AI data centers head for the ocean
This AI newsletter covers Panthalassa's $140M raise to build ocean-based floating data centers powered by wave energy, Anthropic co-founder Jack Clark's prediction that AI will train its own successors before 2029, and both Anthropic and OpenAI launching private equity-backed deployment ventures targeting enterprise adoption.
Summary
The newsletter opens with coverage of Panthalassa, an Oregon-based startup that has secured a $140M Series B led by Peter Thiel, valuing the company at nearly $1B. Panthalassa builds autonomous 85-meter steel floating nodes that convert wave energy into electricity for AI compute, cooled naturally by seawater. The nodes navigate autonomously using hull shape alone and transmit results via Starlink. This development is framed as a direct response to growing public hostility toward land-based data center construction, with commercial deployment targeted for 2027.
Anthropic co-founder Jack Clark published a blog post placing 60%+ odds on AI systems training their own successors before 2029. He cited METR data showing AI's independent task duration grew from 30-second tasks in 2022 to 12-hour tasks in 2026, with 100-hour runs projected by year-end. The SWE-Bench coding benchmark also moved from 2% (Claude 2) to 93.9% (Mythos Preview) in under three years, reinforcing the case for rapid self-improvement capability.
On the enterprise front, Anthropic announced a $1.5B Claude services joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs targeting mid-sized companies. Simultaneously, OpenAI is reportedly raising $4B for its own 'Deployment Company' at a $10B valuation. Both initiatives are characterized as AI-native consulting firms helping large businesses integrate AI into complex existing systems.
Additional items include a tutorial on running local AI models on iPhones via the Locally AI app, a community workflow story about using AI to manage a GoFundMe campaign for a paralyzed friend, and brief news items covering Sierra's $950M raise, Elon Musk's settlement outreach to OpenAI, White House AI oversight proposals, and Anthropic's reported chip talks with Fractile.
Key Insights
- Peter Thiel argues that ocean-based compute infrastructure is a realistic near-term alternative to land-based data centers, framing it as 'opening the ocean frontier' rather than waiting for space-based solutions.
- Jack Clark claims that AI's independent work capability has scaled from handling 30-second tasks in 2022 to 12-hour tasks in 2026, and projects 100-hour autonomous runs by end of year — a trajectory he uses to justify 60%+ odds on self-training AI before 2029.
- The newsletter argues that the primary barrier to enterprise AI adoption is no longer model capability but integration into large, messy business systems — a gap both Anthropic and OpenAI are now moving to fill through PE-backed deployment companies.
- The SWE-Bench benchmark result — moving from 2% to 93.9% accuracy on real GitHub coding tasks in under three years — is presented as concrete evidence that AI coding capability has undergone near-complete saturation in a remarkably short timeframe.
- Public hostility toward land-based data center construction is characterized as a structural constraint accelerating investment in alternative compute infrastructure, with ocean and space-based solutions gaining serious institutional backing as a result.
Topics
Full transcript available for MurmurCast members
Sign Up to Access