Divide Anthropic's February numbers and the detection problem takes on a size: 16 million exchanges spread over 24,000 fraudulent accounts works out at roughly 667 exchanges per account [2][16]. At that per-account volume, nothing in the logs screams extraction rig. It looks like a small team running evaluations against an API. The signal exists in the correlations between accounts, in payment instruments and prompt structure and timing, not inside any one of them. Google's disclosure describes a different profile again, a single campaign of 100,000 queries against Gemini [3].
That is the queue the memo's language actually describes. Tens of thousands of proxy accounts to evade detection, plus jailbreaking to pull out proprietary behaviour [4], is a description of work that belongs to trust and safety and account integrity, the same people already sorting card fraud from real signups. The four things the administration committed to are information sharing, private sector coordination, best practices for identifying and mitigating and remediating, and exploring measures to hold actors accountable [5]. None of those four ships a classifier or pays for the humans reviewing its output.
The asymmetry is that the accounts being hunted are also revenue, and a wrong call suspends a paying customer. Providers absorb that cost; the memo does not offer to. Tom Uren, writing at Seriously Risky Business, argues the promised actions are reasonable and will make some difference but will not stop the campaigns, and reads them as the same information-sharing posture applied to two decades of Chinese IP theft [15]. His sharper complaint is the omission: the memo says nothing about semiconductor export restrictions [8].
The evidence he cites for those controls biting is real and also messy. DeepSeek's V4 was significantly delayed after an unsuccessful attempt to train on Huawei's Ascend, and the company reverted to Nvidia silicon for training while using Huawei parts for inference [9][10]. A Chinese tech blogger attributed V4 being text-only to constraints on computing power and cash [12]. Bulk Ascend production has historically happened in Taiwan in breach of sanctions, and Chinese firms cannot match Nvidia's volume [14]. Both levers depend on enforcement; the difference is that one is enforced at fabs and customs, and the other at a login screen owned by a private company.
The number that sets the price on either is DeepSeek's own: V4 trails frontier models by about three to six months, per its technical report [13]. Closing that with compute costs billions. Closing it with 24,000 accounts costs the price of the API calls. Diplomacy, meanwhile, produced a cable to posts, a demarche to Beijing [6], and an embassy statement calling the accusations pure slander [7].
