The Cartel Machine: SAFA, the FINRA Fallacy, and the Coordination Trap

On September 15, 2026, three of the most valuable companies on earth announced a standards body in the same week that their closest competitors refused to join it. Google, OpenAI, and Anthropic — the "coordinating bloc" — launched a self-regulatory alliance. Meta, xAI, and Nvidia walked away. Nvidia, which supplies the compute substrate for every model in the room, declined on the record.

Set the numbers plainly. Three members in, three out. A leadership slate assembled from politics and venture capital rather than model engineering. No statutory authorization. A live antitrust lawsuit filed against one of the founding members. A legislative exemption killed in the Senate three days after the alliance formed.

That is not a governance layer. That is a coalition with a missing quorum.

I have spent years auditing coordination mechanisms — DAOs, validator sets, algorithmic stablecoins, standards consortia — and they all fail in the same place. The failure is never the code. The failure is the authorization layer underneath the code. SAFA just built a validator set without publishing the consensus rule, and the entire debate is about the members while the empty chair where the law should sit goes unmentioned.

The story surfaced through a blockchain news feed, which is itself a data signal. Coordination failure is not an AI problem. It is a structural problem, and the crypto industry has been running the same experiment in public — with the same outcomes — for eight years.

Context: What SAFA Actually Is

The Safe AI Frontier Alliance — SAFA — is proposed as an industry-led body with four nominal functions: standardized model testing, credentialing of independent auditors, incident-reporting protocols, and material support for third-party evaluation firms. Membership is voluntary. Compliance is a signal, not a constraint. Non-participants face no formal consequence — which is the first thing that should trouble anyone who has read a DAO charter.

The leadership slate tells you what kind of institution this is meant to be. SAFA floated Sriram Krishnan, a former White House advisor; Arati Prabhakar, a former Biden-era technology official; Condoleezza Rice, a former Secretary of State; and the venture investor David Friedberg. That is a cross-partisan, cross-domain bench assembled for political legitimacy, not technical capability. Nobody on that list trains models. The roster is a lobbying asset dressed as a standards committee.

The opposition was immediate. On September 15, at Dreamforce, Meta, xAI, and Nvidia stated their refusal to join. Jensen Huang framed the position bluntly: AI-specific law is "completely unnecessary," and model makers should own their products' outcomes. Read that twice, because it is a liability argument, not a safety argument. Nvidia does not want responsibility for model behavior extended down to the compute layer. As the sole supplier of the substrate, it has the most to lose from any threshold that touches training scale.

The political environment is actively hostile. On September 19, the sitting administration called AI safety a "scam" and stood up an "AI Force" with no safety mandate. A draft executive order was shelved. In the legislature, Senators Hawley and Cruz blocked an antitrust exemption in the National Defense Authorization Act — the single provision that could have given SAFA a legal runway. And in the courts, Buist et al. v. Anthropic PBC et al. alleges coordinated safety standards constitute a cartel. Cohere's CEO called SAFA "a cartel by another name." That is not rhetoric. It is a pleading.

One thread to hold. OpenAI ran a second training pause in this window and released its own third-party evaluation framework on September 22 — one week after the opposition bloc walked. The sequence is the evidence. Whoever defines the methodology first defines the standard, and whoever defines the standard gets to decide who can afford to meet it. I will return to this.

Core: Four Failures, One Architecture

Failure One: The FINRA Fallacy

The defense of SAFA leans on a comparison to FINRA, the Financial Industry Regulatory Authority. It is a strong-sounding analogy and it is structurally false.

FINRA works because Congress handed it a mandate through the Securities Exchange Act. That statutory authorization is what converts a private club into a quasi-public regulator. It grants enforcement teeth. It grants antitrust immunity. Without that authorization, a body of competitors setting standards is not a regulator — it is the exact thing antitrust law was written to break up. Strip the authorization and you do not have FINRA. You have a trade association with subpoena envy.

The reviewed material put it precisely: every standard SAFA sets is simultaneously a safety claim and a potential exhibit in the next cartel complaint. That is not a design flaw. It is the default state for any coordination mechanism without a legal basis.

I learned this early, auditing a token contract in 2018. The contract had three reentrancy vulnerabilities and an integer overflow that could have drained roughly $50,000 in ETH, and the team's published audit said the code was safe. The audit was not wrong because the auditors were incompetent. It was wrong because nobody had forced them to run the same checks. Authorization is the difference between a review and a guarantee.

When I deconstructed the governance architecture of a stablecoin consortium in 2020, the members spent six weeks arguing about quorum thresholds while the one question that mattered — who can compel compliance — went unanswered. In every system I have audited, the answer is the same: whoever holds the legal mandate. Consensus thresholds are decoration without it. On-chain, the mandate is the hard fork. Off-chain, it is the statute. SAFA has neither, and no amount of leadership-branding substitutes for a missing legal primitive.

Failure Two: Capture as a Feature

Now the part that gets buried under the safety discourse. The compliance standard is a moat, and the moat is built by the incumbents who will live inside it.

The mechanics are old and boring. Incumbents define a benchmark. The benchmark carries real cost to meet — auditors, documentation, evaluation runs, legal review. Large labs absorb that cost as a rounding error. Small labs do not. The barrier is never worded as a barrier. It is worded as a safety floor.

Anyone who has watched a token-standard fight on-chain will recognize this instantly. When the dominant validators or the largest protocols converge on a "best practice," the practical effect is to raise the cost of forking or competing. The stated reason is robustness. The outcome is concentration. I have audited enough governance upgrades to know that the stated reason and the outcome rarely contradict each other — they just serve different populations. The developers writing the standard and the teams trying to fork it have different interests, and the standard encodes the interests of whoever wrote it.

Scale it up. If SAFA certification becomes a condition for enterprise procurement, or for access to regulated sectors like finance and healthcare, membership stops being voluntary in any meaningful sense. It becomes a license. And a license issued by an alliance of the three most capable competitors is not a safety instrument. It is pricing power with a compliance badge.

The uncomfortable question is not whether SAFA members want safety. It is whether the safety floor is set at a height where the incumbent can afford the lobbyist and the challenger cannot afford the audit. In traditional finance, that question has a settled answer: standards enacted by statute, applied to everyone the same way. Here, the standard is being drafted by the three firms with the most to gain from a floor only they can clear.

There is a name for this in every crypto post-mortem. It is called "the founding team defines the token." Every launch that failed had a team that wrote the rules, held the keys, and called it decentralization. SAFA is that launch, at the scale of national infrastructure.

Failure Three: The Cartel Problem Is a Consensus Problem

Strip the vocabulary and Buist v. Anthropic is asking a coordination question, not a safety question. When do independent actors setting a shared rule cross from "promoting best practice" into "restraining trade"? Antitrust law has an entire doctrine for this — the rule of reason — and it has spent a century concluding that the answer depends on two variables: market power and whether the coordination has a legitimate pro-competitive justification.

SAFA has both squarely in play. The market power is unconcealed — three firms at the top of the frontier. The justification is arguable — safety coordination could be genuinely pro-competitive if it lowers systemic risk. The problem is the design. A voluntary body with no enforcement cannot prove a safety benefit, because it cannot produce uniform compliance. It also cannot prove it is not exclusionary, because it has no admission standard and no appeal process. It fails the justification test in one direction and the enforcement test in the other. It sits exactly where antitrust law is least forgiving.

This is where contract experience sharpens the read. Voluntary governance in any system is a soft fork with no activation threshold. Participants can signal support; nobody can be forced to run the software. In blockchain terms, SAFA is a proposal every validator is free to ignore. That is acceptable for a feature flag. It is fatal for a security standard, because the security property you actually want is uniformity — everyone running the same rule at the same time. A voluntary body can never guarantee uniformity. The only mechanisms that guarantee it are a hard fork with consensus, or a statute with enforcement. In the real world, that mandate is legislation.

The reviewed analysis framed the trap cleanly: without legislative authorization, every standard SAFA sets is "both a safety claim and possibly evidence in the next cartel complaint." That is the entire problem in one sentence. The alliance is trying to act like law without being law, and antitrust is the doctrinal tripwire that catches exactly that behavior. The three firms did not design a regulator. They designed a proposal, and proposals with market power attached are the thing courts distrust most.

There is a specific detail here that on-chain governance researchers will recognize immediately. SAFA's proposed functions include credentialing auditors and defining incident-reporting protocols — in other words, deciding who counts as a qualified observer. That is not a technical function. It is a gatekeeping function. When I led technical due diligence on a ZK-Rollup design in 2025, the hardest four months were not spent finding the throughput bottleneck in proof generation. They were spent arguing about who was qualified to define "scalable." The metric is never neutral. Whoever defines "qualified" has defined the market.

Failure Four: The Three-Bloc Fragmentation

The American vacuum does not stay American. It exports.

Two mandatory regimes already function. The EU AI Act defines obligations by risk tier and is in force. China's large-model filing and registration regime applies a pre-deployment gate. The United States is debating whether a private club is a cartel. The asymmetry is the story.

If the permissive bloc — Meta, xAI, Nvidia — wins the domestic argument, American frontier labs operate under lighter constraints than their European and Chinese counterparts. That is a regulatory-arbitrage advantage, and it is precisely the kind of advantage that reshapes global competition over a five-year horizon. Security and compliance cost time and money; the players who avoid both get speed and margin, and speed compounds. If the coordinating bloc wins instead, compliance capability becomes a moat, and the EU framework becomes the de facto reference for anyone who wants European market access.

Either outcome fragments global governance. The optimistic read — that voluntary coordination fills the gap mandatory law should occupy — collapses on one detail: a gap filled by a body with no authority is not filled, it is re-labeled. When the United States has no statute and no enforcement mechanism, a voluntary club is a public-relations instrument with a rulebook stapled to it.

Then there is the third possibility almost nobody is modeling. If Buist produces a ruling that coordinated safety standards are an unlawful restraint of trade, the chilling effect will not stop at the frontier. It will suppress the same voluntary coordination happening quietly in open-source release norms, in biosecurity, in model-incident sharing — everywhere a group of competitors tried to agree on a safety floor without a government compelling them to. The most expensive line in this whole affair may not be a fine. It may be the precedent.

Contrarian: The Blind Spot Is the Precedent, Not the Cartel

Everyone is litigating whether SAFA is a cartel. Almost nobody is pricing the second-order effect.

Three things are true at once, and the discourse collapses them into one. First, SAFA as designed is plausibly exclusionary. Second, SAFA as designed is functionally toothless — voluntary compliance with no consequence is not a standard, it is a suggestion. Third, the lawsuit that kills it will also kill the incentive for any future voluntary safety coordination. Everyone is arguing about the first. The third is the one that compounds.

Here is the inversion. If a court holds that competitor coordination is per se unlawful in this context, the rational move for any lab becomes simple. Do not coordinate. Do not join. Do not sign the pledge. Do not share the incident report. Walk alone, and eat the reputational damage if a model misbehaves, because the legal damage from cooperating is larger and more certain. The AI safety movement may deliver, through its own legal defeat, the greatest single reduction in frontier safety coordination the industry has ever seen.

The "revolutionary" framing is also wrong, incidentally. There is nothing new here. Airlines, pharma, credit-rating agencies, the ESG complex — every one ran the same playbook, and every one ended either with a statutory mandate that made capture structurally impossible or with a dead cartel. AI is not special. It is late.

The genuinely contrarian observation is that the dangerous failure mode is not "SAFA becomes a cartel." It is "SAFA is legally smothered, and the industry internalizes that safety coordination itself is litigable." The first outcome concentrates power. The second destroys the coordination mechanism entirely. Concentration plus fragmentation is a strictly worse equilibrium than concentration alone, and it is the one the current litigation path is most likely to produce.

Takeaway

The next eighteen months will answer a question that has nothing to do with model capability: can private actors coordinate on a public safety standard when the legislature has declined to authorize it? The docket currently says no. If the courts confirm that answer, the next frontier lab that tries to build a safety consortium will be negotiating with a precedent, not a competitor.

Watch the Buist docket, not the press releases. The revolutionary moment was never the alliance. The revolutionary question is whether anyone can legally form one.