The AI bubble's probability is debatable. Its severity isn't

Risk is probability times severity, and severity is where the argument actually lives

The standard way any board or insurer sizes a risk is by multiplying how likely it is by how bad it would be if it happened. Almost all public commentary on the AI investment bubble argues about the first term. Analysts debate whether valuations are twenty percent or sixty percent overextended, whether the correction is a year away or five, whether "bubble" is even the right word given how much real enterprise revenue now exists. That is a legitimate and unresolved argument, and this piece is not going to settle it.

What gets skipped is the second term. Severity is usually assumed to scale with the size of the financial write-down, the way it did in 2000 and 2008. That assumption is the mistake. The asset at the centre of this particular bubble is not capital in the conventional sense. It is a set of model weights, and weights do not behave like written-down equity once they leave a controlled environment.

The bubble is Western, and the Chinese comparison sharpens why

The circular financing structure sits inside a small number of US firms: Nvidia, Microsoft, OpenAI, Amazon, Google, and Anthropic, tied together through reported investment and supply relationships several analysts have described as circular. That is not a global phenomenon. It is a specifically Western capital structure, built on a specific bet, that scale and proprietary compute are a durable moat.

Chinese labs have spent the past eighteen months testing that bet directly. A RAND report found Chinese models running at roughly one sixth to one quarter the cost of comparable American systems, through architectural efficiency rather than raw scale, sparse routing, memory compression, lower-precision training. DeepSeek's widely cited training cost for a frontier-competitive model, around $5.6 million against tens or hundreds of millions for Western equivalents, is one data point in that pattern rather than the whole story, and it typically counts only the final training run rather than the full cost of getting there.

The naive conclusion, drawn briefly when this first became visible in early 2025, was that cheap Chinese models would collapse demand for the expensive infrastructure Western firms were building. That conclusion was wrong. Usage expanded as intelligence got cheaper, and Western compute spending hit records anyway. The more precise conclusion, and the one that actually matters here, is narrower: China has shown that proprietary scale is a contestable moat, not an unassailable one, at exactly the moment several Western firms are valued as though it were unassailable. That is a probability-side pressure specific to the West's bet, not evidence that the underlying technology is a bubble.

Why this may be harder to contain than nuclear material, not just as dangerous

The comparison to the end of the Cold War is not a claim that a Western AI firm's insolvency equals the collapse of a state. American and European courts, trustees, and regulators keep functioning through any single company's failure, which Soviet institutions did not. That difference is real and worth stating plainly.

But the comparison was never really about institutional collapse. It was about what happens when a dual-use, catastrophically capable asset stops being reliably held by the institution responsible for it. On that narrower point, the honest version of the argument is not that AI is as dangerous as nuclear material. It is that AI may be harder to contain once it starts moving, for a structural reason: nuclear material is physical.

IAEA safeguards work because fissile material can be measured, weighed, and tracked. Every gram is accountable, and a diversion shows up as a discrepancy someone can find. The Nunn-Lugar Cooperative Threat Reduction program, established through the 1991 Soviet Threat Reduction Act after Senators Sam Nunn and Richard Lugar built the case for it, could inventory Soviet warheads, transport them, and dismantle them, because there was a physical object at every step to inventory, transport, and dismantle. The program also funded something just as important as the material itself: finding gainful employment for thousands of former Soviet scientists with expert knowledge of weapons of mass destruction, specifically to keep them from taking that expertise elsewhere. The entire nonproliferation architecture built over the past seventy years rests on the premise that the dangerous thing has mass, occupies space, and can be counted.

Model weights have none of that. There is no equivalent of a Geiger counter for a frontier model running on a private cluster. A copy is not a diminished fraction of the original. It is the original, indistinguishable and fully capable, made in seconds at zero marginal cost, transmitted anywhere, reconstituted anywhere else. Custody, in the nuclear sense of controlling a finite, trackable, physical quantity, is not really available as a strategy here. What can be built, escrow of an authoritative copy, access controls, monitoring for anomalous transfers, addresses containment before the fact. None of it can do what an IAEA inspection does, verify after the fact that no unauthorized copy exists, because there is no physical inventory to reconcile against.

That is the actual severity claim, and it is harder to sit with than a simple analogy. The nuclear case gives a template for what worked: material and expertise both treated as objects requiring custody, and a coordination channel between rivals built specifically to manage that custody together. It cannot give a template for verification, because the thing it verified was physical and this is not.

The engagement gap is the severity multiplier

Nunn-Lugar did not emerge from nothing. It built on decades of US-Soviet arms control dialogue, backchannels, and verification regimes that existed because both sides had spent the Cold War talking to each other about exactly this category of risk, even while treating each other as adversaries everywhere else. When the crisis came, faster than anyone in Washington had planned for, there was already a relationship to move through.

There is no equivalent relationship between the West and China on AI capability risk specifically. On July 16, 2026, twenty-nine countries signed the founding agreement for the World AI Cooperation Organization, a body China first proposed a year earlier, headquartered in Shanghai. No major Western democracy is among the signatories. Western analysts have read the organization's structure as designed to build Chinese influence over AI standards across the Global South while excluding Western democracies by construction. Chinese officials frame it as a corrective to Western-led governance efforts that concentrate among a small circle of advanced economies and underweight the interests of the rest of the world. Both readings can be true at once.

What matters for severity is narrower than adjudicating that debate: whichever reading is right, the West's absence from this table, and from any equivalent one, means that if a disorderly correction at a major Western AI firm puts frontier capability at risk of uncontrolled dispersal, there is no existing channel through which the US, Europe, and China have ever discussed how to jointly handle that specific scenario. Not general AI competition, not chip export policy, but the narrow, mutually threatening question of what happens to loose frontier capability. That is an anticipation problem, and anticipation requires a relationship built before the event, not during it.

What exists, what's missing, and what to do about it

Current frameworks are real and shouldn't be waved away. The EU AI Act sets systemic-risk thresholds for general-purpose models above a compute threshold. Export controls already restrict advanced chips and, increasingly, some model-adjacent technical data. What none of these cover is custody of model weights through a company's insolvency specifically, which is the precise and narrower gap. A bankruptcy court handling a distressed AI company's assets today has no more obligation to consider where the weights end up than it would for an ordinary patent portfolio.

The remedy has two parts, and neither works without the other. The first is domestic and achievable unilaterally: classify critical frontier model weights as dual-use technology, and build custody that survives corporate failure, the way source code escrow already protects software customers when a vendor fails, and the way safeguards already protect against diversion of nuclear material. The second cannot be built unilaterally at all. Because verification of AI weights may not have a nuclear-style solution, this is a narrower ask than mutual inspection. It is a dedicated channel for shared early warning and agreed response protocols, specifically for the scenario of loose frontier capability, insulated as much as possible from the broader competitive relationship, the way Cold War arms control dialogue was insulated from the rest of the rivalry. Custody without a coordination channel just relocates the problem to whichever jurisdiction the collapse happens in. A coordination channel without domestic custody has nothing concrete to coordinate around.

What this is not

This is not a call to slow AI development, and not a prediction that a correction is imminent. It is not an argument that China's governance approach should be adopted wholesale, or that any specific existing initiative is the answer. The argument is narrower than either of those: unpreparedness compounds with silence, and the West currently has both, on a timeline shorter than the one most institutions are planning around.

illuminem Voices is a democratic space presenting the opinions of leading Sustainability Thought Leaders, their views do not necessarily represent those of illuminem.

The world needs sustainability knowledge. At illuminem, no interest group or shareholder can influence our work. Thank you for supporting our mission to make high-quality and independent sustainability information free for all. Every contribution helps. Thank you for donating today.