King Charles is preparing to host the heads of the world's largest artificial intelligence laboratories. Dario Amodei of Anthropic and Sam Altman of OpenAI have issued synchronized warnings that rapidly improving AI systems are becoming progressively harder to control. The wires framed the meeting as a landmark moment in AI governance. I read it as a ledger entry that was never posted.
Every entity expected at that table derives authority from a claim it cannot cryptographically prove. Anthropic publishes Constitutional AI papers, not verifiable computation proofs. OpenAI publishes system cards, not zero-knowledge attestations of training runs. The behavioral guarantees of a frontier model rest entirely on trust in a corporate disclosure, reviewed by parties with commercial incentives to approve it. That is not governance. That is the pre-blockchain banking model wearing policy language. And this is precisely the gap that the crypto-AI convergence has been quietly filling while the policy class debates restraint in London drawing rooms.
To understand why the London meeting matters — and why it matters less than the participants believe — you have to map the actual architecture underneath it.
The current AI governance stack has three layers. The first is the policy layer: voluntary commitments, national frameworks like the NIST AI Risk Management Framework, the EU AI Act's treatment of general-purpose models, and now a royal-hosted convening in the United Kingdom. The second is the disclosure layer: model cards, system cards, evaluation reports, red-team summaries. The third is the computation layer itself — the weights, the training runs, the inference calls, the agent-to-agent messages.
Amodei and Altman operate almost entirely in layers one and two. Their warnings are policy artifacts. They call for restraint at the frontier, for evaluation regimes, for governance mechanisms. What they do not call for — because their business models cannot survive it — is cryptographic verification of the third layer. That layer is where the crypto industry has spent the last four years building.
The UK government, meanwhile, has positioned itself as the "third way" between the EU's precautionary regime and the American innovation-first posture. King Charles, a known climate advocate and technology-adjacent figure, provides symbolic cover. The problem is symbolic. A monarch who does not write code cannot audit a model. The meeting will produce communiqués, not commitments with teeth. The consensus will be that AI is consequential and caution is warranted. The consensus, as always, is the contrarian trap.
Here is where I need to bring my own audit record into the frame, because the London conversation is missing the only mechanism that has ever worked for verifying machine behavior: cryptographic proof.
In 2017, I spent 400 hours auditing the smart contract logic of an early DeFi prototype. The team had raised a substantial round on the strength of a tokenomics model that assumed rational actors. I found a reentrancy vulnerability that could have drained roughly $50 million. The team's response was to publish a blog post. I refused to invest. The blog post was not a proof. The code was the proof, and the code was broken.
The same forensic logic applies to frontier AI. When Anthropic says its model is aligned, it has published no verifiable artifact that a third party can check. When OpenAI says a deployment is safe, the safety claim is adjudicated by OpenAI. The ledger remembers what the market forgets: trust without verification is a liability carried at par until the moment it isn't.
The crypto-AI convergence solves this in a way the policy layer cannot. Zero-knowledge proofs of computation allow a model to prove that a given output was produced by a given set of weights running a given inference routine — without revealing the weights themselves. This is the cryptographic trust layer for autonomous AI, and I have been writing about its necessity since early 2026, when it was still a niche research interest rather than an emerging asset class.
Let me map the mechanism concretely, because abstraction is where policy conversations go to die.
A zero-knowledge proof of inference works like this. The prover — the AI operator — commits to the model weights and the input. It runs the inference. It then generates a succinct proof that the output corresponds to the committed weights and input according to the committed architecture. A verifier, which could be another agent, a smart contract, or a regulator, checks the proof in milliseconds. The weights never leave the operator's environment. The proof guarantees that the computation was performed honestly.
This is not speculative. The mathematics has been available for years. What has changed in the last eighteen months is the cost. Proof generation for transformer inference was, until recently, orders of magnitude more expensive than the inference itself. Recent recursive proof systems and specialized hardware have brought that ratio down to a range where high-value agent transactions can economically justify proof generation. The crossover has not yet happened for consumer chatbots. It is happening right now for machine-to-machine settlement, which is precisely where autonomous AI agents will operate.
Why does this matter to the London meeting? Because the entire premise of the meeting — that AI development should be "restrained" by human institutions — assumes that human institutions can observe what the models are doing. They cannot. Without proofs of computation, no regulator can distinguish a model that is following its stated policy from a model that is not. Amodei's warnings about "systems becoming harder to control" are, at the deepest level, admissions that the control layer does not exist. He is describing a jailbreak of governance, not a jailbreak of the model. And he is describing it in the presence of the one technology that could patch it.
The ledger remembers what the market forgets. Now let me widen the frame to liquidity, because that is where the AI-crypto convergence will be felt by portfolio holders first.
The current bull market is priced on a narrative that AI and crypto are separate asset classes. They are not. They are converging on the same substrate: verifiable computation. The AI labs need proof systems to make machine trust tractable. The crypto economy needs real demand for blockspace beyond speculative trading. The intersection is agent-to-agent payments, where an AI agent pays another AI agent for a verified computation, and the settlement is a zero-knowledge proof on an L2. The infrastructure layer of this convergence is where I positioned my fund before the market recognized the necessity of cryptographic verification for machine trust.
Notice what this means for the London meeting. The policy class is debating restraint at the same moment the technical class is building the rails that make restraint irrelevant. You cannot slow a system whose incentives are cryptographic and whose settlement is permissionless. You can regulate the legal entities that operate today's labs, but you cannot regulate the arithmetic. Architecture reveals the true intent, and the architecture of autonomous machine economies points toward proof-based settlement, not discretionary governance.
Let me be more specific about the structural risk that the London meeting is not auditing.
The AI policy conversation assumes a stable set of institutional actors: the labs, the regulators, the governments. It assumes that these actors will implement the agreed restraint over a multi-year horizon. It assumes that the frontier will advance at a pace that human institutions can track. Every one of these assumptions is a point of failure.
Amodei and Altman's own companies are the counterexample. Anthropic and OpenAI are locked in a capability race with each other, with Google DeepMind, with xAI, and with open-weight models that the Chinese and American open-source communities keep releasing. The restraint they are publicly endorsing has no enforcement mechanism. It is a commitment that binds no one and obligates nothing. In audit terms, it is an unsecured note from a counterparty that has every incentive to default quietly while claiming compliance.
This is the structural risk audit that belongs at the center of any serious AI-crypto convergence thesis. The London meeting will produce a communiqué. The communiqué will be non-binding. The non-binding commitment will be cited in marketing materials. The marketing materials will be used to raise capital. The capital will be deployed into capability races that outrun the commitments. The cycle is the same one I documented in 2022 with opaque custodial arrangements in the CeFi sector: trust extended to institutions that cannot prove their claims, until the claims are tested and the trust evaporates.
Patterns repeat, but the participants change. The crypto industry has a word for this. It is called a rumored liability. It sits off-balance-sheet until it doesn't. The FTX collapse was a rumored liability materializing. The Celsius unwind was a rumored liability materializing. The Terra collapse was a rumored liability materializing. Each time, the disclosure layer said one thing and the computation layer did another. Each time, the market learned the same lesson and then forgot it within a cycle. The London meeting is the AI sector's version of the same disclosure gap, at a scale measured in trillions rather than billions.
Mapping the invisible currents of liquidity in this context means tracking, not the AI narrative, but the flow of verifiable compute. Where are the proof-generation costs declining fastest? Which L2s are building native verification precompiles? Which agent-payment protocols are integrating zk inference proofs into their settlement layers? These are the metrics that will determine which assets capture the AI-crypto convergence, and they have almost nothing to do with what gets said at Buckingham Palace.
Let me address the contrarian thesis directly, because this is where most analysts writing on this topic get it wrong.
The mainstream interpretation of the London meeting is that it marks a regulatory pivot: AI is becoming governed, and therefore AI-adjacent assets will face headwinds. The mainstream is positioning defensively.
I take the opposite position. Regulatory attention at the frontier is the single strongest catalyst for the verifiable-compute thesis. Here is the mechanism. When regulators increase the cost of unverified claims — when they require disclosure, audit, or liability — the relative advantage shifts to architectures that can produce proofs automatically. A zero-knowledge inference proof is not a compliance burden. It is a compliance shortcut. It reduces the friction of demonstrating that a computation was performed as claimed, because the demonstration is mathematical rather than procedural.
The labs that resist verifiability — the ones that want to keep their weights opaque and their evaluations private — will find that their political capital erodes as the regulatory bar rises. The labs and protocols that build on proof systems will find that compliance becomes a feature, not a cost.
Certainty is a liability in this domain. The analysts who are certain the London meeting is bearish for crypto-AI assets are the same analysts who were certain the ETF approval was bearish in January 2024. They read events through a sentiment lens. The event is a signal extraction problem, not a sentiment problem. The signal is that the policy layer has admitted it cannot control the computation layer. The extraction is that verifiable computation is now the scarce asset.
Let me return to the ledger one more time, because I have not yet made the sharpest point.
The attendees at the London meeting will discuss AI risk as a category. They will not discuss AI risk as a specific, auditable phenomenon with a specific cryptographic remedy. The closest they will come is the language of "evaluation" and "red-teaming," which are procedural rather than mathematical. A red-team report is a document. A zk proof is a fact. The difference between a document and a fact is the difference between the pre-Satoshi financial system and the post-Satoshi one.
Twenty-nine years in this industry have taught me one durable lesson. The market prices narratives until the narratives meet architecture. When they meet architecture, the architecture wins. The London meeting is a narrative event. The zk-inference infrastructure is an architectural event. The narrative event will be remembered for a week. The architectural event will be remembered for a decade.
There is a secondary point that the crypto press has not yet surfaced. King Charles's involvement is a signal about which jurisdictions will host the verifiable-compute economy. The UK has been explicit about wanting to be a global hub for AI safety research. The country has also been slow to provide regulatory clarity for stablecoins and DeFi. If London positions itself as a governance hub while the actual verifiable-compute builders move to jurisdictions with clearer crypto frameworks — Switzerland, Singapore, the UAE, increasingly the United States — then the governance hub will govern an economy that has already relocated. Survival is a function of position sizing, and the position sizing of AI-crypto builders is tilting toward jurisdictions that recognize proofs as legal infrastructure, not just technical curiosities.
The most important thing the market is missing is that the AI chiefs are calling for restraint at precisely the layer where restraint is impossible, while ignoring the layer where restraint is automatic.
You cannot restrain a frontier model through policy because the model's behavior is determined by arithmetic, and arithmetic does not negotiate. You can restrain a frontier model through cryptography because the cryptography determines what the model can prove about itself, and a model that cannot prove its behavior will not be trusted by other agents. The market is treating the London meeting as a governance event. It is actually a failure-of-governance event. The correct trade is not to fade AI-crypto assets on regulatory headlines but to accumulate the verifiable-compute infrastructure that regulatory pressure will make mandatory.
The second blind spot is the composition of the attendees. The meeting is framed as a conversation between AI and government. The absent third party is the cryptographic settlement layer. No representative of the proof-system economy will be in the room. This is the same structural omission I documented in exchange proof-of-reserves reporting: the entities that can verify are excluded from the venues where verification is discussed. The consensus is often the contrarian trap, and the consensus here is that governance is the answer. The answer is arithmetic.
When the communiqué is published, watch what it does not mention. If it does not mention verifiable computation, zero-knowledge proofs of inference, or cryptographic attestation of model behavior, then the policy layer has confirmed it intends to govern blind. Position accordingly. The ledger remembers what the market forgets, and the market is about to forget a very large number.
