The Oversight Mirage: Anthropic's Singapore Office and the Compliance Theater Crypto Already Ran

The code does not lie; only the founders do. Anthropic confirmed it would open a Singapore office — its second Asia-Pacific footprint after Tokyo — and, in the same news cycle, defended its position on "AI usage oversight." That is the entire substance of the announcement. Three information points. A location. A stance. And an argument that the stance is defensible. No office function. No headcount. No investment figure. No memorandum with Singapore's Infocomm Media Development Authority. No definition of what "usage oversight" actually means. I have spent a decade reading code instead of press releases, and here is the auditor's problem with a story like this: you cannot audit a sentence. You cannot stress-test a press release. You can only audit the thing that executes. So I am going to do what I actually do. I am going to take this announcement apart the way I took apart Project Aether's token sale function in 2018 — not to find a vulnerability, but to find the gap between what the narrative claims and what the mechanics deliver. The Setup: A Governance Story With No Governance Code Anthropic is the AI lab founded in 2021 by former OpenAI researchers who left over disagreements about safety direction. It builds the Claude family of models. It sells primarily through an API and enterprise contracts, with a large share of its distribution running through Amazon Web Services' Bedrock platform. Amazon is its largest strategic investor, with roughly $8 billion committed across multiple rounds. Google is also a strategic backer. The company completed a financing round in early 2025 at a valuation in the $60 billion range. Its brand is built on one word: safety. Responsible Scaling Policy. Interpretability research. A public posture of supporting measured AI regulation. When the company speaks about oversight, it is not speaking off the cuff. It is executing a positioning strategy that predates the Singapore office by years. The Singapore office is the second step of an APAC expansion that began in Tokyo in 2024. Tokyo and Singapore together form a recognizable pattern — a mature, high-willingness-to-pay enterprise market paired with a regional headquarters and government-relations hub. Singapore hosts a dense concentration of multinational corporate regional offices, sits at the center of Southeast Asian policy formation, and is an AWS infrastructure node. The office placement maps almost perfectly onto the AWS enterprise customer map. The reasonable inference is that the Singapore team's core mandate is converting enterprise accounts already inside the Amazon ecosystem. None of that is controversial. It is also not what makes this event interesting. What makes it interesting is the pairing: a geographic expansion announcement and a regulatory-defense announcement, delivered together, as if they were one thing. They are not one thing. And the crypto industry has run this exact combination so many times that I can recognize the choreography from the opening bars. I want to be precise about what I am and am not claiming. I am not claiming Anthropic is a scam. I am not claiming the Singapore office is fake. I am claiming that the announcement is a governance narrative, and that governance narratives have a specific failure mode that the crypto industry has spent eight years documenting in public, at a cost of tens of billions of dollars. The failure mode is this: a compliance claim is cheap to make and expensive to verify, and markets routinely price the claim as if it were the mechanism. The Crypto Precedent: How "Audited" Became a Decorative Word In 2018 I manually audited the token sale contract of Project Aether, a popular ICO from the 2017 boom. I found a reentrancy vulnerability in the sale function that allowed an attacker to drain roughly 40 ETH from the treasury before the team patched it. I wrote up the exploit path on GitHub. The founders never engaged. The limited technical community did. That experience taught me the single most useful lesson of my career: the market does not reward verification. It rewards the appearance of verification. A project that paid for a logo on its website saying "audited" was treated as safer than a project that had actually been read line by line, because the market could not tell the difference and had no incentive to try. That is the mechanism behind every compliance cycle crypto has produced since. The ICO era gave us the audit badge economy — a small number of firms issuing certificates that ranged from rigorous to rubber-stamped, and a market that treated all of them as equivalent. The 2020 DeFi Summer gave us "battle-tested" protocols with rounding errors in their interest-rate models. I spent weeks on a local fork of Compound that summer and found a rounding error in the borrow rate calculation that could push a position into insolvency under high volatility. The core developers acknowledged it. They prioritized liquidity incentives over the fix, because in a bull market, incentives were the growth lever and the fix was a cost center. The trade-off between speed and safety is not a DeFi-specific disease. It is the default state of any industry where compliance is a marketing line rather than an engineering constraint. The 2021 NFT cycle gave us MetaBeast, a collection whose minting contract had an owner function with no access control. Any user could pause minting or mint unlimited tokens. Early buyers were warned. The project launched anyway. I shorted the associated governance token. The rug was pulled before the mint even finished, and $2 million evaporated. The 2022 collapse of Terra's algorithmic peg gave us a different lesson: a governance narrative — the "algorithmic backstop" — that was mathematically impossible to sustain, dressed up in enough technical language to pass a casual reading. I audited the peg mechanism post-collapse and documented the oracle manipulation vectors that accelerated the death spiral. Regulators in the EU cited the work as evidence of predatory design. And in 2024, the restaking cycle gave us "slashing guarantees" that were, in many implementations, governance promises rather than enforceable code paths. The guarantee existed on the website. The enforcement mechanism existed in a multisig that a small group of people controlled. Every one of these was a compliance story attached to a mechanism that did not match it. Every one of them was priced as if the story were the mechanism. The pattern is stable across cycles, across geographies, and across technology stacks. Which is why I do not treat Anthropic's announcement as an AI story. I treat it as a compliance story, and I read it with the tools that compliance stories have taught me to use. The Core Problem: "Usage Oversight" Is a Phrase, Not a Function Here is the first thing an auditor notices about "AI usage oversight." It has at least three distinct meanings, and the announcement does not tell you which one applies. Reading one: model abuse monitoring. The company watches how its models are used, detects misuse — attempts to generate harmful content, to automate fraud, to circumvent safety filters — and intervenes. This is an operational security function. It lives in the product. It can, in principle, be tested. Reading two: user data usage transparency. The company discloses how it handles the interactions it collects, whether it trains on user data, how long it retains prompts, and under what conditions it shares data with third parties. This is a privacy and data-governance function. It lives in the contract terms and the logging pipeline. It can, in principle, be audited. Reading three: a government-relations posture. The company stakes out a position on how AI should be regulated, typically in a jurisdiction where rules are still forming, in order to shape them. This is a lobbying function. It lives in policy documents and meetings. It cannot be audited at all, because it produces no mechanism — only a position. The announcement blends all three into a single word and lets the reader choose the most flattering interpretation. That is not an accident. Ambiguity is the point. A specific commitment can be checked and failed. A vague stance can be checked by no one and failed by no one. This is the same structural move that crypto's compliance culture perfected. "Audited" meant nothing until you asked which auditor, which commit hash, which scope, and whether the scope covered the function that actually held the money. "Decentralized" meant nothing until you asked who held the upgrade keys. "Community-governed" meant nothing until you asked whether the vote was binding or advisory. Every one of those words functioned as a blur, and the blur was load-bearing. "Usage oversight" is now doing the same work. The question I would ask Anthropic is not "do you believe in oversight." The question is: show me the function signature. Show me what is monitored, at what granularity, retained for how long, accessible to whom, and enforced by what mechanism when a violation is detected. Show me the audit log format. Show me the escalation path. If oversight is real, it produces artifacts. If it produces no artifacts, it is a position, not a function. The Enforcement Gap: Who Actually Bears the Cost Compliance has a second property that the crypto industry learned the hard way, and it is the one that matters most for understanding why an AI company would volunteer for oversight. Compliance is not a cost that falls evenly. It is a cost that falls hardest on the smallest participants, which means it functions as a moat for the largest ones. Europe's Markets in Crypto-Assets regulation is the cleanest recent demonstration. MiCA gave the continent apparent regulatory clarity, and the clarity came with stablecoin reserve requirements and CASP licensing obligations whose fixed costs are trivial for a large issuer and fatal for a small one. The small projects did not get regulated out of existence because regulators wanted them gone. They got regulated out of existence because the compliance burden was a fixed cost, and fixed costs are a regressive tax. The same dynamic is now visible in AI governance. A framework that requires documented usage monitoring, data-handling disclosure, and audit trails is cheap for a company with a legal department and expensive for a startup with twelve engineers. Every incumbent that publicly supports "responsible regulation" is, whether it intends to or not, supporting a cost structure that favors incumbents. This is not a conspiracy. It is arithmetic. So when Anthropic defends usage oversight in Singapore — a jurisdiction with a relatively mature AI governance framework, including the Model AI Governance Framework and the AI Verify toolkit — it is doing two things at once. It is adapting to the host country's regulatory expectations, which is standard market-entry behavior. And it is helping to shape those expectations in a direction that a well-resourced company can satisfy more easily than a poorly-resourced one. Both of those things are rational. Neither of them is the same as "we believe oversight is good." I want to be fair here. There is a version of this story where Anthropic is sincere and the moat is a byproduct. There is a version where the moat is the point and the sincerity is the byproduct. The announcement does not let you distinguish between them, because it contains no mechanism. And when a compliance claim contains no mechanism, the rational default is to assume the interpretation that requires the least verification — which is the cynical one. That is not my bias. That is how you read a contract with no consideration specified. The Infrastructure Reality: Single Points of Failure Do Not Care About Narratives Let me move from governance to mechanics, because that is where I am actually competent. Anthropic's compute is heavily concentrated in AWS, including custom silicon — Trainium and Inferentia — alongside GPU clusters. Amazon is the largest investor. The distribution runs through Bedrock. The Singapore office sits on top of an AWS region that is a major APAC node. The reasonable inference, which the announcement does not state, is that the office supports low-latency regional inference for APAC enterprise customers. That is a coherent business rationale. It is also a concentrated dependency. I have spent my career finding single points of failure, and I can tell you that a concentration of this kind is a risk that governance narratives routinely obscure. In crypto, the equivalent is a protocol that depends on a single RPC provider, a single bridge, or a single oracle. The protocol can be as decentralized as its whitepaper claims, but if one infrastructure provider going down takes the whole thing offline, the decentralization is a diagram, not a property. Anthropic's relationship with AWS is a strategic asset and a structural exposure at the same time. It is an asset because it provides compute, distribution, and capital from a single counterparty. It is an exposure because that counterparty is also a competitor's landlord, because pricing power sits on one side of the table, and because any disruption to the relationship is a systemic event for the company. In 2025, I led the audit of a major ETF issuer's cold storage solution and found a side-channel vulnerability in their multi-sig implementation that could leak private keys through timing attacks. I demanded a full rewrite of the signing logic. It cost the client $500,000 in delays. It prevented a potential billion-dollar breach. The lesson from that engagement is not that dependencies are bad. It is that dependencies are invisible until they fail, and the organizations that manage them well are the ones that document them before anyone asks. Nothing in the Singapore announcement documents the dependency. That is not a scandal. It is a gap. And gaps are where the interesting questions live. The Valuation Question: Narrative Premium and the Multiples Problem Anthropic's valuation sits in the $60 billion range against annual recurring revenue that, by public reporting, climbed toward the $1 billion mark during 2024. That is a multiple that would make a public-market analyst reach for the smelling salts. It is justified, if it is justified at all, by growth rate and by the option value of the underlying technology — not by current cash flows. This is a structure crypto investors know intimately. A token with a $10 billion fully diluted valuation and $5 million of annual protocol revenue is priced on narrative, not on earnings. The narrative can be correct — the project can genuinely become important — and the valuation can still be wrong, because the multiple embeds a set of assumptions that must all hold simultaneously for years. When one assumption breaks, the multiple compresses faster than any fundamentals can recover. The Singapore expansion is, in capital-markets terms, an incremental narrative item. It does not change the model, the revenue trajectory, or the capital structure. What it does is provide the kind of "global commercialization" milestone that is useful in the next financing conversation. AI companies, like crypto protocols, run on milestone storytelling because their valuations are forward-looking and their current financials cannot support them. Opening a new office is a cheap way to produce a milestone. It costs a lease and a hiring plan. It buys a paragraph in the next pitch deck. I am not saying the office is fake. I am saying that an office is an operating expense, not a valuation inflection, and the market will treat it as the former until the latter is demonstrated. The demonstration would be regional revenue data, customer counts, and retention. The announcement contains none of them. That is the tell. The Data Sovereignty Collision Nobody Is Discussing Here is the piece of this story that the AI press is missing, and that a crypto auditor notices immediately because we have been living in it for years. "Usage oversight" in a Southeast Asian context does not float free of data governance. It collides with it. Singapore and its neighbors have been building out data localization expectations and cross-border data transfer rules for years, driven by sovereignty concerns and by the desire to capture the economic value of data generated within their borders. An AI company that monitors usage is, by definition, processing interaction data. Where that processing happens, where the logs live, who can access them, and whether they cross borders are all questions that regulators in the region will ask. The oversight claim and the data-sovereignty regime are in tension. Strong oversight requires centralized visibility into usage across regions — you cannot monitor what you cannot see. Data localization requires that visibility to be fragmented along national lines. A company can resolve this tension, but only with explicit architecture: regional processing zones, jurisdiction-specific retention, and disclosure of the trade-offs. The announcement resolves nothing. It asserts a stance and leaves the architecture unspecified. In crypto, we call this the compliance-versus-custody problem. A regulated custodian that holds assets across jurisdictions must satisfy every jurisdiction's rules simultaneously, and those rules frequently contradict each other. The projects that survived did so by making the architecture explicit — by publishing where the keys were, who held them, and what happened under conflicting legal demands. The projects that did not survive said "we are compliant" and let the reader imagine the rest. The Governance Without Teeth Problem One more parallel, and it is the sharpest one. Crypto spent years learning that a governance token is not governance. A vote that does not bind is a survey. A treasury that is controlled by a multisig is not a treasury that is controlled by the token. The industry is littered with projects whose governance mechanisms were decorative — designed to produce the appearance of community control while leaving actual authority with a small group of insiders. The AI governance equivalent is an oversight claim that is not backed by an enforcement mechanism. If Anthropic monitors usage but the consequences of a detected violation are discretionary, then the oversight is advisory. If the consequences are automatic, then the oversight is real, and there will be a documented trigger, a documented response, and a documented audit trail. The difference between those two states is the difference between governance and theater, and it is the only difference that matters. The Contrarian Case: Why the Cynics Might Be Wrong I have spent this piece dismantling the announcement. Let me now do the harder thing and argue against myself, because a critique that cannot survive its own strongest counterargument is not a critique. It is a mood. Here is the counterargument, and it is good. The crypto parallel breaks in one specific and important place. Crypto's governance was cosmetic by architecture. The token holders never had a mechanism to bind the core team, because the chain could not enforce off-chain promises and the contracts were upgradeable by design. The governance was decorative because the technology made it decorative. AI is different. Anthropic does not just write the model. It runs the inference. The oversight it claims can, in principle, be enforced at a layer crypto never had access to — the model weights, the inference pipeline, the training data pipeline, the deployment controls. An AI company can refuse to serve a request. It can suppress a capability at the weight level. It can instrument every interaction by default. The enforcement mechanism that crypto lacked is, for an AI lab, physically present in the stack. That means the governance could be real in a way crypto's never was. And it means the compliance moat is not purely a marketing construct — a company that genuinely invests in usage monitoring, interpretability, and audit infrastructure is building a capability, not just a brand. If that capability is real, then the moat is real, and the incumbent advantage is earned rather than merely extracted. I find this argument persuasive in principle. It is also unverified in this specific case, because the announcement provides no evidence of the mechanism. The counterargument establishes that Anthropic could be sincere. It does not establish that it is. And the burden of proof sits with the claimant — always. A company that asserts oversight and provides no artifacts has not demonstrated oversight. It has demonstrated the assertion. The reason I still write the critique, even granting the counterargument, is that the crypto industry's most expensive mistakes were made by people who correctly identified that a mechanism could work and then failed to check whether it did. The capability was plausible. The implementation was absent. Plausibility is not verification. It has never been verification. The Takeaway: Audit the Function, Not the Statement So here is where I land, and it is not where a headline wants me to land. Anthropic opening a Singapore office is unremarkable. Anthropic defending usage oversight in a jurisdiction with a mature AI governance framework is unremarkable. Both are standard market-entry behavior, executed by a competent company with a coherent strategy. The combination becomes interesting only because it is presented as a governance story, and governance stories have a documented failure mode: they get priced before they get verified. The test is not whether Anthropic believes what it says. The test is whether the oversight it describes produces artifacts that an outside party can inspect. A monitoring system produces logs. An enforcement mechanism produces triggers and responses. A data-governance policy produces retention schedules and jurisdiction maps. A lobbying posture produces nothing, and a posture that produces nothing should be valued at nothing. I don't trust the audit; I trust the gas fees. In this case, the gas fees are the logs. When someone shows me the function signature for "usage oversight" — the scope, the retention, the enforcement trigger, the audit format — I will read it the way I read Project Aether's sale function. Line by line. Until then, the announcement is a sentence, and sentences do not execute. The question for the next twelve months is not whether AI companies can govern their models. It is whether anyone outside them will ever be able to check.