On May 14, 2024, Jan Leike—the co-lead of OpenAI's Superalignment team—walked out the door. Three days later, Ilya Sutskever, the company's chief scientist, followed. The official narrative blamed "creative differences." The technical reality was simpler: two of the most prominent AI safety researchers alive had concluded that OpenAI could not be trusted to build safely. Predictability is a myth; only volatility is real. The market, fixated on GPT-5 launch dates and revenue multiples, priced this information at approximately zero.
This is not an isolated personnel matter. It is a structural failure mode that I have documented before—in cryptocurrency protocols that collapsed under the weight of incompatible incentive architectures. The parallel is not metaphorical. It is architectural.
Anthropic and OpenAI are now trapped in a recursive death spiral identical to the one that destroyed UST's seigniorage model in 2022. In Terra's case, the mechanism was explicit: algorithmic stablecoins required continuous expansion to maintain parity. Growth was not optional—it was structural. When expansion stalled, the entire mechanism reversed into a catastrophic contraction. The result was $60 billion in market capitalization vaporized in 72 hours.
The AI frontier labs face an equivalent dynamic, though the mechanics are more diffuse. Both companies have raised capital at valuations ($157 billion for OpenAI, $18 billion for Anthropic) that implicitly price in continued capability advancement. Investors did not pay these multiples for incremental improvements. They paid for AGI. The moment either company signals that safety constraints will meaningfully slow development, the valuation architecture faces revaluation risk. The pressure to "keep building" is not a strategic choice—it is a structural necessity baked into the financing architecture.
The Composability Problem Nobody Wants to Discuss
In blockchain systems, composability—the ability for protocols to stack and interact—is often celebrated as a feature. It is also what makes systemic contagion possible. Aave, Compound, and MakerDAO learned this lesson in June 2020 when a 20% price drop in ETH triggered cascading liquidations that wiped out $500 million in value. The protocols were individually sound. The system, composed of multiple interacting components, was fragile.
Anthropic and OpenAI have their own composability problem, except the components are organizational rather than technical. Consider the following structural dependencies:
First, the safety-commercial dependency. Anthropic's "safe AI" narrative is its primary market differentiator. The company's Constitutional AI framework and RLHF depth are positioned as engineering features that justify premium pricing and investor confidence. But this narrative creates a binding constraint: if a Claude instance is demonstrably jailbroken, or if a high-profile safety failure occurs, the company faces a double collapse—loss of customer trust AND loss of investor narrative. The safety positioning is load-bearing. It cannot afford to crack.
Second, the geopolitics-commercial dependency. American chip export restrictions on H800 and A100 GPUs have created what amounts to a de facto moat for OpenAI and Anthropic. Chinese AI developers face structural compute disadvantages that slow their capability advancement. This geopolitical protection is not neutral—it creates implicit obligations. The US government has invested significant diplomatic and economic capital in maintaining AI leadership over China. Companies that receive the benefit of this protection face pressure, formal or informal, to align their development timelines with national strategic objectives. The question is whether those objectives are compatible with rigorous safety evaluation.
Third, the talent-safety dependency. Jan Leike's departure was not an isolated event. It was a symptom of a broader structural problem: the people most qualified to evaluate AI safety risks are precisely the people most likely to object when safety constraints are relaxed for commercial reasons. This creates a perverse selection dynamic where safety-conscious researchers become flight risks during periods of accelerated development. Each departure weakens internal safety capacity, which increases the probability of safety incidents, which increases the probability of further departures.
Why the 2024-2025 Timeline Is Different
History does not repeat, but it rhymes in binary. The AI industry in 2024-2025 is structurally analogous to DeFi in mid-2020, not because the technologies are similar, but because the incentive failure modes are identical. In both cases, you have a technology with genuine transformative potential, operating under financing structures that require continued growth to maintain solvency, with safety considerations that create tension against that growth imperative.
The critical difference is timescale. DeFi protocols collapsed in hours or days. The AI infrastructure being built today will remain operational for years. A safety failure in an AI system is not a liquidation event—it is a slower-moving, harder-to-reverse erosion of trust and regulatory credibility. The Superalignment team's dissolution at OpenAI is not equivalent to a flash crash. It is equivalent to removing the emergency shutdown mechanism from a nuclear power plant and then continuing to increase output.
The Regulatory Trigger Point
The EU AI Act has moved from legislative text to enforcement reality. High-risk AI systems—those used in employment decisions, credit scoring, biometric identification, and critical infrastructure—are now subject to mandatory conformity assessments. The question is not whether compliance requirements will tighten. The question is whether the frontier labs will be caught in the compliance net before they achieve commercial maturity.
My assessment: the probability of a significant safety incident triggering accelerated regulation before 2026 exceeds 60%. The structural conditions for such an incident are already present. Internal safety capacity is degrading. Development velocity is increasing under investor and geopolitical pressure. The models being developed are operating in capability regimes that safety evaluation frameworks were not designed to handle.
The Infrastructure Valuation Blind Spot
Wall Street analysts covering AI are making the same mistake that crypto analysts made in 2021: conflating market capitalization with infrastructure integrity. The $157 billion valuation of OpenAI is a claim about future cash flows, discounted to present value. It assumes that the company can continue operating at its current strategic posture indefinitely. It does not price in the tail risk of a regulatory intervention or a catastrophic safety failure.
The same dynamic exists in blockchain infrastructure. When I evaluate custody solutions for institutional clients, the first question is never "what is the AUM?" It is "what are the failure modes?" The failure mode analysis for Anthropic and OpenAI reveals a concentration risk that their valuations do not reflect: both companies are single points of failure in a global AI infrastructure that is being built around them. If either company experiences a credibility crisis—through a safety incident, a regulatory action, or an internal governance failure—the ripple effects will propagate through every downstream application that depends on their APIs.
What Happens Next
Three scenarios, ranked by probability:
Scenario A (50% probability): Incremental accommodation. Both companies survive, but development velocity is constrained by regulatory requirements and internal safety investments. Valuations reprice downward by 20-40% as growth assumptions are revised. The "AGI by 2027" thesis recedes to 2030+.
Scenario B (30% probability): Safety incident triggering regulatory clampdown. A high-profile failure—perhaps a model used for autonomous decision-making that causes significant harm—forces emergency regulatory intervention. The industry enters a "regulatory winter" similar to post-UST crypto regulation. Smaller players with cleaner safety profiles benefit.
Scenario C (20% probability): Geopolitical override. The US government formally or informally designates frontier AI development as a strategic national interest, overriding safety concerns in favor of competitive acceleration. The labs become quasi-governmental entities, with corresponding governance implications.
None of these scenarios are positive for current investors at current valuations. The market has not priced in the structural tension between growth imperative and safety constraint. When that tension resolves—through regulation, incident, or revaluation—only those who understood the underlying architecture will be positioned correctly.
The AI industry is building infrastructure for the next decade on a foundation that has not been stress-tested against its own growth dynamics. I have seen this before. The protocols that failed in crypto were not stupid. They were sophisticated systems operating under incentive structures that made failure inevitable. The AI frontier labs are not stupid. They are sophisticated organizations operating under structurally identical conditions. The question is not whether the pressure will resolve. It is how much value will be destroyed when it does.
