AI and National Security: What Each Position Is Protecting

On October 7, 2022, the U.S. Bureau of Industry and Security issued a set of export control regulations that, in the words of semiconductor historian Chris Miller, author of Chip War, represented "the most sweeping export controls ever imposed by any country." The rules restricted sales of advanced AI chips — specifically Nvidia's A100 and H100 — to China. They blocked U.S. citizens from working at Chinese advanced semiconductor fabs. And they expanded the foreign direct product rule to cover chips made anywhere in the world using American software or equipment.

The strategic logic behind the rules had been stated explicitly one month earlier by National Security Advisor Jake Sullivan. In a September 2022 speech, Sullivan announced that the United States was abandoning its prior doctrine of maintaining a "relative advantage" over adversaries in key technologies. In force-multiplier technologies like AI, semiconductors, and quantum computing, Sullivan said, "we want to maintain as large a lead as possible." The phrase captured a real shift: not just slowing China down, but actively widening the gap.

Then, on January 20, 2025 — the day of Donald Trump's second inauguration — a Chinese AI startup called DeepSeek released a model called R1. By January 27, it had become the most downloaded free app in the United States, overtaking ChatGPT. That same day, Nvidia lost more than $600 billion in market capitalization in a single session — the largest single-day market cap loss in U.S. stock market history. DeepSeek had achieved performance comparable to OpenAI's flagship models, at a fraction of the cost, using the chips Nvidia had designed specifically to comply with the export controls: the downgraded H800, restricted from the market in October 2023, but stockpiled in the intervening year.

The same week, Pentagon logs showed Project Maven — the AI-powered targeting system now operated through a $1.3 billion Palantir contract — had helped compress kill chain decisions from hours to minutes in strikes across the Middle East. By March 2026, the Trump administration had designated Anthropic, whose CEO had refused to remove human-authorization requirements from AI targeting systems, a "supply chain risk to national security."

These events — the chip controls, DeepSeek, Maven, the Anthropic standoff — are not separate debates. They are the same debate, asked four different ways: What does it mean to maintain technological supremacy? Who bears the cost of pursuing it? What constraints, if any, should govern its military application? And who gets to decide any of this?

What security-first export control hawks are protecting

The recognition that in transformative general-purpose technologies, a small lead can become a decisive structural advantage — and that maintaining it requires active, aggressive policy, not just market competition. Jake Sullivan's "as large a lead as possible" formulation was a repudiation of a prior assumption: that the United States could maintain its technological edge through innovation alone, and that export controls only needed to prevent adversaries from leaping ahead, not from keeping pace. The 2022 controls operationalized a different theory: that AI chips are to the 2020s what uranium enrichment was to the 1940s — a chokepoint technology whose access determines not just who wins a market but who shapes the next century's military and political order.

The hawk case for export controls rests on a specific empirical claim: that without the controls, China's AI capabilities would be substantially more advanced. The TSMC "die bank" episode is the clearest evidence. Before the controls took effect, Huawei illegally stockpiled approximately 2.9 million advanced chip dies at TSMC. That stockpile funded Huawei's Ascend AI chip production through 2024 and 2025. DeepSeek used H800 chips stockpiled before October 2023. In the hawk view, both of these facts prove the opposite of what critics claim: the controls are working well enough that China had to spend years stockpiling to circumvent them, and a China without those pre-ban stockpiles would be meaningfully further behind.

The technology ceiling argument is the deeper version of this case. SMIC, China's most advanced domestic chip foundry, is currently producing chips at approximately 7nm-equivalent density. The global frontier, currently held by TSMC, is 2nm. SMIC cannot reach sub-7nm without ASML's extreme ultraviolet lithography machines, which are export-controlled (the Netherlands, under U.S. pressure, restricted ASML exports to China in 2023). Even a 5nm gap in process technology translates to roughly two to three generations of performance. The ceiling is real. The question is whether it matters enough to justify the cost of maintaining it.

Dario Amodei of Anthropic made the hawk case explicitly in a Wall Street Journal op-ed in January 2025: the controls should be tightened, not loosened, and the argument that they are backfiring by spurring Chinese innovation "gets the causation backwards." Chinese stockpiling before controls took effect was predictable and should have been anticipated; the solution is faster and more comprehensive enforcement, not retreat. The Semiconductor Industry Association's warning that controls could harm U.S. firms should be taken seriously, but not treated as dispositive — in 1945, the argument that nuclear controls would harm American physics research would have been true and irrelevant.

What economic and industry critics of controls are protecting

The market logic that made American semiconductor leadership possible in the first place: global revenue funds the R&D that generates the lead, and removing the revenue base doesn't slow China down — it accelerates China's domestic build-out while weakening the American firms that the controls are supposed to protect. Jensen Huang, Nvidia's CEO, estimated China as a "$50 billion annual opportunity" for Nvidia in calendar year 2025. That figure represents lost research capacity. A New York Federal Reserve study found that the October 2022 announcement alone caused a persistent 2.5 percent stock valuation decline for affected U.S. firms, representing approximately $130 billion in aggregate market capitalization. The Information Technology and Innovation Foundation estimated that full decoupling would cut U.S. semiconductor research and development spending by roughly 24 percent — about $14 billion per year — and eliminate 80,000 or more direct jobs, with five to six downstream jobs lost for each semiconductor position.

The industry critique has a harder version: the controls may be producing exactly the outcome they were designed to prevent. Before October 2022, Huawei was a chip customer — dependent on TSMC and other non-Chinese suppliers. After the controls, Huawei was forced to become a chip manufacturer. By 2024, it had become the fourth largest wafer fabrication equipment buyer in the world, with $7.3 billion in capital expenditure, up 27 percent year over year. A company that was purchasing chips from the West is now building the infrastructure to produce them domestically. The Semiconductor Industry Association's position — that controls should be "narrowly targeted" and developed "with sufficient industry expertise" — is not a lobbying position, it is an operational observation: controls designed without understanding the actual supply chain create loopholes, generate stockpiling incentives, and drive the adversary toward self-sufficiency faster than arms-length competition would.

DeepSeek crystallized the efficiency argument. MIT Technology Review's investigation found that export controls forced DeepSeek's engineers to optimize radically for compute efficiency — not because they wanted to, but because they had to. The H800 chip they used has roughly one-third the bandwidth of the H100 for the operations that matter most in AI training. The team developed novel memory management techniques, a mixture-of-experts architecture that activates only a fraction of its parameters at once, and training protocols that squeezed far more performance from less hardware than U.S. labs considered worth pursuing. In the Brookings Institution's analysis, "DeepSeek shows the limits of U.S. export controls on AI chips" — not because the controls failed to restrict hardware, but because hardware restriction drove algorithmic innovation that hardware access had made unnecessary. The controls made China better at AI under resource constraints, and resource constraints may turn out to be the permanent condition of AI deployment, not its frontier.

What international and multilateral governance advocates are protecting

The possibility that the risks posed by AI — autonomous weapons capable of acting without human authorization, surveillance systems that can identify and track dissidents at scale, disinformation systems that can destabilize democratic elections — require international frameworks that no unilateral technology lead, however large, can address on its own. The multilateralist position is not primarily that export controls are wrong. It is that export controls are the wrong tool for the most important problems, and that pursuing technological supremacy as the organizing frame of AI strategy crowds out the governance work that actually needs to happen.

The evidence for this concern is in what the international forums have and haven't produced. The November 2023 Bletchley Park AI Safety Summit — 28 countries plus the European Union — issued the Bletchley Declaration and launched the International Network of AI Safety Institutes (eventually including the U.S., UK, EU, Japan, Singapore, South Korea, Canada, France, Kenya, and Australia). This was genuine diplomatic infrastructure for AI governance. The February 2025 AI Action Summit in Paris, which France and India co-chaired with approximately 100 countries and 1,000 stakeholders in attendance, was notably renamed from "AI Safety" to "AI Action" — and its final declaration included no substantial commitments to AI safety. The Paris summit marked the moment when the international frame shifted from managing AI's risks to competing for AI's benefits.

The multilateralist critique identifies a specific mechanism by which technology competition undermines governance. The U.S. AI Action Plan (July 2025) explicitly pursues "plurilateral" controls — coordinating with Japan, the Netherlands, and South Korea outside multilateral treaty bodies, rather than working through the Wassenaar Arrangement or seeking UN-level agreements. This approach is deliberately chosen: multilateral bodies include China, which can block or water down agreements. But the cost is that China and Russia, excluded from the design of these rules, have no incentive to comply and every incentive to position themselves as the alternative for the Global South. In November 2025, 156 nations supported a UN resolution calling for a legally binding treaty on autonomous weapons systems. The United States was not among them. Russia voted no, alongside North Korea and Belarus — and is reportedly conducting approximately 300 AI-assisted strikes per day in Ukraine. The multilateral infrastructure that might eventually constrain that is being built without U.S. leadership, and may be weaker for it.

Georgetown's Center for Security and Emerging Technology has documented that U.S. allies face a genuine legal problem: most of them lack the domestic statutory authority to implement AI and semiconductor export controls comparable to U.S. BIS rules. The plurilateral approach requires each ally to build its own legal infrastructure — a slow, politically contentious process that creates enforcement gaps and inconsistency. RAND's analysis of the AI Diffusion Rule framework (now rescinded) found that the rule's most ambitious goal — using chip access as leverage to bind Tier 2 countries into U.S. data center governance standards — was also its most fragile, because it depended on countries voluntarily accepting U.S. oversight conditions that reduce their digital sovereignty. China's counter-narrative — that it offers access without conditions — is structurally appealing to governments skeptical of American tech hegemony.

What democratic accountability and anti-militarism critics are protecting

The principle that when governments use AI to make targeting decisions that result in the deaths of specific human beings, democratic accountability — not just technical oversight — is what "meaningful human control" actually requires: known standards, public debate, congressional authorization, and the ability to hold specific people responsible when something goes wrong. The accountability critique is not primarily about whether AI targeting systems work. It is about who knows they exist, who authorized them, what standards they operate under, and who is liable when they fail.

Project Maven — launched in April 2017, now operated under a Palantir contract that grew from $480 million in May 2024 to $1.3 billion by May 2025, with a separate $10 billion Army framework announced in July 2025 — is described by the Pentagon as "human on the loop": humans authorize each engagement, but AI recommends targets, ranks them, and fuses intelligence from more than 150 data sources simultaneously. The Brennan Center for Justice, in its 2025 report on the business of military AI, found that "even the most basic information about the types of systems the Pentagon is adopting is often hidden from Congress and the public." The FY2025 National Defense Authorization Act, signed December 2024, requires an annual congressional report listing all autonomous weapons systems approved and deployed under DoD Directive 3000.09 — the first time Congress has mandated such disclosure, after seven years of Maven's operation without it.

The Anthropic standoff of early 2026 crystallized the accountability argument in a way that abstract policy debates cannot. Anthropic signed a $200 million DoD contract in July 2025 with two explicit contractual conditions: Claude models would not support fully autonomous lethal targeting without human authorization, and they would not support domestic surveillance of U.S. citizens. These were not policy preferences — they were hard red lines Anthropic refused to remove. When the Pentagon demanded their removal and Anthropic refused, the Trump administration designated Anthropic a "supply chain risk to national security" in March 2026, ordering federal agencies to phase out Claude within six months. The implications are significant: a company drawing a line against autonomous weapons targeting was treated not as acting on legitimate ethical grounds but as a threat to the national security apparatus itself.

DoD Directive 3000.09, updated in January 2023, is widely mischaracterized as a ban on lethal autonomous weapons systems. It is not. It requires "appropriate levels of human judgment over the use of force" and senior-level review before deploying autonomous weapons — but explicitly does not define what "appropriate" means, and Pentagon officials have stated that the U.S. "may be compelled to develop [LAWS] if U.S. competitors choose to do so." This is not a guarantee against autonomy; it is a procedural framework that defers the core ethical question to a race-to-the-bottom dynamic. If Russia or China deploy fully autonomous weapons, the directive provides the rationale for the U.S. to follow. The accountability critics argue that this is precisely the wrong governance structure for a technology whose errors, in combat, cannot be undone.