On-chain data rarely lies. Off-chain governance, however, is a different beast entirely.
Consider this anomaly: two of AI's largest model providers—Anthropic and OpenAI—are reportedly coordinating with the incoming Trump administration to define the evaluation standards for frontier AI systems. On the surface, this reads as a necessary step toward safety alignment. Dig deeper, and the on-chain signature of incentives becomes clear. This is not collaboration. It is a fork.
Context: The Protocol of Trust
In the blockchain world, we define trust through code and consensus. A proposal that passes a governance vote is executed—no exceptions. AI safety standards, by contrast, have no such deterministic finality. They are subjective, negotiable, and infinitely malleable. The joint effort between Anthropic and OpenAI to co-author an "AI Model Evaluation Plan" with a new administration is effectively a governance proposal—but one where the voting power is allocated by political proximity, not token holdings.
The key metadata here is timing. Both companies have spent years competing—Anthropic’s constitutional AI vs. OpenAI’s RLHF, their differing stances on open-sourcing, their public spats over safety culture. Yet now, they whisper in the same ear. Why?
Core: Deciphering the Hidden Geometry of Influence
Let me reconstruct the transaction flow. First, we observe that the Trump administration is historically skeptical of heavy-handed regulation, but hawkish on national security. The AI Evaluation Plan, therefore, is not about protecting consumers—it is about defining a whitelist of acceptable actors.
Follow the trail of outliers that others ignore. The outlier here is that Anthropic and OpenAI are _not_ the only two frontier labs. Google DeepMind, Meta, Microsoft, Mistral, and numerous open-source projects are also developing advanced models. Yet only these two have a seat at the table. This is not a technical decision—it is a political allocation of validation authority.
I mapped the incentive vectors: Anthropic needs capital and legitimacy to sustain its safety-research-heavy burn rate. OpenAI needs to maintain its de facto regulatory capture to justify its closed-source business model. Both benefit from a standard that requires massive compliance overhead—the equivalent of setting a gas limit that only whales can afford.
The empirical data supports this. In Q4 2024, Anthropic raised $4B at an $18.4B valuation—a cash pile that burns at $2B annually, largely on safety research. OpenAI’s revenue is projected at $3.7B in 2024, but it still loses money on compute. A government-endorsed evaluation framework that mandates Red-teaming, bias audits, and interpretability certifications will force smaller labs to either partner with the incumbents or pay licensing fees for compliance tooling. This is not safety—it is a toll bridge.
The algorithm does not lie, but it may omit. What the press release omits is the economic moat. If the evaluation standard requires access to proprietary alignment data (which Anthropic and OpenAI guard closely), then every newcomer must beg for API access or license data from them. The result is a centralized validation oracle under the guise of public safety.
Let’s quantify the moat. According to my back-of-the-envelope model based on GPU compute costs and researchers’ salaries, the compliance cost for a mid-size AI startup to meet a hypothetical “grade A” evaluation standard could exceed $50M annually—nearly 40% of a typical Series A round. Compare that to the $0 they spend today. The standard itself becomes a barrier to entry.
Contrarian: Correlation ≠ Causation
Here is the counter-intuitive angle: this cooperation might actually _increase_ systemic risk, not reduce it.
Think about the DAO collapse in 2016. The cause was not too little governance—it was that governance was concentrated in a small set of auditor contracts that all used the same flawed code. By centralizing the evaluation process under a single political and corporate alliance, we create a monoculture of safety assumptions. If that standard is flawed—say, it over-penalizes certain model behaviors while ignoring adversarial vector attacks—then everyone who complies is vulnerable to the same failure.
Moreover, the current bull market euphoria around AI capabilities masks this technical flaw. The narrative is “safety through partnership,” but the code of incentives reveals a different story. The partnership is with a political entity that has a one-term horizon. Any standard forged today can be repealed or weaponized after a regime change. Voting power in this governance system is not eternal—it expires when the administration does.
I’ve seen this pattern before, in the 2021 NFT floor price anomaly. When 60% of trading volume was wash-traded, the market believed it was demand. The reality was a single bot cluster inflating prices. Here, the bot cluster is the collusion between two dominant labs and a transient political majority. The volume of “safety” noise is high, but the genuine liquidity of good-faith global cooperation is thin.
Takeaway: Next-Week Signal
Watch the open-source community—Meta’s Llama, Mistral, and community-driven fine-tunes. If they begin creating their own evaluation standards outside the Anthropic-OpenAI axis, we will see a fork. The real market signal is not the collaboration itself, but the reaction of those excluded from it.
The question I leave with readers: when the evaluation oracle is centralised, who audits the auditor?