We are trying to govern a ghost. Washington and Beijing keep talking about putting guardrails on artificial intelligence, but they are using an outdated playbook. They want something like Cold War non-proliferation treaties. It won't work. Nuclear weapons require massive, highly visible enrichment facilities that satellites can spot from orbit. High-end AI models require code and silicon. You can hide a trillion-parameter model on a stack of server racks inside an unmarked warehouse.
If we want serious multilateral AI arms control, we have to stop focusing on the software. We need to track the physical hardware instead. Don't forget to check out our earlier coverage on this related article.
Diplomats spend hours debating international frameworks at global summits. They talk about shared ethics, red lines, and algorithmic safety. This is mostly theater. The real battleground isn't diplomatic prose. It is the global semiconductor supply chain. If global leaders refuse to adapt their inspection strategies to the realities of advanced computing, any treaty they sign will be completely useless.
The illusion of the nuclear analogy
People love comparing AI to the atomic bomb. It feels big, dangerous, and existential. It also leads to terrible policy. To read more about the background of this, MIT Technology Review offers an informative summary.
Traditional arms control relies on physical verification. The International Atomic Energy Agency sends inspectors to look at centrifuges. They count warheads. They check seals on nuclear material. This works because weapons-grade uranium doesn't just appear out of nowhere. It requires an industrial footprint that cannot be hidden easily.
Software is completely different. A cutting-edge neural network is just a file filled with billions of numbers. You can copy it onto a thumb drive. You can transmit it across borders in seconds over an encrypted connection. Once a model is trained, running it requires a fraction of the power it took to build it.
This creates a massive verification trap. If a country promises not to build an autonomous cyberweapon, how do you verify that? You cannot send inspectors to look at every hard drive in Virginia or Shenzhen. No sovereign nation would allow foreign intelligence agencies to audit their proprietary codebases. The level of state intrusion required to verify a software treaty makes the treaty dead on arrival.
Chokepoints and the physics of compute
We cannot track code. We can track chips.
Training frontier intelligence requires an enormous amount of physical infrastructure. It takes tens of thousands of specialized accelerators, massive data centers, and gigawatts of electricity. This infrastructure cannot be easily hidden. This is where multilateral AI arms control must start.
The global hardware pipeline is highly centralized. Only one company in the world, ASML in the Netherlands, makes the extreme ultraviolet lithography machines needed to print the most advanced chips. Only a handful of foundries, primarily TSMC in Taiwan, can manufacture them at scale. This extreme concentration gives global regulators a unique lever.
Instead of policing what engineers write on their keyboards, international agreements must police the factories.
True multilateral governance means turning these industrial hubs into tracking centers. Every advanced chip could be embedded with a secure cryptographic identifier at the factory level. Think of it as a digital license plate. This hardware-level tracking would allow an international registry to monitor where clusters of high-end chips are being assembled. If an unauthorized data center suddenly starts drawing massive amounts of power with tens of thousands of untracked chips, it flags an immediate treaty violation.
Why the current diplomatic approach is broken
Right now, international discussions are split into fragmented camps. You have the UN AI Advisory Body trying to build a global consensus. You have bilateral talks between the US and China. Then you have regional initiatives like the EU AI Act.
Most of these efforts are missing the point. They focus heavily on regulating the outputs of the technology. They want to ban deepfakes, stop algorithmic bias, and prevent discriminatory hiring practices. Those are real societal problems, but they are not arms control.
True arms control is about preventing catastrophic misuse by state actors. We are talking about automated bioweapon design, autonomous drone swarms, and self-replicating cyber warfare tools. You cannot prevent these threats by asking developers to fill out compliance paperwork before they release an app.
The current approach also ignores the competitive incentives driving the US and China. Both nations see this technology as a zero-sum race for geopolitical dominance. Washington wants to maintain its lead by cutting off Beijing's access to advanced silicon. China is pouring billions into domestic semiconductor manufacturing to bypass those restrictions. In this environment, vague agreements about safety principles are just a stalling tactic. Neither side will voluntarily slow down its research based on a pinky promise.
The technical reality of verification
If we want a treaty that actually binds nations, we need to talk about zero-knowledge proofs and secure enclaves. These are the tools that make compliance possible without compromising national security.
A major hurdle in tech diplomacy is intellectual property. If the US wants to prove to China that its newest model doesn't violate a ban on offensive cyber tools, it cannot simply hand over the weights of the model. Doing so would give away a multi-billion-dollar state asset.
This is where advanced cryptography comes in. Using zero-knowledge proofs, a state can mathematically prove that a piece of software adheres to specific constraints without revealing the underlying code. An international inspection team could run a verification program on an isolated server inside a secure enclave. The program would check the model's capabilities against a list of banned functions. It would output a simple yes or no.
This shifts the discussion from politics to math. It removes the need for blind trust. It allows states to maintain their secrets while still providing verifiable proof of compliance. This is not science fiction. These cryptographic tools exist today. They just need to be integrated into international law.
Building a realistic inspection framework
What does a working treaty actually look like? It doesn't look like a grand global ban on smart machines. It looks like a gritty, technical framework focused on three specific areas.
First, we need a global cap on compute clusters dedicated to single training runs. If nations agree that models above a certain computational threshold are too dangerous, they must agree to limit the size of the data centers tracking those models. This is equivalent to limiting the number of silos for intercontinental ballistic missiles.
Second, we need mandatory hardware-level logging. Tech firms must design advanced chips to report their operational status to a decentralized ledger. If a chip cluster is redirected toward a classified military network, the international registry would know instantly.
Third, we must establish joint research facilities. The US, China, Europe, and other tech powers need to fund shared labs focused purely on defensive alignment. If we can make defensive systems outpace offensive ones, the strategic value of deploying rogue tech drops significantly.
Your next steps for understanding the tech race
Do not get distracted by the high-profile political declarations coming out of global summits. If you want to track where global tech governance is actually going, you need to monitor the physical supply chain.
- Watch the export controls. Pay close attention to the specific technical thresholds set by the US Bureau of Industry and Security. These numbers define the boundary lines of global tech containment.
- Follow the hardware tracking research. Look into papers published by institutions like the Center for Security and Emerging Technology. They outline how physical chip tracking can be enforced without violating privacy.
- Monitor energy grids. Advanced machine learning requires unprecedented amounts of electricity. Tracking the construction of massive data centers and their proximity to nuclear or hydro power plants will show you exactly where the next frontier models are being built, regardless of what any public treaty says.
