Diplomacy often wears the costume of history. When American and Chinese security researchers sit across from each other in quiet conference rooms to debate artificial intelligence treaties, they reach instinctively for the most familiar playbook on the shelf. That playbook belongs to the mid-twentieth century. It smells of uranium, heavy lead containers, and the International Atomic Energy Agency.
The proposal sounds neat on paper. Establish international inspection teams, cap computational power the way treaties once capped warheads, and create verification protocols to stop rogue models from going rogue. Headlines praise the pragmatism. Analysts talk about bipartisan consensus and cross-border cooperation in the shadow of existential risk.
It is a comforting performance. It is also entirely divorced from technological reality.
Software is not enriched uranium. You cannot smell a data center from orbit with a satellite sensor, nor can you lock a cluster of high-performance graphics processors in a concrete vault with a tamper-proof seal without completely halting the engine of modern commerce. Treating the current machine learning race through the lens of Cold War arms control fundamentally misunderstands how code spreads, how hardware is manufactured, and why sovereign states view computational supremacy as a matter of national survival.
The Myth of Physical Containment
Uranium has mass. It requires physical mines, vast enrichment facilities filled with centrifuges that consume megawatts of power and emit distinct isotopic signatures, and physical shipping routes that intelligence agencies can monitor. You can track a pound of yellowcake from the ground to a warhead casing.
Algorithms possess no mass. They are matrices of floating-point numbers distributed across arrays of silicon. A model with billions of parameters can be compressed, quantized, and transmitted across a fiber-optic cable in seconds. It can be copied onto an encrypted hard drive that fits inside a coat pocket.
Proposals for international inspection regimes assume that computing clusters resemble nuclear reactors. They picture massive, centralized installations that require armies of technicians and constant external oversight. While frontier training runs do require concentrated clusters of specialized chips, inference workloads and smaller, highly capable open-weight models do not.
Imagine an inspection regime granted access to a major data center in Virginia or Guangdong. Inspectors verify that the power draw matches expectations, that the cooling systems are operating within declared parameters, and that no unauthorized training runs are underway on those specific nodes. Meanwhile, a research group in an undisclosed basement splits a smaller training run across thousands of consumer-grade graphics cards rented through decentralized cloud providers. The physical verification model collapses the moment the compute requirement drops below the threshold of suspicion.
History shows that arms control treaties work best when verification is absolute and non-compliance carries an immediate, undeniable physical footprint. The nuclear test ban treaties relied on seismic monitoring networks that could detect underground detonations anywhere on Earth. There is no seismic sensor for a rogue gradient descent calculation.
The Semiconductor Chokepoint Illusion
Proponents of nuclear-style oversight frequently point to hardware restrictions as the anchor of any future treaty. By controlling the supply chain of extreme ultraviolet lithography machines and advanced semiconductor manufacturing, governments can theoretically throttle the capability of any adversary.
This argument confuses a temporary supply chain bottleneck with permanent structural control.
Semiconductor manufacturing is currently concentrated in a handful of high-precision foundries located in Taiwan, South Korea, and the United States. This concentration has tempted policymakers into believing that hardware is the ultimate control point. If you control the silicon, you control the intelligence.
Markets do not tolerate artificial scarcity for long. When exports of high-end accelerator chips are restricted, affected nations pour billions into domestic fabrication research, alternative architectures, and analog computing concepts. Furthermore, efficiency gains in software often outpace hardware limitations. Methods that allow models to train effectively with lower-precision data types mean that older or less sophisticated hardware can achieve outputs previously thought to require frontier chips.
Relying on hardware chokepoints to enforce a treaty is akin to trying to stop the spread of firearms by controlling the supply of specialized steel mills. Eventually, innovators find ways to work with cheaper materials, simpler designs, and distributed networks. The chokepoint narrows, pressure builds, and workaround ecosystems emerge in the shadows where inspectors cannot reach.
The Prisoner Dilemma of Sovereign Paranoia
Even if technical verification were possible, the underlying political architecture required for a nuclear-style treaty does not exist.
The United States and China are not partners managing a mutual existential hazard. They are strategic competitors locked in a zero-sum race for economic and military dominance. In a nuclear standoff, both sides share a common interest in mutual assured destruction remaining mutual. Neither superpower wants a stray warhead to trigger a global exchange because both know they will burn in the fire.
Artificial intelligence operates under a different calculus. The first nation to achieve decisive superiority in automated cyber warfare, autonomous logistics, and algorithmic intelligence gains an asymmetric advantage that alters the global balance of power without necessarily triggering a cataclysmic exchange.
If Washington signs a treaty capping its computational growth, Beijing has every incentive to cheat in secret, knowing that verification protocols will be evaded more easily in a closed society than an open one. If Beijing signs the same treaty, its military planners will calculate the risk of American defection and conclude that compliance is an invitation to strategic subjugation.
Trust requires transparency. Transparency requires verifiable verification mechanisms. Verifiable verification mechanisms require a level of mutual trust that neither government possesses toward the other. The diplomatic effort becomes a circular trap where every proposed inspection protocol is viewed by the opposing side as an espionage vector designed to steal proprietary trade secrets or map critical infrastructure vulnerabilities.
The Danger of Regulatory Capture
There is another, quieter reason why large technology firms often welcome discussions of nuclear-style safety frameworks.
Regulation creates moats.
When industry giants participate in summits with government officials to draft stringent compliance rules, mandatory auditing standards, and multi-million-dollar safety testing protocols, they are not acting solely as public-spirited citizens. They are building a tollbooth.
Compliance is expensive. A startup with three engineers and an open-source ethos cannot afford a dedicated compliance department, third-party algorithmic safety audits, or government-certified containment facilities. By framing AI risks in terms that evoke nuclear disasters, established corporations encourage lawmakers to codify regulations that only well-capitalized incumbents can satisfy.
The result is not safety. The result is market consolidation disguised as risk mitigation. Smaller research groups are priced out of the frontier race, driving open-source development underground or into jurisdictions with laxer oversight, while a handful of corporate behemoths cement their status as government-sanctioned gatekeepers.
The rhetoric of existential threat serves the financial interests of those who wish to pull up the drawbridge behind them.
What Real Oversight Requires
If international treaties modeled on the Cold War are bound to fail, what remains?
Effective governance must abandon the fantasy of centralized prohibition and embrace the messy reality of decentralized resilience. Instead of trying to lock up the code or halt the hardware, policymakers should focus on defensive engineering, liability frameworks, and operational transparency in critical deployment sectors.
We do not regulate civil aviation by banning the invention of new wing shapes or limiting the number of wind tunnels a nation can build. We regulate aviation by enforcing rigorous maintenance standards, training requirements, air traffic control protocols, and strict liability when an aircraft fails.
In the digital domain, this means shifting attention from preemptive compute caps to post-deployment accountability. If a model is deployed in a critical infrastructure network or a financial system and causes catastrophic harm, the deploying organization must face severe, unambiguous legal and financial consequences.
It means investing heavily in defensive cybersecurity, cryptographic watermarking, and model interpretability tools that allow operators to audit decision pathways after the fact, rather than pretending inspectors can read the minds of neural networks during training runs.
It means accepting that code will proliferate, algorithms will improve, and the genie cannot be put back into the bottle through diplomatic communiques issued by delegations posing for photographs in Geneva.
The desire for a neat, centralized fix is understandable. The alternative requires living with chronic uncertainty, managing vulnerabilities through constant vigilance, and abandoning the comforting illusion that a treaty can repeal a mathematical reality.
The next intelligence explosion will not be contained by a treaty signed by men in suits who believe software behaves like enriched uranium. The sooner policymakers stop pretending otherwise, the sooner real work can begin.
