Who eats the compliance bill: Sacks and Amodei argue thresholds while OpenAI pauses training

Dario Amodei used a lengthy post on X, a platform he rarely posts on, to defend Anthropic's approach to regulation and its habit of talking about AI's risks ^[1]. The post answered investor Gavin Baker on the "All In" podcast, cohosted by former Trump AI czar David Sacks, who is no fan of the company ^[2], and it matters because the same week's ledger includes Anthropic tracking toward $65 billion in annual revenue ^[13] and OpenAI detailing a two-week training pause plus additional security controls after a Hugging Face hack ^[14].

The provocation: Baker said multiple people he trusts told him Amodei had claimed Anthropic was so confident in AI's potential and its own position that "Anthropic might be the only private company in the world at some point" ^[3]. Baker called that a "maximalist" vision in which only Anthropic and the U.S. government would decide who gets access to very powerful AI ^[4]. Sacks called it "hubristic" and repeated his charge of regulatory capture: that Anthropic uses fear of AI risk to get rules only Anthropic can comply with easily, squeezing out startups and open-source models ^[5]. Anthropic's chief brand and communications officer, Sasha de Marigny, called the account "Complete and utter nonsense" on X ^[6]; Sacks noted that Amodei himself never addressed the claim directly ^[7].

Strip the personalities out and what remains is an argument about who pays. Amodei's defence is a design claim: Anthropic tries "very hard to make proposals that disadvantage (slow down) frontier AI companies while *advantaging* smaller competitors," pointing to rules it has backed that exempt companies below a revenue threshold or below a training-spend threshold, or that apply only to cutting-edge models ^[8][9]. He also argued that compute requirements concentrate economic power on their own, that this is not the same as predicting one or two surviving firms, and that open weights only partly solve it because they still need compute ^[10], and said he wants "rules of the road" that "leave room for open-weights models while also addressing the specific risks that they bring" ^[11].

Thresholds decide who is nominally exempt. Scale decides who actually feels the cost. At a run rate near $65 billion, Anthropic is far above any small-company carve-out in the proposals it says it supports, so it would sit on the paying side of its own thresholds ^[17] - which is the strongest fact in its favour and also the reason those carve-outs are cheap to offer. For a company below the line, an exemption is a deferral, not a discharge: the threshold is crossed by whoever succeeds, and the controls have to be built before the crossing.

OpenAI's disclosure is the useful control experiment. Two weeks of paused training and a tightened security posture arrived from an attacker, not a statute ^[14]; Fortune's summary lists the pause and the new controls without attaching a cost figure ^[18]. Frontier overhead is being set by incidents whether or not legislatures act, and enterprise buyers tend to ask smaller vendors for the same controls regardless of exemption language. Baker's separate complaint, that Amodei's risk talk has fed negative public sentiment and local opposition to data center construction, is a cost argument too, just one denominated in permits rather than payroll ^[15].

Watch for specifics: whether Anthropic names the revenue and training-spend numbers in the bills it endorses, whether OpenAI quantifies the pause in compute or schedule, and whether other labs disclose comparable stoppages. Fortune's own framing is that Sacks says Amodei wants a "DMV for AI," while plenty of regulated industries do fine ^[16]; the test is which line the thresholds actually fall on.