Global Military Affairs | Silicon Valley "Enlists": Accelerating the Militarization of AI

Silicon Valley "Enlists": Accelerating the Militarization of AI

■ Ke Weiyin

U.S. military personnel operating autonomous systems for target acquisition.

A Ukrainian soldier at the front line preparing to connect to the Starlink system.

The YFQ-44A "Collaborative Combat Aircraft," developed with participation from Anduril Industries, flying in formation with a crewed fighter jet.

Recently, the U.S. Department of Defense announced cooperation agreements with leading technology enterprises including SpaceX, OpenAI, Google, Microsoft, NVIDIA, Amazon Web Services (AWS), Oracle, and Reflection, deploying each company's AI technologies, computing models, cloud platforms, and communications security technologies comprehensively across U.S. military IL-6 Secret-level and IL-7 Top Secret-level networks, for lawful application in military scenarios including intelligence assessment, battlefield awareness, command decision-making, and all-domain operations. The large-scale entry of Silicon Valley-style innovation into the defense establishment has been vividly described by media as Silicon Valley "enlisting," marking the U.S. military's transformation toward AI-driven combat power entering a substantive and large-scale stage.

From the release of the Accelerating AI Strategy (《人工智能加速战略》) in January of this year to the conclusion of "enlistment" agreements with leading Silicon Valley technology enterprises, the Department of Defense completed a critical integration of the military AI supply chain in just a few months—at a pace far exceeding outside expectations—drawing intense international attention.

Foundational Restructuring: Eight Companies with Differentiated, Complementary Roles

The eight companies selected have been explicitly designated by the U.S. side as "military AI technology suppliers," with clearly defined and mutually complementary roles that together constitute a full military AI industrial chain.

At the algorithm and model layer, Reflection, Google, and OpenAI serve as the "brain" of the U.S. military's AI intelligence architecture, bearing responsibility for the most core intelligence analysis and decision-support functions. OpenAI's general-purpose large language model excels at multilingual intelligence translation, mass-text screening, complex information synthesis, and preliminary generation of operational plans, providing commanders with concise and actionable analytical conclusions. Google focuses on satellite remote-sensing image analysis, battlefield target recognition, and terrain situation assessment, responsible for identifying enemy force deployments, equipment positions, and fortification facilities. Reflection concentrates on the U.S. military's core requirement for "AI autonomous operations" (AI自主作战), specializing in military intelligent agent (智能Agent) automated execution systems to achieve full-process automation of intelligence assessment, resource scheduling, and after-action review.

At the computing power and hardware infrastructure layer, NVIDIA serves as the "computing heart" of military AI—the only U.S. enterprise capable of providing military-grade ultra-high-density GPU clusters, AI acceleration chips, and classified computing architectures—capable of supporting large model training, real-time intelligence inference, all-domain data computation, and drone swarm target recognition within U.S. military IL-6 and IL-7 networks. It can be said that NVIDIA's computing ceiling determines the capability ceiling of the U.S. military's AI intelligence architecture.

At the cloud services platform layer, Microsoft, AWS, and Oracle—three of the world's top cloud service providers—are building dedicated, isolated military classified cloud networks for the U.S. military. AWS hosts the storage and scheduling of the U.S. military's global mass raw intelligence data. Microsoft focuses on military office applications, command-and-control systems, and AI model integration. Oracle specializes in the construction and operation of high-security, high-stability classified databases, ensuring that multidimensional battlefield data cannot be tampered with and can be efficiently accessed. If Microsoft and AWS, leveraging their respective cloud platforms, have built a "dual-cloud" architecture covering the U.S. military globally, then Oracle is embedded within it as an additional security valve.

At the communications and transmission layer, SpaceX's Starlink and Starshield systems are regarded by the U.S. military as the "nervous system" of the battlefield. Relying on a satellite communications architecture with no dead zones, low latency, and high throughput, they are responsible for real-time, all-domain battlefield enablement. Starlink ensures real-time connectivity for frontline individual soldiers, drones, ships, and aircraft, while Starshield provides the U.S. military with intelligence reconnaissance, satellite surveillance, and classified data transmission services, closing the loop between rear-area AI assessment and frontline response, and allowing AI intelligence capabilities to truly reach the battlefield.

Beyond the technology enterprises that signed agreements this time, Palantir and Anduril—as America's premier defense AI enterprises—have long been deeply bound to the U.S. military under long-term contracts. Palantir is the strategic hub and master systems integrator of the U.S. military's AI combat systems, responsible for coordinating and scheduling all technical resources. Anduril focuses deeply on edge AI, drone swarms, and counter-unmanned systems, serving as the frontline execution terminal of the U.S. military's AI combat systems.

System Leap: Five-Layer Closed Loop Accelerates Throughput

If the technology suppliers selected this time are assembled together with systems integrators and tactical execution units, a clear map of the intelligentized warfare (智能化作战) system architecture comes into view.

The first layer is the Department of Defense, the top-level decision-making core of the system, responsible for coordinating strategic planning, classified network access permissions, and budget approval. The second layer is the "strategic brain" represented by Palantir, responsible for receiving DoD requirements and completing intelligence integration and kill chain construction. The third layer consists of the component and technology suppliers represented by the enterprises that signed agreements this time, providing foundational technical support in computing power, algorithms, cloud services, and communications security. The fourth layer encompasses Anduril and traditional defense industry giants, responsible for tactical execution and hardware integration including unmanned operations and manned-unmanned teaming. The fifth layer consists of terminal units such as drones, satellites, ships, and individual soldier terminals, executing reconnaissance and strike missions while simultaneously transmitting battlefield data back to close the loop.

The disruptive nature of this architecture is reflected in three notable characteristics. First, "speed supremacy" (速度至上) has become the core logic. The U.S. side released the Accelerating AI Strategy with the aim of rapidly building an AI foundational support architecture and driving accelerated AI technology integration across all departments. Since its official AI platform GenAI.mil went online in December of last year, it has already covered approximately 1.7 million DoD users. This time, Silicon Valley enterprises went from negotiation to signed "enlistment" in less than three months—whereas classified network access approvals have typically required more than 18 months—demonstrating the intensity and resolve with which the U.S. military is driving civil-military integration (军民融合).

Second, "deep penetration" (深度突破) has become possible. Previous civil-military cooperation, whether cloud service procurement or algorithm development, remained at the level of "peripheral enablement" (外围赋能). Technology companies provided tools; the military used them within its own systems; and data flows always had clear isolation barriers. This time, the AI technologies of eight enterprises have been directly deployed into classified network environments, participating in the highest-level operational decision-making and weapons control.

Third, "eliminating single-point dependency" (去单一依赖) highlights wartime preparedness considerations. Whether it is the multi-source supply approach adopted at the algorithm and model layer or the dual-cloud parallel mode at the cloud services platform layer, the primary purpose is to avoid core AI interruption during wartime due to system failures, cyberattacks, or supply chain problems, thereby enhancing the resilience and sustained combat capability of the intelligentized warfare system.

All indications suggest that the intelligent systems of Silicon Valley's technology giants are no longer tools used by the military, but have become integral components directly embedded in the military's decision-making nervous system.

AI-Enhanced Capabilities: Leaps in Capability Raise Alarms

In fact, whether in the "Operation Absolute Resolve" (绝对决心行动) launched by the United States against Venezuela earlier this year, or in the illegal military strikes conducted by the United States and Israel against Iran, the U.S. military has made extensive use of artificial intelligence technology for target recognition and precision strike operations. This signals that the logic of warfare is accelerating its evolution from "human-led linear killing" toward "algorithm-led networked closed-loop killing" (算法主导的网状闭环杀伤), developments that have drawn international attention and alarm.

Iterating intelligence capabilities supports the formation of intelligence-based battlefield shaping power (情报战场塑造力). Leveraging multimodal large model capabilities, the U.S. military's AI systems can automatically parse satellite imagery, monitor communications signals, and decode text information, enabling automatic target identification, correlation analysis, behavioral prediction, and trend extrapolation—precisely screening high-value intelligence from massive datasets and even anticipating the decision-making tendencies of opposing commanders and the operational intentions of their forces. This supports the U.S. military's intelligence capabilities in upgrading from traditional "reconnaissance and search advantage" to "cognitive advantage" (认知优势), forming battlefield shaping power. In the U.S.-Israeli strikes against Iran, the U.S. military employed multiple AI platforms to achieve full-process intelligent control from intelligence assessment, target identification, and situational decision-making to precision strike, effectively supporting continuous target acquisition.

Cloud-based algorithm enablement supports distributed unmanned autonomous operations. Anduril's tactical integration system, leveraging Reflection's lightweight open-source model and SpaceX's Starshield encrypted satellite network, bridges the data interfaces of unmanned equipment of different types and different services, potentially enabling autonomous grouping and coordinated division of labor among unmanned ground vehicles, reconnaissance-strike integrated drones, and unmanned underwater vehicles. A single commander can simultaneously command and control large numbers of unmanned platforms, forming a swarm-based systematic destruction capability (集群体系化毁伤能力).

Shortening the OODA loop creates "asymmetric" operational advantages. In the U.S.-Israeli-Iranian conflict, AI-enabled intelligence analysis efficiency and operational planning efficiency were confirmed to have improved substantially. Now that the U.S. military has fully integrated civilian AI technology, multimodal large model capabilities are continuously compressing decision cycles, pushing intelligence distillation toward the second-level timeframe. The leap in computing power and the dual-cloud architecture may completely break down data barriers across U.S. global garrisons, overcoming edge computing bottlenecks, while low-latency space-based communications and high-security databases provide the frontline fire network with anti-jamming links and stable data support. The U.S. military believes that only by binding itself to the iterative technological advantages of civilian AI can it seize greater "asymmetric" operational advantages in future conflicts and firmly maintain global military hegemony.

Layout design: Hu Yunyan

Data compilation: Huang Tibiao, Li Qingxu

All images are archival photographs.

"Algorithms Chambered": Multiple Strategic Risks Lurk Beneath the Surface

■ Wu Yang

On the morning of June 12, the ticker "SPCX" flashed on the NASDAQ electronic display. Leveraging its technological barriers, supply chain advantages, and military satellite network and intelligence contracts with the Department of Defense, SpaceX became the largest IPO in U.S. history on its first day of trading. This phenomenon is far more than an ordinary commercial event; it is a microcosm of how, in the age of artificial intelligence, the U.S. defense industrial system is accelerating its shift from the traditional military-industrial complex model toward deep integration with Silicon Valley's innovative forces.

For a long time, the Department of Defense has played the role of Silicon Valley's "angel investor." Technologies that changed the world—computers, the internet, GPS, drones—almost all originated from military requirements, underwritten by defense contracts, before spilling over into civilian markets once mature. After the outbreak of the Russia-Ukraine conflict in 2022, cloud service giants such as Microsoft and AWS quickly intervened, migrating the Pentagon's data architecture to the cloud, while Starlink has long served as the communications hub for Ukrainian drone strikes. The fact that these technology companies' commercial technologies have participated in and influenced the course of the conflict to varying degrees has become an undeniable reality for all parties. In May of this year, the Department of Defense announced cooperation agreements with leading Silicon Valley technology enterprises and stated plainly its intent to "accelerate the pace of the U.S. military's transformation into an AI-first combat force and enhance warfighters' ability to maintain decision advantage across all operational domains."

Unlike the production model of traditional defense industry giants, America's Silicon Valley "new arms merchants" are not contractors for individual components but definers of system standards and builders of system architecture and software-hardware integration. The Defense Innovation Board, led by former Google CEO Schmidt, has elevated Silicon Valley from a "subcontractor" to a "co-designer" of top-level architecture. Some analysts argue that the U.S. side's accelerated push for military AI applications—allowing technology giants to bypass layers of procurement processes and cut directly into the most core battlefield functions—conceals multiple strategic risks beneath this seemingly efficient and impressive integration model.

"Capital profit-seeking" (资本逐利) continuously erodes the boundaries of military operations. Silicon Valley enterprises are driven by scale expansion, commercial valuation, and market share, oriented toward profit maximization; the foundational principles of a combat system are full-process controllability and stable reliability, with national defense security as the supreme criterion. The two operate on inherently opposing logics with fundamentally different value orientations. As U.S. military all-domain operational processes become deeply embedded in Silicon Valley's technical architecture, profit-oriented commercial entities are taking on core defense functions, and the business considerations of commercial capital are quietly permeating critical nodes including strike decision-making, force employment, and operational assessment, directly affecting how the U.S. military conducts operations. According to disclosed information, during this year's U.S. military strikes against Iran, SpaceX executives proposed raising the Starlink connection fee for the "Lucas" drone from the $5,000 previously proposed by the military to $25,000, and the Pentagon—then intensifying its air campaign—had no choice but to agree. As the U.S. military's binding with Silicon Valley deepens, U.S. military operations may be continuously constrained at the technical level.

"Algorithms chambered" (算法上膛) conceals risks of loss of control. In order to seize the opportunity of the military's intelligentized transformation and consolidate their share of defense cooperation, Silicon Valley enterprises tend to prioritize deployment and iteration efficiency, simplifying adaptation testing for military scenarios. Commercial AI technologies that have not undergone high-intensity adversarial testing are highly susceptible, when facing adversarial means such as data poisoning, model deception, and electromagnetic interference in wartime, to evolving into fatal problems such as target misidentification and fire distribution errors. At the same time, massive volumes of battlefield data and operational instructions are output at high speed through algorithms, making it difficult for human commanders to individually verify and validate AI computational results. The "human-in-the-loop" (人在回路) control mechanism is continuously weakening, human decision-making authority is passively ceded to machine algorithms, and the risk of intelligentized warfare losing control and order is substantially elevated.

"Speed first" (速度优先) may trigger a new round of AI arms races. The U.S. military has placed artificial intelligence at the forefront of military competition, proactively breaking down the security barriers of its combat systems and opening high-level classified networks to absorb commercial technology. The core objective is to leverage top AI technology resources to help its own side seize battlefield initiative and form "asymmetric" operational advantages. This development approach—trading security compromises for speed of iteration—is saturated with typical arms race thinking and is highly susceptible to imitation. At the same time, the U.S. military is promoting the diffusion of military AI intelligence capabilities to allies through multilateral mechanisms such as AUKUS and NATO, further exacerbating the risks of AI-domain arms competition and geopolitical security rivalry. As the United States accelerates its "AI-first" positioning, the world's major military powers will inevitably follow suit rapidly, the intensity of the arms race will continue to escalate, and the uncertainty of the international security situation may continue to rise.

Observing Silicon Valley's Path to "Enlistment" Through Shield AI

■ Li Lu, Feng Zhiqi

In June of this year, U.S. defense technology enterprise Shield AI announced it had received a U.S. Air Force Collaborative Combat Aircraft (CCA) program contract to provide the "Hivemind" swarm intelligence platform for the program, enabling it to autonomously execute missions and operate in contested and denied environments. From a small company launched with $100,000 in seed funding to becoming an indispensable technology "unicorn" in the U.S. military's intelligentized combat chain, Shield AI's rise has become an important lens through which to observe the evolution of America's Silicon Valley innovation ecosystem.

Shield AI's founder, Brandon, was a U.S. Army Special Forces soldier who deployed to Afghanistan multiple times during his service and had direct experience with the limitations of traditional reconnaissance methods in urban close-quarters combat. When Brandon and his partners founded Shield AI in San Diego, California in 2015, they had a clear objective: to build an unmanned system that could autonomously sense, plan, and execute missions under conditions of complete communications blackout—without relying on GPS, pre-loaded maps, or remote control. This objective closely aligned with the U.S. military's real-world requirements on the Middle Eastern battlefield, providing the enterprise with an important opportunity to enter the military's field of view.

Because traditional defense industry giants had long monopolized equipment supply, and given the uncertainty and long cycle times of defense contracts, Silicon Valley enterprises were once not particularly enthusiastic about defense orders. Over the past decade or more, the Department of Defense has adopted a series of reform measures to expand cooperation with Silicon Valley, including establishing dedicated institutions such as the Defense Innovation Unit (DIU), forming the Defense Innovation Board, and continuously streamlining defense procurement processes—deepening cooperation with technology enterprises and facilitating the participation of non-traditional defense contractors in defense innovation. In 2016, Shield AI—less than a year old—obtained $1 million in research and development funding through DIU and entered the U.S. military procurement system. In 2018, its first product, the Nova tactical unmanned aerial vehicle, was deployed to the Middle East for reconnaissance support missions in special operations, breaking the long-standing monopoly of traditional defense industry giants over tactical unmanned equipment. Today, Shield AI is a leader in airborne unmanned systems and an important partner in the U.S. Air Force's development of manned-unmanned teaming (有人与无人协同作战) technology.

For Silicon Valley companies, survival anxiety is also an important driver accelerating their cooperation with the Pentagon. Since the outbreak of the Russia-Ukraine conflict, unmanned platforms, autonomous systems, and related technologies have been accelerating their transition from laboratory to battlefield. Against this backdrop, the U.S. military is promoting the application of emerging technologies, underwritten by defense contracts, attracting a surge of capital. At the same time, cooperation with the military provides access to unique "non-public data" and extreme use-case scenarios—irreplaceable scarce resources for Shield AI's entrepreneurs in terms of technology iteration. Faced with this stable and enormous "cake," Shield AI proactively aligned itself with military standards to obtain qualifications for classified projects; brought in multiple senior defense executives to build a procurement and project management system adapted to military rules; and implemented a "software-defined hardware" model, with the "Hivemind" intelligent platform as its core, compatible with multiple types of crewed and unmanned equipment, reducing equipment integration costs and iteration cycles. By virtue of its advantages in standards compliance, rapid iteration, and lightweight flexibility, Shield AI has consistently maintained competitiveness in military procurement.

It can be said that Shield AI's rise is a mirror of the deepening cooperation between the U.S. defense establishment and Silicon Valley's innovative forces, reflecting the U.S. military's strategic intent to accelerate the generation of intelligentized combat capabilities by nurturing technology enterprises. What merits serious reflection is this: when Silicon Valley elites who once aspired to "change the world" willingly make weapons for the Pentagon, this combination will bring enormous and unpredictable security risks to the world. Especially in the frenzy of chasing speed, disregarding the constraints and oversight of AI military applications makes it extremely easy to breach the ethical boundaries of warfare and the baseline of international humanitarian law.

Knowledge Corner

The "Revolving Door" Mechanism

In June of this year, the U.S. Army Reserve's 201st Detachment completed the commissioning of a second cohort of Silicon Valley executives, with executives from Cloudflare, FAIR, and several other companies receiving reserve lieutenant colonel commissions. At the same time, in recent years multiple retired senior DoD acquisition and intelligence officials have flowed into Silicon Valley enterprises such as Palantir and Anduril, leveraging their connections to lobby the military and win large AI contracts, forming a two-way personnel flow of corporate executives entering military service and senior military officials entering enterprises. Although the U.S. military has introduced conflict-of-interest provisions, it lacks a routine third-party audit mechanism. Executives simultaneously hold dual roles as corporate decision-makers and military technical advisors, and the chain of interests through which capital intervenes in AI armament development has already taken shape.

The "Compliance" Test

In 2018, a defense program involving Google—which used AI to analyze drone imagery to improve U.S. military strike accuracy—was leaked, immediately triggering a joint protest signed by more than 3,100 Google employees. Under pressure, Google announced it would not renew the defense program and pledged not to apply AI to weapons development. In March of this year, Anthropic was excluded from the supply chain by the Department of Defense after refusing to apply its AI technology to autonomous lethal weapons. In May, a group of Silicon Valley giants reached cooperation agreements with the Department of Defense, fully aligning with the latter. The trajectory of the entire industry sends a clear signal: only enterprises that pass the "compliance" test (服从性测试) can obtain DoD contracts. The forcefulness and urgency with which the U.S. military pursues AI military advantage is plain to see.

"Bubble-Style" Financing

AI defense concept stocks tend to surge sharply when conflicts break out, forming a cycle of "war narratives driving up stock prices—higher stock prices enhancing financing capacity—financing reinvested in R&D." SpaceX, with $18.6 billion in revenue, at one point carried a market capitalization of nearly 2 trillion dollars; Palantir's price-to-earnings ratio has long remained elevated; Anduril continues to double its financing despite annual losses of over a billion dollars—these figures have long since departed from traditional profitability logic. Following SpaceX, deeply loss-making OpenAI and Anthropic are also making pre-IPO sprints. However, once listed, when major shareholders' lock-up periods expire and performance fails to materialize, the pressure of valuation normalization will be transmitted through index funds to ordinary people and may even cause systemic shocks to global financial markets.