Finished Intelligence Assessment | Defense Technology & Force Modernization
Reporting Period: February 2025 – Present (Analytical Cutoff: May 2025)
Classification: UNCLASSIFIED // Open-Source Derived
Source Reliability Note: This assessment draws exclusively on open-source material. Confidence levels reflect source quality, corroboration, and analytic tradecraft—not classified access. PLA-controlled outlets (PLA Daily, CCTV, CNKI conference papers) are treated as primary evidence of intent and claimed capability, not as independently verified operational reporting. Readers should apply corresponding source-reliability discounts throughout.
KJ-1. The PLA has moved from doctrinal aspiration to active procurement, research integration, and prototype demonstration of DeepSeek and other domestic large language models across command-and-control, battlefield decision support, drone swarming, and autonomous ground systems. This transition is documented in verifiable procurement records, peer-reviewed research, and public demonstrations within roughly twelve months of DeepSeek-R1's open-weight release in January 2025. (Medium-High confidence; based on corroborated open-source procurement and research evidence, with caveats on PLA-controlled reporting)
KJ-2. Available open-source evidence supports R&D, prototype, and demonstration-stage integration but does not support claims that DeepSeek is currently controlling lethal autonomous PLA swarms in operational combat units. The most accurate characterization is "confirmed procurement and demonstration; unverified large-scale operational fielding." (High confidence in the negative bound; the absence of evidence for operational deployment is itself analytically significant)
KJ-3. U.S. export controls have imposed real costs on PRC AI compute access but have been partially circumvented—DOJ enforcement actions through mid-2025 document hundreds of millions of dollars in attempted illicit transfers of restricted Nvidia GPUs. The full scale of successful evasion remains uncertain; published estimates carry wide confidence intervals and do not disaggregate PLA-specific from broader PRC commercial AI consumption. (Medium confidence)
KJ-4. DeepSeek's optimization work targeting Huawei Ascend processors, documented in engineering publications through early 2025, represents a strategically significant trend: China is actively building an AI software stack with reduced U.S. hardware dependencies. The maturity and military relevance of this stack remain uncertain. (Medium confidence; trajectory is clear, endpoint is not)
KJ-5. China's open-weight release strategy creates a structural vulnerability in U.S. export control architecture. The Entity List cannot prevent PLA-affiliated entities from downloading frontier-class models from public repositories; this gap has no current legal remedy and is analytically distinct from compute-access controls. (High confidence)
KJ-6. Assessed PLA AI integration represents acceleration of a pre-existing trajectory, not a discontinuous break. The PLA was integrating commercial AI into C4ISR, unmanned systems, and decision support before DeepSeek's emergence; DeepSeek's open-weight availability and efficiency characteristics have lowered barriers and accelerated timelines but did not originate the program. (Medium-High confidence)
Assessing DeepSeek's significance requires establishing what PLA AI integration looked like before January 2025. This baseline is essential to distinguish acceleration from origination.
Pre-2025 foundations. Xi Jinping designated intelligentization (智能化) a PLA modernization priority in 2017 and formalized it in the 2020 edition of The Science of Military Strategy. By 2023–2024, the PLA had already:
The 2025 DoD China Military Power Report assesses that the PLA conducted AI experimentation through 2024 across unmanned systems, ISR analysis, decision support, cyber operations, and influence operations, and that PRC AI development was constrained in 2024 by limited high-performance accelerator access—prompting workarounds via lower-grade chip optimization, indigenous alternatives, and licit and illicit acquisition.
What DeepSeek changed. DeepSeek-R1's open-weight release in January 2025 lowered three barriers simultaneously: cost (no licensing fees), deployment flexibility (local hosting without vendor dependency), and reasoning capability (chain-of-thought performance competitive with frontier closed models). The Reuters review finding approximately a dozen PLA tenders referencing DeepSeek versus only one referencing Alibaba Qwen—filed within months of R1's release—indicates that DeepSeek's specific characteristics, not merely the availability of domestic LLMs, drove rapid PLA adoption. The shift is real; it is an acceleration, not an origin.
The most authoritative open-source evidence comes from the PLA Procurement Network (plap.mil.cn). The following contracts are drawn from Reuters reporting and Jamestown Foundation analysis of that database; they represent the verifiable floor of PLA DeepSeek procurement, not a comprehensive inventory.
Confirmed contracts (corroborated by multiple open-source outlets):
Contracts reported by single sources (treat with caution):
Additional tenders referencing DeepSeek for theatre-level C2 and drone-swarm applications have been reported in Chinese-language defense media. These are noted in relevant sections below with appropriate confidence caveats. Contract numbers and vendor names from single Chinese-language sources are not reproduced here as primary evidence given inability to independently verify.
A Reuters review found DeepSeek referenced in approximately a dozen PLA tenders filed in 2025—versus only one referencing Alibaba's Qwen—indicating clear PLA preference within the documented procurement record. The Jamestown Foundation's 2025 review of the PLA Procurement Network (Sunny Cheung) characterizes DeepSeek not as a single product but as an evolving system architecture intended for integration across the C4ISR chain.
Note on vendor attribution. Vendor names cited in PLA procurement records should not be assumed to correspond to DeepSeek (Hangzhou DeepSeek Artificial Intelligence Co., Ltd., owned by High-Flyer Quant) as a corporate entity. PLA contracts reference DeepSeek models—open-weight software—deployed by third-party integrators. DeepSeek/High-Flyer's public posture does not explicitly endorse military applications; the company has not publicly commented on PLA procurement. Whether PLA contracts represent endorsed integration or downstream redistribution of open-weight models by independent vendors is analytically unresolved.
A CSET analysis of 338 entities winning AI-related PLA contracts—covering the period through approximately 2024, predating the DeepSeek-specific procurement surge—found that close to three-quarters were nontraditional vendors with no self-reported state-ownership ties. This inverts the legacy SOE-dominated PLA procurement model and complicates U.S. sanctions targeting. Whether this pattern holds specifically for DeepSeek-era contracts has not been independently confirmed; the CSET finding is offered as structural context, not as a characterization of current DeepSeek procurement specifically.
A case reported by Jamestown: a Shanxi-based private IT firm identified as holding classified-project clearance and DeepSeek integration contracts was reportedly banned from military procurement for one year for falsified bidding materials. If accurate, this indicates the PLA is accepting elevated counterintelligence and quality-assurance risk in exchange for procurement speed—a pattern consistent with the broader non-SOE vendor trend.
Battalion-Level Decision Support. A peer-reviewed paper from researchers affiliated with the National University of Defence Technology (NUDT) describes an LLM-based command information platform that processes battlefield reports, identifies critical information requirements, and presents commanders with decision options. In simulation, the system reportedly outperforms human commanders on speed of options generation. The paper is cited in Chinese-language defense journals; independent English-language verification of specific performance claims has not been established.
Rapid Battlefield Planning. Xi'an Technological University researchers claimed in May 2025 that a DeepSeek-powered system evaluated 10,000 battlefield scenarios in 48 seconds—a task estimated at 48 hours for a human staff. Reuters explicitly noted it could not independently verify this claim. It is treated here as directionally indicative of research ambition, not as a confirmed capability benchmark.
Multimodal C4ISR Architecture. A 2025 article in Command Information System and Technology describes a distributed architecture using DeepSeek-R1 and DeepSeek-VL2 for multimodal fusion of reconnaissance imagery, spectrum data, audio, navigation, and structured operational data, with DeepSeek-R1 generating commander-reviewable courses of action. This is a research publication describing a proposed architecture; it does not confirm operational deployment.
Rocket Force Decision Support. Chinese-language conference papers report DeepSeek deployment for missile salvo planning simulations at a Rocket Force installation, evaluating large numbers of trajectories in seconds. (Low-Medium confidence; single Chinese-language source, specific performance figures unverifiable, and PLA-controlled reporting incentivizes capability inflation. Treat as indicative of research direction, not confirmed operational capability.)
These represent the highest-confidence integrations and serve as the operational baseline from which combat-system integration should be assessed:
These applications are significant as proof of institutional adoption and local-deployment capability, but they are administratively distinct from tactical or operational military systems.
Evidence does not support uniform PLA enthusiasm. PLA Army Artillery and Air Defense Academy experts have noted that DeepSeek-R1's parameter scale precludes edge deployment on small autonomous platforms; the lighter V3 variant is more deployable but has been characterized in PLA research as producing unstable tactical outputs in some simulation conditions. A January 2025 PLA Daily article warned explicitly that AI "cannot replace human decision-making on the battlefield."
Identified PLA-side technical concerns in the open-source record include: latency variance under network stress, output irreproducibility, adversarial vulnerability from open-source model weights, and black-box decision chains incompatible with accountability requirements. These concerns are consistent with challenges documented in Western military AI integration programs and should be weighted against optimistic capability claims from the same PLA-controlled outlets.
In February 2025, Norinco unveiled the P60 autonomous combat-support vehicle, publicly identified as DeepSeek-powered by integrator Landship Information Technology. Landship's associated white paper claims DeepSeek-based capability to rapidly identify targets from satellite imagery while coordinating with radars and aircraft. This claim originates from vendor promotional material and has not been independently verified. No evidence of serial production or unit fielding has surfaced. The P60 demonstration is significant as a public signal of PLA-industry intent; it should not be treated as evidence of fielded capability.
Beihang University Patent (2025). Beihang University—one of the "Seven Sons of National Defense" institutions with formal PLA research ties—filed a 2025 patent applying DeepSeek to drone-swarm decision-making against "low, slow, small" (LSS) targets, the standard PRC term for hostile drones and light aircraft. Patent filing indicates research investment; it does not confirm prototype testing or deployment.
NUDT Swarm Demonstration (January 2026, per CCTV). CCTV broadcast footage of a single PLA operator directing more than 200 fixed-wing drones launched simultaneously from multiple vehicles, with reported autonomous task division across reconnaissance, jamming, and strike roles. CCTV is a PLA-controlled outlet with strong incentives to publicize capability for deterrence signaling; the footage should be treated as evidence of a demonstration event, not necessarily of routine operational capability. Independent verification of the autonomous coordination claims embedded in the broadcast narration is not available.
LLM Integration in Swarm Research. Chinese-language defense publications, including papers from NUDT-affiliated researchers, describe integration of DeepSeek-series models into swarm coordination modules, with reported performance metrics for anti-jamming path replanning. (Medium confidence in the existence of this research program; low confidence in specific quantitative metrics, which display precision levels characteristic of promotional reporting and cannot be independently verified. Named authors and DOIs from single Chinese-language sources are noted but not reproduced as primary evidence.)
Atlas Swarm System. Public demonstrations document a multi-vehicle architecture capable of launching dozens of drones with onboard autonomy for task allocation, target re-identification, and dynamic rerouting under single-operator supervision. The specific role of DeepSeek versus other AI components in this system has not been independently confirmed.
PLA Exercise Reporting. PLA-controlled media has reported theatre-level exercises integrating AI-enabled drone swarms and robot dogs against LSS targets, with claimed decision-generation and hit-rate metrics. (Low-Medium confidence; single-source PLA-controlled reporting; specific figures are unverifiable and should be treated as capability signaling rather than confirmed performance data.)
PLA procurement tenders have solicited AI-powered quadruped robots for pack scouting, threat identification, and explosive-ordnance clearance. Unitree-manufactured quadrupeds have been documented in PLA exercises and public demonstrations. The specific AI models integrated into these platforms have not been publicly confirmed.
A March 2025 paper from Beijing Institute of Technology researchers, funded under a military-civil fusion research program, explicitly proposes full-chain distributed autonomous decision-making from high-value target identification to strike for drone swarms in urban warfare. This is a research proposal, not a description of a fielded system. Its significance lies in the doctrinal direction it signals: PLA-affiliated researchers are actively developing the conceptual and technical architecture for lethal autonomy, and this work is funded through formal military-civil fusion channels.
China's official position on lethal autonomous weapons systems sets the threshold for prohibited LAWS sufficiently high to permit development of systems that engage without real-time human judgment in the loop. This doctrinal permissiveness, combined with the research trajectory documented above, represents the most consequential long-term signal in the dataset.
Xi Jinping established intelligentization (智能化) as a PLA modernization priority in 2020, formalized in the 2020 edition of The Science of Military Strategy. The 2025 DoD China Military Power Report assesses that the PLA continued AI experimentation through 2024 across unmanned systems, ISR analysis, decision support, cyber operations, and influence operations.
Key doctrinal developments in the reporting period:
PLA Daily articles in early 2025 cited Ukraine and Middle East conflicts to argue that AI-driven swarming will supersede the current FPV-drone era, with AI overcoming electronic warfare and satellite-communications vulnerabilities. This framing reflects PLA doctrinal learning from observed conflicts, not confirmed PLA capability.
Articles in Ordnance Industry Automation by NUDT and PLA National Aviation University authors outline shipborne unmanned systems for amphibious landings, including manned-unmanned teaming with carrier-based drones and amphibious autonomous urban-combat platforms. The amphibious focus is directly relevant to Taiwan contingency planning as a doctrinal signal.
PLA theoretical writing on "Dissipative Warfare" reframes victory not as material attrition but as maintaining order in one's own decision system while inducing disorder in the adversary's—an explicit theoretical justification for AI-enabled C2 and for targeting adversary AI systems.
The PLA articulates a long-term transition from a "human-centric force with unmanned systems in support" to "a force centered on unmanned systems with humans in support." The timeline for this transition is not specified in open-source doctrine and should not be assumed to be imminent.
Acknowledged limitations in PLA doctrine and capability. The PLA lacks the operational data volume the U.S. has accumulated through decades of expeditionary warfare—a gap that affects AI training data quality for combat applications. The PLA also faces unresolved tension between the decentralized authority that effective AI-enabled operations require and its centralized command culture. These are structural constraints that open-source PLA research acknowledges; they are not resolved by model capability alone.
Information operations caveat. PLA-controlled media, including PLA Daily, CCTV, and official conference proceedings, have strong institutional incentives to publicize AI capability for deterrence signaling and domestic legitimacy. This does not mean reported capabilities are fabricated, but it means that PLA-sourced performance claims require independent corroboration before being treated as ground truth. This assessment applies that filter throughout; readers should apply it to any single-source PLA reporting not flagged here.
BIS rules dating from October 7, 2022, reinforced October 17, 2023, and clarified in April 2024, restrict PRC access to advanced computing chips and semiconductor manufacturing tools, explicitly citing military-civil fusion and AI military applications. The 2025 DoD CMPR assesses that PRC AI development was constrained in 2024 by limited high-performance accelerator access, prompting workarounds via lower-grade chip optimization, indigenous alternatives, and licit and illicit acquisition.
The following cases are drawn from DOJ indictments and BIS enforcement actions available in open-source reporting. Indictments allege conduct; they do not constitute convictions. Specific figures represent alleged amounts in charging documents.
Operation Gatekeeper (DOJ, December 2025). Indictments alleging at least $160 million in attempted illegal exports of Nvidia H100 and H200 GPUs from October 2024 to May 2025. Alleged methods included straw purchasers, false paperwork, relabeling of Nvidia GPUs as generic computer parts, and shipments to PRC and Hong Kong destinations.
Super Micro-Related Indictment (early 2026, per Reuters). Charges alleging conspiracy to ship Nvidia-equipped AI servers to PRC customers via Taiwan and Malaysia, using forged documents and dummy equipment to defeat audits. Alleged value in the billions of dollars. (Reported by Reuters; specific figures and timeline treated as alleged, not confirmed.)
ALX Solutions (DOJ, August 2025). Two PRC nationals charged with illegal exports from October 2022 through July 2025, including multiple shipments via Singapore and Malaysia freight forwarders.
Linxing Tech (reported April 2026). A Shenzhen-based firm reported to have PLA procurement relationships was indicted for smuggling Nvidia H200 GPUs via Malaysia, falsified as network cards. Reporting alleges the destination was a PLA-affiliated research laboratory and that the GPUs were used to train swarm-related AI models. (Medium confidence; reported by multiple outlets but specific linkage to named PLA programs rests on seized documents not independently reviewed. This is the most direct alleged linkage between a smuggling network and a specific PLA AI program in the open-source record, and warrants priority collection attention.)
Epoch AI has published estimates of H100-equivalent GPUs smuggled to China through end-2025, with a reported 90% confidence interval spanning roughly 290,000 to 1.6 million units and a median estimate of approximately 660,000. Three critical caveats apply:
First, this range is extremely wide—the upper bound is more than five times the lower bound—reflecting genuine uncertainty, not precision.
Second, these estimates cover all PRC AI compute acquisition through illicit channels and do not disaggregate PLA-specific from commercial or civilian AI consumption. The fraction attributable to PLA programs is unknown.
Third, the estimates are produced by a research organization using indirect inference methods; they have not been validated against classified intelligence.
The estimates are useful as an order-of-magnitude indicator that export control evasion is occurring at significant scale. They should not be cited as precise figures or as evidence of PLA-specific compute accumulation.
DeepSeek's engineering teams have published work on optimizing model inference for Huawei's Ascend processors via the CANN framework, reducing dependence on Nvidia CUDA. This is documented in engineering publications and represents a genuine strategic development: China is actively building an AI software stack with reduced U.S. hardware dependencies.
The performance gap between Huawei Ascend and Nvidia H100 is real. Published comparisons suggest Ascend 910-series chips deliver meaningfully lower throughput on standard AI workloads, though specific figures vary by benchmark, workload type, and precision regime. "DeepSeek's own engineering benchmarks" is not a citable independent source; readers should consult published third-party benchmark comparisons for specific figures. The gap is narrowing; the rate of narrowing is uncertain.
The strategic implication is directional and significant: if China achieves a competitive domestic AI chip stack, the long-term leverage of U.S. compute controls is structurally reduced. This trajectory is more important than any specific performance figure.
U.S. export control policy on advanced AI chips has undergone multiple adjustments in the 2024–2025 period, including expansions of restricted chip categories, modifications to licensing frameworks, and legislative proposals to mandate physical location-tracking hardware on advanced chips. The policy environment reflects genuine interagency disagreement on optimal restriction depth and on the tradeoff between denial and economic costs to U.S. semiconductor firms.
Multiple reports indicate that PRC AI developers have used large numbers of accounts to systematically extract outputs from U.S. frontier AI models, including Claude and GPT-series models, to train domestic models. OpenAI has separately alleged that PRC-linked actors circumvented access restrictions to harvest model outputs at scale. The specific companies involved, the scale of extraction, and the degree to which this constitutes a meaningful capability transfer are disputed; Anthropic and OpenAI have not published detailed public disclosures that would allow independent verification of specific figures.
The analytic significance of this vector is that it is orthogonal to hardware controls: even if compute access is restricted, behavioral and reasoning capabilities of frontier models can potentially be transferred through output extraction. The military relevance of distilled civilian model capabilities is not established in open-source reporting and should not be assumed.
No publicly available satellite imagery has been definitively linked to DeepSeek-specific military infrastructure. AI compute facilities lack distinctive overhead signatures absent paired procurement or telemetry data. Commercial imagery has documented unmanned-platform test activity at known PLA test ranges, including facility expansions consistent with increased unmanned-systems testing, but these are platform indicators, not payload indicators—they do not confirm which AI systems are being tested.
A notable adjacency: PRC commercial satellite-AI firms have published AI-tagged imagery of foreign military installations and infrastructure. Pentagon assessments have alleged that some of this commercial OSINT capability has been used for operational targeting by state and non-state actors. The specific role of DeepSeek-series models in commercial satellite-AI applications has not been confirmed in open-source reporting.
Prior reporting on PLA DeepSeek integration has produced genuine disagreement, partly because "operational deployment" is used inconsistently across sources. The following framework distinguishes tiers of integration by evidence quality:
| Tier | Characterization | Evidence Basis | Confidence |
|---|---|---|---|
| 1 | Procurement, R&D contracts, academic integration, non-combat administrative use | Multiple corroborated open-source procurement records; peer-reviewed publications; SCMP, Reuters, Jamestown reporting | High |
| 2 | Prototype demonstration and laboratory testing, including swarm demonstrations with claimed LLM integration | CCTV footage (PLA-controlled); Chinese-language defense publications; Norinco P60 public unveiling | Medium — demonstrations confirmed; specific AI integration claims unverified |
| 3 | Theatre-level exercise integration and unit-level decision-support deployment | Single-source PLA-controlled reporting; unverified contract documents | Low-Medium — reported but not independently corroborated |
| 4 | End-to-end lethal autonomous control in operational combat units | No open-source evidence | Not confirmed |
Synthesized judgment: The accurate characterization of PLA DeepSeek integration as of mid-2025 is Tier 1 confirmed, Tier 2 partially confirmed with caveats, Tier 3 reported but unverified, and Tier 4 not evidenced. The gap between Tier 3 and Tier 4 is the single most important intelligence question for the next reporting period.
On prior analytic contradictions. Some prior assessments characterized PLA DeepSeek integration as either "no verified operational deployment" or "confirmed operational testing." Both characterizations are defensible depending on how "operational" is defined. The contradiction is partly semantic and partly reflects genuine evidence gaps. This assessment does not claim to resolve those gaps; it maps them explicitly. Claims that specific evidence "resolves" the central question should be treated skeptically when that evidence is itself single-source PLA-controlled reporting.
1. Export controls remain necessary but face structural limits. Documented smuggling operations confirm that hardware controls are being evaded at meaningful scale. Simultaneously, China's active development of a domestic AI chip and software stack means that hardware controls buy time rather than permanent denial. The logical policy evolution is toward embedded telemetry and end-use verification mechanisms, though these face counter-tampering challenges and require international coordination to be effective.
2. The open-weight problem has no current legal remedy. PLA-affiliated entities can download DeepSeek-R1 and subsequent open-weight releases from public repositories regardless of Entity List status. Export controls on model weights are legally and technically distinct from chip controls and remain underdeveloped. Any comprehensive U.S. response must address model-weight diffusion as a distinct vector.
3. Vendor ecosystem complexity undermines sanctions targeting. The dominance of non-SOE vendors in PLA AI procurement means that traditional sanctions targeting of state-owned defense firms will not capture the majority of PLA AI integration activity. Effective targeting requires mapping the private-sector vendor ecosystem, which is less visible and more dynamic than the SOE procurement base.
4. DoD unmanned-systems programs face a competitive timeline. PLA swarm development is progressing on a documented trajectory. The U.S. unmanned-mass advantage that programs like Replicator were designed to establish should be assessed against PLA development timelines, not assumed. Specific comparative assessments require classified data not available in open-source reporting.
5. Taiwan-contingency relevance. PLA procurement records referencing aviation and flight-management applications in Yunnan, doctrinal articles on amphibious unmanned systems, and theatre-level exercise reporting collectively indicate that AI-enabled C4ISR is being developed on platforms and in theatres relevant to a Taiwan contingency. This is a doctrinal and procurement signal, not confirmed operational deployment. The distinction matters for policy response calibration.
6. Deterrence signaling complicates assessment. PRC information operations have strong incentives to publicize PLA AI capability. Some portion of the open-source reporting on PLA DeepSeek integration may be intended to signal capability for deterrence purposes rather than to accurately describe operational status. Analysts should weight this possibility when evaluating PLA-controlled sources, particularly for claims that cannot be independently corroborated.
The following gaps represent the highest-priority collection requirements emerging from this assessment. They are listed in order of analytic urgency.
PIR-1 (Critical). What is the actual operational status of DeepSeek-series LLMs in PLA combat units—specifically, are any units authorized to use LLM-generated outputs for targeting or fire-control decisions without human review? This is the Tier 3-to-Tier 4 gap identified throughout this assessment and the single most consequential unresolved question.
PIR-2 (High). What fraction of smuggled advanced compute (Nvidia H100/H200 equivalents) is being directed to PLA or PLA-affiliated research programs versus commercial PRC AI development? Current estimates do not disaggregate these populations, making it impossible to assess PLA-specific compute accumulation.
PIR-3 (High). What is the actual performance of Huawei Ascend 910-series chips on military-relevant AI workloads (inference latency, throughput under constrained power budgets, performance on multimodal sensor fusion tasks)? Published benchmarks are insufficient for military capability assessment.
PIR-4 (High). What is the production rate and yield of Huawei Ascend chips under current semiconductor manufacturing constraints? The strategic significance of the Ascend/DeepSeek stack depends critically on whether Huawei can produce chips at scale.
PIR-5 (Medium). What is the actual corporate relationship between DeepSeek/High-Flyer and PLA procurement entities? Are PLA contracts for DeepSeek models endorsed by the company, or do they represent downstream redistribution of open-weight models by independent vendors without company knowledge or participation?
PIR-6 (Medium). What AI decision-support systems were in operational use in PLA combat units before January 2025, and how has DeepSeek integration changed those systems? Establishing this baseline is essential for assessing the magnitude of the DeepSeek-driven change.
PIR-7 (Medium). What are the actual technical specifications and operational parameters of the swarm coordination systems demonstrated in PLA exercises? CCTV footage and PLA-controlled reporting provide insufficient technical detail to assess capability.
PIR-8 (Lower). What is the scale and systematic organization, if any, of PRC efforts to extract capabilities from U.S. frontier AI models through output harvesting? Current reporting is fragmentary and contested.
China's integration of DeepSeek AI into PLA systems is real, documented, and strategically significant—but the open-source evidence base supports a more cautious characterization than some reporting has offered. What is confirmed is a rapid and broad procurement and research integration program, spanning administrative applications, decision-support tools, and prototype autonomous systems, that accelerated sharply following DeepSeek-R1's open-weight release in January 2025. What is not confirmed is operational deployment of LLM-driven lethal autonomous systems in PLA combat units.
The most consequential developments are structural rather than tactical: the open-weight release model that makes export controls on model access legally unenforceable; the active development of a domestic AI stack reducing U.S. hardware leverage over time; and the doctrinal trajectory toward lethal autonomy that PLA-affiliated research is actively pursuing. These structural factors will shape the competitive landscape regardless of the current operational status of any specific system.
The central analytic uncertainty—whether PLA AI integration has crossed from demonstration and exercise use into operational combat deployment—cannot be resolved from open-source reporting alone. This assessment maps that uncertainty explicitly rather than resolving it artificially. Policymakers and operators should plan for a trajectory in which that gap closes, while avoiding assessments that treat the gap as already closed.
This assessment was produced from open-source materials only. All confidence levels reflect open-source evidence quality and corroboration. PLA-controlled sources are treated as evidence of intent and claimed capability, not as independently verified operational reporting. Specific quantitative claims from single Chinese-language sources are flagged throughout and should not be cited as confirmed figures without independent corroboration.