Anthropic Calls Pause; Its AI Builds Code.

Mazen Al Sabbagh

— 7 min read
Share

TL;DR

  • Strategic Paradox: Anthropic published a proposal for a coordinated industry pause on frontier AI development, citing risks of recursive self-improvement, concurrently mending a months-long rift with the White House after being blacklisted for refusing military application of its models.
  • Legislative Friction: The House "Great American AI Act" immediately faced bipartisan skepticism, with Republicans resisting broad regulation and Democrats opposing state preemption, as states like Vermont and California accelerated their own distinct AI legislative agendas.
  • Sovereignty Assertions: Canada launched a C$2.3 billion national AI strategy explicitly framed around technological sovereignty from US platforms, following similar declarations from the EU and Japan, signaling a global shift towards distinct national AI postures.
  • Capital Allocation: Supabase raised $500 million at a $10.5 billion valuation for backend infrastructure, and Ramp secured $750 million at $44 billion for AI token spending management, indicating late-stage capital is consolidating in critical AI infrastructure and cost governance solutions.
  • Government Equity: The Trump administration confirmed preliminary discussions regarding federal equity stakes in frontier AI companies like OpenAI, echoing Senator Bernie Sanders' recent proposal, marking a potential bipartisan convergence on state ownership that could fundamentally alter industry-government relations.

Lead Story: Anthropic Says the Industry Needs a Kill Switch — While Mending Fences with the White House

Anthropic released a comprehensive post on Thursday, "When AI Builds Itself," advocating for a verifiable industry mechanism to slow or halt frontier AI development. The call centers on the escalating risk of recursive self-improvement outstripping human safety evaluation capacity. Reuters concurrently reported Anthropic’s contention that such autonomous system evolution is no longer purely hypothetical.

Internal data from Anthropic reportedly substantiates this concern. By May 2026, over 80% of code integrated into the company’s codebase was generated by Claude, with human engineers primarily overseeing rather than authoring, as detailed by Cybernews. Decrypt observed Anthropic's characterization of human involvement as increasingly a "slowing factor" in technical work. This framing bolsters the argument for structured pause mechanisms, yet simultaneously functions as a potent demonstration of Claude’s advanced capabilities.

The timing of this announcement prompted immediate market scrutiny. Anthropic recently closed a $65 billion Series H at a $965 billion valuation and confidentially filed its S-1 with the SEC within days. OpenTools highlighted the inherent tension: "The company urging a global pause is also the one spending billions to stay ahead."

A Reuters exclusive further complicated the narrative Friday. The Defense Department had designated Anthropic a "supply-chain risk" in March – an unprecedented label for a US company – after its refusal to permit military use of its models for domestic surveillance or fully autonomous weapons. This designation restricts tens of thousands of contractors from utilizing Anthropic's AI in government projects. Reuters confirmed that tensions have eased following CEO Dario Amodei's mid-April White House visit and discussions with National Cyber Director Sean Cairncross concerning Mythos and critical infrastructure protection. The composite image is of a company simultaneously advocating for an industry-wide pause, preparing a multi-billion-dollar IPO, and navigating its removal from a government blacklist – each move strategically coherent, yet the confluence raises pertinent questions about their driving imperatives.

Separately, Inc. reported a "rare moment of agreement" between Sam Altman and Dario Amodei on biosecurity risks. Both joined Demis Hassabis and Mustafa Suleyman in a joint letter to Congress urging mandatory screening for synthetic DNA and RNA orders, marking the first collective policy stance by all four major lab leaders.

In Other News

The House "Great American AI Act" encountered significant resistance within 24 hours of its introduction. Politico reported Friday on Republican skepticism regarding sweeping federal regulation and Democratic pressure to prevent state preemption. The ACLU formally opposed the draft's three-year state preemption clause. Concurrently, state legislatures are advancing their own frameworks: the Transparency Coalition's weekly update noted Vermont’s ban on AI therapy bots, Illinois sending five AI bills to its governor, and California progressing 30 AI bills into second-chamber committees. The defining tension in US AI governance is now the collision between federal preemption efforts and a rapid surge in state-level legislation.

Canada unveiled a C$2.3 billion "AI for All" national strategy, marking the third sovereign AI initiative from a major democracy this week. Prime Minister Mark Carney announced the plan Thursday, encompassing a C$500 million growth fund for domestic AI companies, a national supercomputer, free AI literacy programs, and restrictions on AI-powered surveillance pricing. The New York Times contextualized it as Carney's "wary of U.S." influence, while Politico positioned Canada to "lead middle powers in the AI sovereignty race." The strategy aims for 250,000 jobs and a 3% GDP increase by 2031, yet notably proposes no new AI regulations – a contrast to the EU's recent Tech Sovereignty Package, which combined industrial policy with aggressive regulatory mandates. On Friday, Japan’s Digital Minister Hisashi Matsumoto warned the country risks becoming an "AI colony" without sufficient progress, defending a bill to amend data protection laws to allow AI developers access to medical and criminal records without individual consent. Three nations, three continents, three distinct sovereignty strategies – all within a single week.

Significant capital continued flowing: Supabase closed at a $10.5 billion valuation, Ramp at $44 billion, and Generalist AI at $2 billion. Supabase secured $500 million in a Series F, led by GIC, effectively doubling its valuation in seven months. CNBC attributed this surge to the demand for backend infrastructure underpinning the rapid deployment of AI-generated applications, where Supabase’s open-source Postgres platform has become a default. Ramp raised $750 million in a round led by ICONIQ, GIC, and Ontario Teachers' Pension Plan, strategically positioning itself to manage AI token spending as a nascent corporate expense category. Bloomberg reported that Nvidia-backed Generalist AI secured $400 million at a $2 billion valuation for real-world robotics intelligence, led by Radical Ventures.

X / Social Pulse

Anthropic's "When AI Builds Itself" publication initially dominated the discourse, though the subsequent Reuters report on its government blacklist status pivoted the conversation Friday afternoon. The prevailing narrative shifted from a simple critique of "raise $65 billion then ask everyone to slow down" to a more nuanced assessment: a company that previously declined to weaponize its technology for the Pentagon is now advocating for voluntary industry restraint while simultaneously repairing its relationship with the White House. Safety researchers largely lauded the specificity of Anthropic’s pause proposal, emphasizing a verifiable mutual slowdown over abstract open letters. Less attention was afforded to the bioweapons letter, despite its potential significance: all four frontier lab CEOs aligning on a written request to Congress for mandatory DNA synthesis screening. On the regulatory front, state-level advocates strongly criticized the "Great American AI Act's" preemption clause, leveraging Vermont's swift ban on AI therapy bots as evidence of states’ more agile and targeted legislative capacities compared to Congress.

One to Watch

The US government is actively confirming its interest in acquiring equity stakes in frontier AI companies. NOTUS reported Thursday that senior administration officials have engaged in preliminary discussions with major AI firms, including OpenAI, regarding federal ownership shares. On Friday, President Trump directly informed reporters that his team would "look into" the concept, elevating it from staff-level exploration to a presidential priority. The administration previously indicated in May its intent to pursue $2 billion in equity stakes. These discussions align with Senator Bernie Sanders’ recent proposal for public ownership of leading AI companies, with Benzinga noting this bipartisan convergence on the idea, albeit from opposing ideological standpoints. Should government equity in frontier AI labs materialize, it would fundamentally redefine the relationship between Washington and Silicon Valley, introducing potential conflicts of interest across regulation, procurement, and national security oversight.

Quick Hits

  • Leading AI lab CEOs signed a joint letter to Congress advocating for mandatory screening of synthetic DNA orders, highlighting the risk of AI tools circumventing existing biosecurity protocols — the first collective policy statement from Altman, Amodei, Hassabis, and Suleyman (Wired)
  • Significant investor overlap exists between OpenAI and Anthropic, with nearly half of OpenAI's investors also holding stakes in its competitor, reflecting a diversified capital allocation strategy across leading frontier models (Wired)
  • Palantir announced Google Cloud Marketplace availability at AIPCon 10, featuring two-way BigQuery-Foundry integrations and deep Gemini-AIP connectivity, alongside a Kirkland & Ellis partnership for AI-driven fund formation workflows (Bloomberg Law)
  • Perplexity AI unveiled a hybrid inference orchestrator at Computex, designed to automatically route AI tasks between local hardware and cloud frontier models without user configuration, with reported revenue at $500 million (MarkTechPost)
  • OpenAI launched "Dreaming V3," a redesigned ChatGPT memory architecture that synthesizes conversational context in the background, replacing the prior saved-memories list. This enhancement is rolling out to Plus and Pro users in the US, with a 5x compute efficiency gain enabling a free-tier rollout within weeks (OpenAI)

The week concludes with a defining paradox. The entity arguably developing the world’s most advanced AI system is advocating for an industry-wide agreement to halt development under specific conditions – while concurrently rehabilitating its relationship with the government that blacklisted it for refusing military application of that same technology. Concurrently, the US Congress struggles to forge a federal AI framework before states fill the regulatory vacuum; within 24 hours, its bill faces opposition from both political flanks, while Vermont, California, and Illinois demonstrate proactive, state-specific legislative momentum. Three nations across three continents – the EU, Canada, and Japan – have used this same week to declare AI sovereignty a national imperative, each charting distinct strategies with varied trade-offs concerning privacy, investment, and reliance on American platforms. Furthermore, a US president confirmed his exploration of equity stakes in the very companies he would also regulate. Apple’s WWDC on Monday promises to reveal how 2 billion devices will route AI requests, and SpaceX prices its IPO next Thursday, notably bundling a frontier AI lab within its valuation. The fundamental question is no longer if AI will be governed, but rather how many conflicting governance frameworks can productively coexist before market and architectural cohesion degrades.

Sources

Lock in. M. mazen@thorterminal.com

Read more