Four Battlegrounds

Paul Scharre’s Four Battlegrounds arrived in early 2023, five years after his earlier book Army of None had become a standard reference on autonomous weapons, and a few months after ChatGPT had pulled the broader conversation about artificial intelligence into the centre of public life. Scharre wrote from his perch at the Center for a New American Security in Washington, where he is the director of studies and where he had spent the previous decade translating Pentagon debates about emerging technology for a wider policy audience. Before that he had been an Army Ranger with deployments to Iraq and Afghanistan and a civilian analyst in the Office of the Secretary of Defense, working on the directives that still govern how the Pentagon thinks about autonomy in weapons. The book carries the marks of all three roles. It reads as a Washington briefing in long form, written by someone who has sat through both the operator’s after-action reviews and the policy office’s whiteboard sessions, and who has spent enough time in Beijing, Shenzhen, and the campuses of Stanford and Carnegie Mellon to treat the US–China contest in AI as the central organising story of the decade.

The thesis is straightforward. Artificial intelligence, Scharre argues, is becoming a general-purpose technology in the same family as electricity and the internal combustion engine, and like those earlier general-purpose technologies it will reshape economic and military power across the board rather than producing a single decisive weapon. The competition that matters is therefore not a race to build a particular system but a race to build the deeper capacity to keep building such systems for decades. That capacity rests on four foundations, and the book takes its title from them: data, compute, talent, and institutions. A country that leads on all four can absorb setbacks on any one and still pull ahead. A country that lags on any one will find the others harder to turn into military advantage. The United States, in Scharre’s reading, leads on talent and institutions while struggling to put data and compute to work inside its defence enterprise. China leads on data and on the political will to mobilise it, while still depending on American and allied tools at the bottom of the semiconductor stack. The contest will be settled by which side closes its gaps first.

The book is organised in five parts that walk the reader through that argument. The opening section sets the stage by tracing the moment when American defence officials first began to treat AI as a strategic problem rather than a research curiosity. Scharre returns to the work of the Defense Innovation Board under Eric Schmidt, the 2017 launch of the Algorithmic Warfare Cross-Functional Team, better known as Project Maven, and the establishment of the Joint Artificial Intelligence Center in 2018. He gives the same careful treatment to the parallel Chinese decisions: the State Council’s 2017 New Generation Artificial Intelligence Development Plan, Xi Jinping’s speeches on the “intelligentisation” of warfare, and the Military-Civil Fusion strategy that Beijing uses to direct the work of nominally private AI firms toward the People’s Liberation Army. From there the four battleground sections each take a chapter or two: data, then compute, then talent, then the institutional and doctrinal work of turning all three into military capability. A final section steps back to consider the implications for deterrence, escalation, and the prospect of arms control on systems that the public cannot see.

The substance of the book is in its cases. On data, Scharre devotes long passages to the surveillance architecture China has built in Xinjiang, drawing on reporting and on leaked documents to describe the integrated joint operations platform used by police in the region, the role of firms like SenseTime, Megvii, Yitu, and Hikvision in providing facial and gait recognition, and the way the system has been pitched as a counter-terrorism tool while functioning as a mechanism of mass internment of Uyghurs. The point of the chapter is not only humanitarian. Scharre treats Xinjiang as a working laboratory in which Chinese firms gathered the volumes and varieties of human data that Western competitors, bound by privacy law and democratic scrutiny, cannot easily replicate. He pairs this with the “Sharp Eyes” rural surveillance programme and the Skynet urban camera network, and with the data-sharing arrangements that flow between Chinese AI firms and the PLA’s Strategic Support Force.

On compute, the book reconstructs the semiconductor supply chain in patient detail. Scharre walks through the division of labour among the Taiwan Semiconductor Manufacturing Company, the Dutch lithography firm ASML, the American chip designers Nvidia, AMD, and Qualcomm, and the equipment makers Applied Materials, KLA, and Lam Research. He explains why extreme ultraviolet lithography, mastered only by ASML, has become the single physical chokepoint in the global AI economy, and why TSMC’s Hsinchu and Tainan fabs sit at the centre of Washington’s strategic anxiety about Taiwan. He covers the gradual tightening of American export controls from the 2019 Entity List actions against Huawei, through the 2020 foreign direct product rule, to the October 2022 controls that for the first time blocked the export of the most advanced AI accelerators and the tools to make them. The chapter does not pretend the controls will be airtight; it argues instead that they buy time, and that the value of that time depends on what the United States and its allies do with it.

On talent, Scharre traces the flow of researchers between American universities, American firms, and Chinese institutions. He uses the work of MacroPolo’s Global AI Talent Tracker to show that a substantial share of the top AI researchers working in the United States were born and trained in China, and that the United States benefits enormously from the long tradition of Chinese students staying after graduate school. He covers the Thousand Talents Plan and the way American counter-intelligence services have responded to it, the prosecutions of researchers like Charles Lieber at Harvard, and the chilling effect that broad investigations have had on Chinese-American scientists. The argument cuts against any temptation toward a wholesale decoupling on people: the United States, Scharre writes, wins the talent battleground in large part because it is open, and policies that close it off risk surrendering the one battleground on which the country’s lead is most secure.

The institutional chapters are where Scharre’s Pentagon experience shows most clearly. He recounts the Project Maven controversy in which several thousand Google employees signed a letter in 2018 protesting their company’s contribution of computer vision models to a Defense Department programme that helped analysts sort drone footage, leading Google to decline to renew the contract. He uses the episode to make a broader point about the gap between the culture of Silicon Valley and the culture of the Pentagon, and about how Microsoft, Amazon, Palantir, and Anduril came to fill the space that Google vacated. He covers the wandering history of the JEDI cloud contract, eventually cancelled and replaced by the Joint Warfighting Cloud Capability, and the more recent struggles of the Chief Digital and AI Office to absorb the JAIC and become the department’s institutional centre of gravity on AI. He devotes pages to DARPA’s AlphaDogfight Trials in 2020, in which an AI agent built by Heron Systems defeated an experienced F-16 pilot in simulated within-visual-range air-to-air combat, and to the follow-on Air Combat Evolution programme that put a similar agent on a modified L-39 trainer. He works through the Aircrew Labor In-Cockpit Automation System, the Replicator initiative for attritable autonomous platforms that the Deputy Secretary of Defense Kathleen Hicks would announce later that year, and the slow march of the Pentagon’s responsible AI policy, including the 2020 ethical principles and the updated Directive 3000.09 on autonomy in weapons.

He treats the Chinese institutional story with the same granularity. The PLA’s interest in what its theorists have called a “battlefield singularity” — a tempo of decision driven by machines beyond human ability to keep up — is set against concrete programmes: swarms of small drones tested at Chinese trade shows, autonomous surface vessels exercised in the South China Sea, the use of AI in early-warning radar processing, and the integration of commercial face-recognition firms into PLA contracts. Scharre is careful not to overstate the readiness of these systems. He notes the gap between Xi’s rhetoric on intelligentisation and the operational record of a force that has not fought a major war since 1979, and he points out that the PLA’s incentives to advertise capability ahead of reality are at least as strong as those of any American defence contractor.

The final section is the one most likely to dictate how the book is remembered. Scharre is sceptical of grand AI arms-control proposals, in part because the technology is too diffuse to verify and in part because the underlying models are dual-use down to the chip. He argues instead for narrower, behaviour-based measures: confidence-building agreements on nuclear command and control, prohibitions on specific classes of antipersonnel autonomous weapons, transparency about testing regimes. He returns repeatedly to the risk of accidents and unintended escalation between forces relying on opaque models, citing the Stanislav Petrov episode of 1983 as the kind of human intervention that brittle AI systems may not provide. He closes by warning against complacency in either direction: against the assumption that American technological lead is durable without sustained investment, and against the despair that says China’s data advantage has already settled the contest.

The reception has tracked the book’s centrist Washington stance. Foreign Affairs and Foreign Policy ran broadly favourable reviews. Henry Kissinger, before his death later in 2023, cited Scharre approvingly in his own work with Eric Schmidt and Craig Mundie. Inside the AI safety community, readers welcomed Scharre’s willingness to take catastrophic-risk arguments seriously without folding them into the policy chapters; outside it, defence analysts like Christian Brose and Elsa Kania engaged in friendly disagreement over how far the United States can rely on commercial dynamism rather than industrial policy. Some China watchers, including Jeffrey Ding and Matt Sheehan, pushed back on the portrait of Chinese data abundance, arguing that the country’s surveillance corpus is less useful for general-purpose AI than the four-battleground framing implies, and that Beijing’s compute deficit will bite harder than Washington’s data deficit. Critics on the disarmament side, including the campaign to Stop Killer Robots, found Scharre too willing to accept the Pentagon’s framing on lethal autonomy. The book has nonetheless become a standard syllabus item in graduate programmes on technology and statecraft from Georgetown to Tsinghua.

For a reader working through the literature on AI in war, Four Battlegrounds sits in a specific niche. It is the policy companion to Scharre’s own Army of None, which remains the better starting point on the narrower question of autonomous weapons. It pairs naturally with Christian Brose’s The Kill Chain and Chris Miller’s Chip War, the first for the institutional argument about how American defence acquisition lost the speed it had during the Cold War and the second for the deeper history of the semiconductor industry that Scharre touches on more briefly. It does not try to be a technical primer; readers wanting to understand transformer architectures or the mechanics of reinforcement learning will need other books. It also does not engage at length with the European, Indian, or Israeli ecosystems, treating them mostly as supporting characters in a bilateral story. Anyone reading widely on European defence will need to supplement it with separate sources on Project Maven’s transatlantic analogues and on the way Ukraine’s war with Russia has accelerated commercial drone integration into NATO armies.

Parts of the book have aged faster than others. The chapter on compute, written before the late-2023 and 2024 expansions of American export controls and before the surprise of DeepSeek’s models in 2025, reads as a snapshot of an earlier phase of the chip war. The discussion of large language models, anchored in GPT-3 and early GPT-4, predates the policy ferment around frontier models, the Bletchley and Seoul summits on AI safety, and the executive orders and EU AI Act that followed. What is likely to age well is the framing itself — the insistence that data, compute, talent, and institutions are the four things to watch — and the patient, case-by-case treatment of how those four interact inside real bureaucracies. Scharre’s deeper claim, that the contest will be won by the side that learns faster rather than the side that announces more, has so far survived the news cycle that has tried to overtake it.