reuters, citing one unnamed U.S. official; underlying dataset not disclosed)

Reconciliation: Estimates of DeepSeek-PLA Procurement Volume

The figures below are not directly comparable — they measure different datasets, time windows, and inclusion criteria. They are not summed. The gap between the Jamestown "dozens" figure and the two URL-anchored records in this evidence base reflects a search-coverage difference, not a factual dispute; Jamestown conducted a systematic plap.mil.cn scrape that this product has not independently replicated in full.

Source Estimate Time Period Unit Definition Notes
U.S. State Department official (cited by Reuters) 150+ references Through June 2025 "References in procurement records" (PLA + DIB) Reuters could not verify the dataset; single unnamed official
Jamestown Foundation (Cheung/Lau) "Dozens" May–Oct 2025 Distinct procurement documents calling for DeepSeek tools Based on plap.mil.cn scrape
Reuters investigation (Baptista/Potkin) ~12 PLA tenders Calendar 2025 PLA-entity tenders referencing DeepSeek Methodology for count not detailed; compared to ~1 referencing Qwen
CSET / ETO Multiple discrete RFPs documented 2024–2025 Dated RFPs with text excerpts Primary-source review
Fully URL-anchored in this evidence base 2 Apr–May 2025 Notices with full ID and URL Lhasa + 2025-JK22-F1015

Documented Use Cases: What the PLA Is Actually Doing With DeepSeek

Military medical and staff support (stage 5 — operational fielding, confirmed). The PLA Central Theater Command General Hospital publicly announced deployment of DeepSeek-R1-70B to assist physicians with treatment-plan generation, with all data processed on local servers. The 301 Hospital in Beijing — where senior Chinese officials and military officers are treated — is, per a statement by senior software engineer Ren Hao published on the CHIMA (China Health Information Management Association) website, working with Huawei to deploy DeepSeek-R1 on Ascend hardware to build a local knowledge database. (single source — CHIMA website statement by one named engineer; not independently corroborated by a second source) The Nanjing National Defense Mobilization Office published a DeepSeek user manual covering emergency evacuation planning, defense education, and resource assessments. (single source — the manual's existence and contents have not been independently verified beyond the initial report) The Hainan People's Armed Police political work department posted on official social media about soldiers using DeepSeek for anxiety management and exercise plans. (single source — one official social-media post; elevated here to "confirmed deployment" only in the narrow sense that the post itself is a primary-source PLA institutional statement, not an independently verified operational assessment) (High confidence on the Central Theater Command Hospital announcement; Moderate confidence on the 301 Hospital, Nanjing, and Hainan cases individually)

Battle-scenario simulation (stage 3–4 — prototype/demonstration). A Chinese research team led by Professor Fu Yanfang at Xi'an Technological University, which trains PLA ordnance officers, reported using DeepSeek to generate 10,000 simulated battlefield scenarios in 48 seconds — a task estimated to require 48 hours by human staff. The claim was reported in the South China Morning Post on May 16, 2025 and amplified by Global Times. Reuters explicitly stated it could not independently verify the performance claim. Global Times is a Chinese state-media outlet treated here as deliberate signaling rather than independent reporting. (contested — Reuters could not verify; Global Times amplification is consistent with deliberate capability signaling) (Moderate confidence in the claim's existence; Low confidence in performance validity)

Battalion-level command integration (stage 3 — prototype). A team led by NUDT research scientist Bo Huang published in Command Control & Simulation on March 12, 2025 a system combining a large language model with a dynamic battlefield map to identify "critical information requirements," reportedly tested in an amphibious-landing simulation relevant to Taiwan contingencies. The team claimed decision response time improved by 43%, with over 90% recall accuracy under communications jamming, and stated the system "has already been integrated into a command information platform capable of supporting battalion-level operations." (single source — Bo Huang et al., peer-reviewed but author-affiliated with PLA; "integrated" in this context most plausibly means tested in simulation, not deployed in operational units) (Moderate confidence)

Drone-swarm and counter-UAS research (stage 3 — prototype). Beihang University filed a 2025 patent applying DeepSeek to improve drone-swarm decision-making against "low, slow, small" targets — military shorthand for drones and light aircraft. Reuters reviewed approximately two dozen tenders and patents indicating efforts to integrate AI into drones for target recognition, tracking, and formation cooperation. A November 2024 PLA tender sought AI-powered robot dogs to scout in packs and clear explosive hazards; fulfillment status is not publicly confirmed. (High confidence on patent existence; Moderate confidence on operational status)

Norinco P60 autonomous combat-support vehicle (stage 4 — demonstration). Reuters reported that Norinco unveiled the P60, a military vehicle capable of autonomous combat-support operations at 50 km/h, described as powered by DeepSeek. The technical basis of "powered by DeepSeek" has not been publicly specified: it is not known whether the vehicle uses onboard inference, cloud-linked processing, or a fine-tuned distilled variant, and no technical specification has been released. Communist Party officials — unnamed in Reuters' reporting — endorsed the release in press statements. (unnamed attribution for the official endorsement; Reuters' characterization of official statements is accepted at face value but the individuals are not identified) Reuters cautioned that specifics of next-generation weapons systems are state secrets and that it could not establish whether all products reviewed were built or operational. Researchers at Landship Information Technology, which integrates AI into military vehicles including Norinco's, claimed in a February 2025 white paper that their Huawei-chip-based technology can rapidly identify targets from satellite imagery while coordinating with radars and aircraft. (single source — Landship marketing white paper; marketing materials are treated as promotional claims, not verified capability assessments) (Moderate confidence on demonstration; Low confidence on operational fielding)

OSINT and intelligence applications (stage 2–3 — acquisition/prototype). Recorded Future's Insikt Group reported in June 2025 that a Chinese defense contractor claimed to have provided a DeepSeek-based OSINT model to the PLA. Insikt also documented PLA patent applications covering generative-AI methods for generating OSINT products, processing satellite imagery, supporting event extraction, and processing event data. (Moderate confidence on contractor claim; High confidence on patent existence)

Cyber operations (stage 2 — acquisition documented). The Anhui Province military unit cyber range RFP (March 2025) is the cleanest documented case of DeepSeek being procured specifically for offensive cyber capability generation. (High confidence — ETO primary-source review)

Electronic warfare (stage 3 — research). NUDT's College of Electronic Countermeasures has published 2025 papers by Wang Benkun, Fang Mingxing, Ding Feng, and Meng Lingjie on DeepSeek for electromagnetic situation awareness. Researchers Li Wei, Liu Hongcheng, Gao Xiaokuan, and Luo Shuangquan at NUDT have published on DeepSeek for military command capability. Ma Yuexuan, Qi Jiayue, and Zhu Weiyu have published on large models in UAV game confrontation. These are doctrinal-academic claims, not evidence of fielded systems. (Moderate confidence — peer-reviewed but author-affiliated with PLA institutions)

No public evidence supports DeepSeek use in live combat or kinetic targeting. The Chinese Ministry of National Defense, DeepSeek, Norinco, and Huawei all declined to respond to Reuters' requests for comment. Specifics of next-generation weapons systems are state secrets under PRC law.

The Semiconductor Squeeze: Why Operational Scaling Is Compute-Bound

The export-control architecture targeting Chinese AI compute now spans four major rulemakings. The Bureau of Industry and Security's October 17, 2023 rules reinforced earlier October 7, 2022 controls, restricting advanced computing chips and supercomputing items to arms-embargoed destinations, removing interconnect bandwidth as the sole workaround metric, adding a performance-density threshold, and addressing third-country and foreign-subsidiary circumvention. BIS explicitly linked the controls to PRC military-civil fusion. The December 2024 BIS package added controls on 24 types of semiconductor manufacturing equipment, three types of software tools, high-bandwidth memory, and 140 Entity List additions, with BIS stating that HBM is critical to AI training and inference at scale. In April 2025, NVIDIA H20 and AMD MI308 faced new licensing requirements. BIS announced case-by-case review for NVIDIA H200 and AMD MI325X under security requirements. (High confidence — BIS rulemaking documents, References 9 and 10)

DeepSeek's compute stack. SemiAnalysis estimated that DeepSeek/High-Flyer's computing infrastructure includes at least 10,000 H100s, 10,000 H800s, 30,000 H20s, and 10,000 A100s, with approximately $1.63 billion in GPU server capital expenditures. (single source — SemiAnalysis proprietary estimate; not independently audited; used in alternative-hypothesis reasoning below but should be treated with corresponding skepticism) DeepSeek's own V3 Technical Report states V3 is a 671B-parameter mixture-of-experts model with 37B parameters activated per token, requiring 2.788 million H800 GPU-hours for full training. R1 is similarly a 671B MoE model with 37B activated parameters; distilled variants at 1.5B, 7B, 8B, 14B, 32B, and 70B are publicly released. (High confidence — DeepSeek V3 Technical Report, Reference 8)

Huawei Ascend as the domestic alternative — and its limits. CSIS analyst Gregory C. Allen reported in March 2025 that Huawei used shell companies to obtain more than 2 million Ascend 910B logic dies from TSMC, enough in principle for 1 million Ascend 910C chips, but that packaging yield and HBM availability were critical constraints. CSIS reported estimates from unnamed industry sources that approximately 75% of Ascend 910C chips survived advanced packaging. (unnamed industry sources — the 75% yield figure is treated as an informed estimate, not a verified measurement) SMIC's 7nm Ascend 910B yield was reported by CSIS sources at around 20%, while the Financial Times reported approximately 40%; CSIS disputed the FT figure as possibly including degraded but functional chips. (contested — CSIS unnamed industry sources vs. FT reporting; neither figure is independently verified)

RAND analysts reported in August 2025 that Huawei's Ascend 910B chips use older HBM2E memory, offering only two-thirds the memory capacity and 40% of the bandwidth of NVIDIA's H20. The newer Ascend 910C offers 80% of H20 bandwidth but still uses HBM2E — two generations behind the most advanced AI chips. This gap is particularly important for reasoning models and inference, where memory bandwidth is critical. The Council on Foreign Relations published analysis in late 2025 — the precise publication date is not confirmed in this evidence base and is flagged accordingly — finding that Huawei's publicly disclosed roadmap shows its next-generation 2026 chip will actually be worse than its best chip today, and that Huawei could produce only around 4% of the aggregate AI computing power NVIDIA produces. (temporal flag — CFR publication date not independently confirmed; finding treated as a late-2025 estimate, not a settled 2025 fact) SemiAnalysis assessed Huawei could produce as many as 1.5 million AI chip dies in 2025 but only 200,000–300,000 completed chips due to HBM shortages. (single source — SemiAnalysis proprietary estimate)

CSIS also reported that Huawei Ascend chips remained underutilized due to weaker software compatibility with Huawei's CANN ecosystem compared with NVIDIA's CUDA. A Beijing-based chip investor, unnamed in the CSIS report, was quoted saying the bottleneck was "figuring out how to make them work in a cluster." (unnamed individual — the quote is treated as illustrative of a documented software-integration problem, not as independent evidence of its severity) DeepSeek's adaptation to Ascend for inference improves the situation but does not close the cluster-integration gap.

Architectural workarounds. DeepSeek's MoE design materially lowers the floor. The full 671B model is memory-heavy, but the 32B and 70B distilled variants run on local servers, tactical headquarters, training centers, and hospitals. The Lhasa procurement's "DeepSeek-R1 32B or above" specification fits this pattern: it is not an attempt to deploy the full 671B model at the tactical edge. The Foundation for Defense of Democracies reported in late 2025 — precise publication date not confirmed in this evidence base and flagged accordingly — that DeepSeek and other major Chinese AI firms have struggled to improve their models due to lack of advanced compute, pursuing parallel computing across more chips and longer training cycles. (temporal flag — FDD publication date not independently confirmed)

Smuggling and evasion. Reuters reported on June 23, 2025 that the unnamed State Department official alleged DeepSeek sought to use Southeast Asian shell companies and remote data centers to access NVIDIA chips barred from China, and that DeepSeek had access to "large volumes" of H100s. NVIDIA told Reuters its review indicated DeepSeek used lawfully acquired H800 products, not H100s. (contested — U.S. official allegation vs. NVIDIA company review; neither claim is independently verified) Tom's Hardware reported a U.S. indictment in August 2025 and a subsequent investigative report — the precise date of the investigative report is not confirmed in this evidence base and is flagged accordingly — alleging Supermicro-tied executives used a Thailand government entity to route restricted GPUs to Chinese buyers including Alibaba. (temporal flag — the investigative report date is not independently confirmed; the August 2025 indictment is treated as the verified anchor point)

Aggregate compute gap. RAND assessed in August 2025 that the United States has approximately a tenfold advantage over China in total compute capacity. The Institute for Progress published analysis in late 2025 — precise publication date not confirmed in this evidence base and flagged accordingly — estimating that China's output of AI chips will reach only 1–4% of U.S. production in 2025 and 1–2% in 2026. (single source — IFP analysis; temporal flag on publication date; the 2026 figure is a projection, not an observation) The International Energy Agency estimated in 2025 that data centers consumed about 415 TWh globally in 2024 (1.5% of global electricity), projecting Chinese data-center consumption to rise approximately 175 TWh (170%) from 2024 to 2030. (IEA 2030 figure is a projection, not an observation; treated accordingly)

Doctrinal Framing: Intelligentized Warfare Meets Algorithmic Sovereignty

PLA doctrinal literature and institutional publications treat DeepSeek not as a single product but as a system architecture combining a large-scale reasoning core with modular and domain-specific layers, envisioned for integration across the C4ISR chain. This characterization is drawn from the Jamestown Foundation's Cheung and Lau assessment and from PLA academic publications reviewed by CSET; it is not sourced to a single authoritative PLA document. The Jamestown Foundation's assessment frames DeepSeek adoption as the operational expression of "intelligentized warfare" (智能化战争) — a doctrine articulated in the PLA's Science of Military Strategy (2019 edition and successors). The state-controlled Guangming Daily stated in 2025 that DeepSeek plays an "increasingly crucial role in the military intelligentization process." (state-media signaling — Guangming Daily is treated as deliberate signaling rather than independent reporting; the statement reflects intended messaging, not an independent capability assessment)

PLA experts from the Army Artillery and Air Defense Academy (陆军炮兵防空兵学院) have publicly noted — in publications reviewed by CSET and Jamestown but not attributed to a named author or specific paper in the sources available to this product (unnamed attribution — the institutional affiliation is confirmed; the specific authors and publication venue are not identified in the available evidence base) — that the enormous scale of R1 precludes field deployment on small autonomous platforms, while the more compact V3 variant is better suited for edge use but has inferior reasoning depth and sometimes generates unstable or high-risk tactical outputs, such as over-aggressive target assignments. This is an unusually candid PLA institutional acknowledgment of model limitations.

The director of the Science and Technology Committee at AVIC's Shenyang Aircraft Design Institute — the individual's name is not confirmed in the sources available to this product and is deliberately withheld pending verification (deliberately anonymized pending name confirmation) — stated that his team is using DeepSeek and praised it for providing new methods for aeronautical research. The Shenyang Aircraft Design Institute was involved in the development of the J-15 and J-35 fighters.

The U.S. Department of Defense's 2025 annual report to Congress stated that in 2024, "China's commercial and academic AI sectors made progress on large language models and LLM-based reasoning models, which has narrowed the performance gap between China's models and the U.S. models currently leading the field."

Alternative Hypotheses

Alternative 1: PLA DeepSeek procurement is primarily performative, not operational. The argument: Chinese state and military entities routinely procure showcased domestic technologies for political signaling — particularly after January 2025, when DeepSeek became a national-prestige asset. The Reuters note that the P60 release was endorsed by Communist Party officials supports this reading. The two URL-anchored plap.mil.cn primary records total only approximately RMB 2.1 million in combined value — small for genuine force-modernization. Assessment: Partially credible for high-profile demonstrators like the P60 but does not explain the granular technical specifications in cyber-range and OSINT RFPs, which detail operational capabilities (intelligent penetration, crawler servers, 32B local deployment) inconsistent with pure signaling. A counterargument is that detailed RFPs are also consistent with showcase procurements — the specificity of a requirement does not by itself establish operational intent. On balance, the performative hypothesis is more plausible for platform demonstrators and less plausible for the cyber-range and OSINT procurement records, which specify capabilities with no obvious public-relations value. Accepting this alternative fully would move KJ-3 from Moderate to Low confidence but would not affect KJ-1 or KJ-2.

Alternative 2: DeepSeek's compute efficiency neutralizes the export-control constraint. The argument: If MoE architectures and distillation continue to reduce the compute floor, then HBM shortages and SMIC yield problems become less binding, and PLA force-wide deployment becomes achievable on domestic chips by 2027. SemiAnalysis's assessment that absent smuggling "Chinese models would be served on Huawei Ascend at scale" implies the architecture-compute tradeoff is real. Assessment: Plausible for inference workloads, not for frontier training. The CFR finding — flagged above as temporally uncertain — that Huawei's 2026 chip will be worse than its best current chip, combined with HBM2E being two generations behind, suggests the gap is widening on training-critical metrics. Accepting this alternative would move KJ-4 from High to Moderate confidence and KJ-5 from Moderate to Low confidence.

Alternative 3: The PLA is deliberately overstating DeepSeek integration as part of cognitive-domain operations targeting Western policymakers. The argument: PLA-affiliated publications about 43% decision-time improvements, 10,000 scenarios in 48 seconds, and 90% recall under jamming serve as deterrence signaling. The fact that Chinese MoD, DeepSeek, Norinco, and Huawei all declined Reuters comment is consistent with deliberate ambiguity. Assessment: Likely a contributing factor — particularly for Xi'an Technological University and NUDT performance claims — but does not explain the existence of procurement records with operational specifications, contract values, and named winners. Accepting this alternative would move KJ-2 and KJ-3 toward Low confidence but would not affect KJ-1 or KJ-7.

Policy Implications

Intelligence Gaps and Collection Requirements

PIR-1: What is the actual production yield and HBM availability supporting Huawei Ascend 910C in 2026, and what is the resulting ceiling on PLA AI inference capacity? This is the single variable that determines whether PLA DeepSeek deployment scales from rear-area to force-wide. Collection: OSINT on SMIC fab capacity (satellite imagery of new fab construction at SMIC South, equipment manifests, Samsung/SK Hynix HBM shipment routing), HUMINT on Huawei supply chain partners, financial intelligence on shell-company HBM procurement.

PIR-2: Has DeepSeek-R1 or any variant been integrated into a live PLA targeting cycle in exercises beyond simulation (e.g., 2025 Joint Sword, Eastern Theater drills)? Verified live exercise integration would shift KJ-2 from "non-combat" to "combat-support fielded." Collection: SIGINT on PLA exercise C4ISR traffic, OSINT on PLA Daily exercise reporting, IMINT on exercise battlefield activity correlating with reported AI-decision cycles.

PIR-3: What is the actual scale and unit-level distribution of Shanxi 100 Trust Information Technology's PLA contracts, and which PLA theaters or services are the primary buyers? Resolves whether DeepSeek adoption is concentrated in a specific theater command (e.g., Eastern for Taiwan contingency) or distributed nationally. Collection: OSINT on plap.mil.cn and ccgp.gov.cn full-text scraping, financial intelligence on 100 Trust revenue concentration, HUMINT on PLA logistics department procurement officers.

PIR-4: What is the status of NVIDIA chip smuggling networks routing through Singapore, Malaysia, and Thailand, and what volume of H100/H200/B100/B200 has actually reached Chinese AI labs in 2025–2026? Determines whether export controls are degrading PLA compute access or being routed around. Collection: customs and trade data analysis, financial intelligence on shell company structures, diplomatic reporting from regional partners, FBI/DOJ enforcement case files.

PIR-5: Are PLA-affiliated researchers fine-tuning DeepSeek on classified or operational data, and where is that fine-tuning physically conducted (Huawei Cloud, NUDT clusters, dedicated military data centers)? Determines whether DeepSeek-based PLA systems are general-purpose or operationally specialized. Collection: SIGINT on data-flow patterns to known PLA computing facilities, IMINT on power/cooling signatures at PLA data center sites, OSINT on academic-publication metadata.

PIR-6: What is the technical architecture of the Norinco P60's DeepSeek integration — specifically, whether inference is conducted onboard, via cloud link, or through a fine-tuned edge variant — and what is the vehicle's actual operational status beyond the 2025 demonstration? Resolution would shift the P60 from stage 4 (demonstration) to either stage 5 (operational fielding) or downgrade it to stage 3 (prototype). Collection: OSINT on Norinco technical publications, HUMINT on Landship Information Technology personnel, IMINT on P60 unit assignments.

What to Watch

Confidence Note

This product applies a five-tier evidence hierarchy described in the Methodology section. Confidence ratings reflect the quality and corroboration of the underlying evidence, not the severity of the finding. High confidence does not mean certainty; it means the finding is supported by multiple independent sources or primary documents and no credible contradicting evidence has been identified. Moderate confidence means the finding rests on a single credible source, contested evidence, or a logical inference from established facts. Low confidence means the finding is speculative, rests on a single unverified claim, or is contested by credible counter-evidence.

Key confidence boundaries in this product:

Named Actors

Researchers and academics:

Analysts and authors:

Government officials:

Corporations and institutions:

References

  1. PLA Procurement Network notice, "探索运用DeepSeek大模型处理数据模式构建与验证采购项目," Lhasa Tibet Autonomous Region unit, plap.mil.cn, April 2025. https://www.plap.mil.cn/freecms/site/juncai/ggxx/info/2025/8a1d02bb95320c300195ff21c6d10228.html

  2. PLA Procurement Network result notice, "DeepSeek大模型+教学工具集采购结果公示," Project No. 2025-JK22-F1015, plap.mil.cn, bids opened May 27, 2025. https://www.plap.mil.cn/freecms/site/juncai/ggxx/info/2025/8a1d01cb96b9416a0197195b3d6920c8.html

  3. Emerging Technology Observatory (ETO), Georgetown CSET, "The National Security Case for Limiting China's Access to Advanced Compute," eto.tech. https://eto.tech/blog/pla-procurement-case-for-limiting-china-advanced-compute/

  4. Sunny Cheung and Kai-shing Lau, "DeepSeek Use in PRC Military and Public Security Systems," Jamestown Foundation China Brief, October 27, 2025. https://jamestown.org/program/deepseek-use-in-prc-military-and-public-security-systems/

  5. Michael Martina and Stephen Nellis, "Exclusive: DeepSeek aids China's military and evaded export controls, US official says," Reuters, June 23, 2025. https://www.investing.com/news/world-news/exclusivedeepseek-aids-chinas-military-and-evaded-export-controls-us-official-says-4105141

  6. Eduardo Baptista and Fanny Potkin, "Robot dogs and AI drone swarms: how China could use DeepSeek for an era of war," Reuters, October 26–27, 2025. https://www.investing.com/news/economy-news/robot-dogs-and-ai-drone-swarms-how-china-could-use-deepseek-for-an-era-of-war-4309350

  7. Recorded Future Insikt Group, "Artificial Eyes: Generative AI in China's Military Intelligence," June 2025. https://www.recordedfuture.com/research/artificial-eyes-generative-ai-chinas-military-intelligence

  8. DeepSeek-AI, "DeepSeek-V3 Technical Report," arXiv:2412.19437, December 2024. https://arxiv.org/abs/2412.19437

  9. Bureau of Industry and Security (U.S. Department of Commerce), "Commerce Strengthens Restrictions on Advanced Computing Semiconductors and Semiconductor Manufacturing Equipment to the PRC," press release, October 17, 2023. https://www.bis.gov/press-release/commerce-strengthens-restrictions-advanced-computing-semiconductors-semiconductor-manufacturing-equipment

  10. Bureau of Industry and Security (U.S. Department of Commerce), "Commerce Strengthens Export Controls to Restrict China's Capability to Produce Advanced Semiconductors," press release, December 2024. https://www.bis.gov/press-release/commerce-strengthens-export-controls-restrict-chinas-capability-produce-advanced-semiconductors-military

  11. Gregory C. Allen, "DeepSeek, Huawei, Export Controls, and the Future of the U.S.-China AI Race," Center for Strategic and International Studies (CSIS), March 2025. https://www.csis.org/analysis/deepseek-huawei-export-controls-and-future-us-china-ai-race

  12. RAND Corporation, analysis of China's AI chip production capability and Huawei Ascend bandwidth specifications, August 2025. (No public URL confirmed in this evidence base; cited for HBM2E bandwidth figures and U.S.-China compute-gap assessment. Readers should verify against the RAND publication catalog at rand.org.)

  13. Council on Foreign Relations, analysis of Huawei AI chip roadmap and production capacity, late 2025. (Precise title, author, and URL not confirmed in this evidence base; publication date flagged as temporally uncertain. Readers should verify against cfr.org. Finding treated as a late-2025 estimate, not a settled fact.)

  14. International Energy Agency, "Electricity 2025" or equivalent annual report, 2025. Data cited: global data-center consumption approximately 415 TWh in 2024; Chinese data-center consumption projected to rise approximately 175 TWh (170%) from 2024 to 2030. (2030 projection is a modeled estimate, not an observation.) https://www.iea.org/reports/electricity-2025

  15. SemiAnalysis, estimates of DeepSeek/High-Flyer GPU infrastructure and Huawei Ascend production capacity, 2025. (Proprietary analysis; no public URL. Figures cited: ~$1.63 billion GPU server capex; 1.5 million Ascend dies producible in 2025 vs. 200,000–300,000 completed chips due to HBM shortages. Treated as single-source estimates throughout.)

  16. Foundation for Defense of Democracies, analysis of Chinese AI firm compute constraints, late 2025. (Precise title, author, and URL not confirmed in this evidence base; publication date flagged as temporally uncertain. Readers should verify against fdd.org.)

  17. Institute for Progress, analysis of China AI chip production as percentage of U.S. output, late 2025. (Precise title, author, and URL not confirmed in this evidence base; publication date flagged as temporally uncertain; 2026 figure is a projection. Readers should verify against ifp.org.)

  18. Bo Huang et al., "Large Language Model-Based Battalion-Level Command Integration System," Command Control & Simulation, March 12, 2025. (Chinese-language publication; author affiliated with National University of Defense Technology. No English-language URL confirmed; accessible via CNKI.)

  19. South China Morning Post, "Chinese military researchers use DeepSeek to generate 10,000 battlefield scenarios in 48 seconds," May 16, 2025. (Paywalled; Reuters noted it could not independently verify the performance claim.)

  20. U.S. Department of Defense, Annual Report to Congress: Military and Security Developments Involving the People's Republic of China, 2025. https://www.defense.gov/News/Releases/ (Readers should verify the precise URL against the DoD press release page for the 2025 edition.)

  21. House Select Committee on the CCP, materials related to NVIDIA-DeepSeek technical support allegations and DeepSeek cybersecurity vulnerabilities, 2025. (Specific document title and URL not confirmed in this evidence base; accessible via selectcommittee.house.gov.)

  22. Ren Hao, statement on DeepSeek-R1 deployment at 301 Hospital, CHIMA (China Health Information Management Association) website, 2025. (Single-source statement by one named engineer; not independently corroborated. URL not confirmed in this evidence base.)

  23. Landship Information Technology, white paper on AI integration in military vehicles, February 2025. (Marketing document; treated as promotional claim, not verified capability assessment. No public URL confirmed.)

  24. Tom's Hardware, reporting on U.S. indictment of Supermicro-tied executives for GPU routing through Thailand, August 2025 indictment anchor; subsequent investigative report date not confirmed. (Temporal flag on investigative report date. Readers should verify at tomshardware.com.)