TeraWulf signs a 20-year, $19B Anthropic lease, then cashes out of its own joint venture

Good morning 👋 TeraWulf just signed the biggest AI infrastructure lease of the month. The number everyone will quote isn't the part worth your attention. What the company did on the same page is.

In today's issue:

• 🔭 TeraWulf's $19B Anthropic lease, and the exit buried in the same release
• 🧠 Sakana ships real-time translation, MiniMax lands on Bedrock
• 🔬 Verification as the next scaling axis, plus two papers worth the read
• 🏛️ Illinois passes the first mandatory AI safety audit law
• 🛠️ Turn today's lease into a job search
• 🔗 Nova learns to forget, Hugging Face signs its kernels, fusion cashes in on the power crunch

🔭 THE ONE THING

⚡ TeraWulf signs a 20-year, $19B Anthropic lease, then cashes out of its own joint venture

TeraWulf just landed the AI-infrastructure deal of the week: a 20-year lease with Anthropic worth roughly $19 billion in contracted revenue, building out ~401 MW of critical IT load at its Justified Data campus in Hawesville, Kentucky. First power lands in the second half of 2027, full capacity by early 2028. Buried further down the same release: TeraWulf sold its 50.1% stake in the Abernathy joint venture to a Fluidstack-led group, cashing out the $450 million it put in at a premium instead of holding the capacity it just developed. That's the part worth sitting with. A bitcoin miner turned data-center landlord signs the biggest AI lease of the month, then takes the exit on a project of its own rather than ride it for twenty years, which says more about how the smart money is pricing this buildout than the headline number does.

🧠 MODELS & RELEASES

• 🌐 Sakana AI shipped Sakana Translate, a free JP/EN/CN tool baked into Sakana Chat and built on its own Namazu models. Translate handles long text in real time, proofread tracks every change it makes, and ask lets you interrogate a word choice instead of just accepting it. On the WMT 2024 benchmark it scores in frontier-model range. Sakana Translate
• 🧩 MiniMax landed its M2 family on Amazon Bedrock: M2 with a 1M-token context window, plus M2.1 and M2.5, a 230B-parameter mixture-of-experts model tuned for agent-native execution with only 10B active per token. Three pricing tiers, standard, priority for lower latency, flex for cheaper batch runs, mean you can tune cost against speed without leaving Bedrock. AWS blog

🔬 RESEARCH HIGHLIGHTS

• LLM-as-a-Verifier A new framework argues that verification, not more pretraining or longer chains of thought, is the next scaling axis: train a model to grade solutions well, and you get a cheaper lever than training it to produce them. Holds up in-paper. Whether it holds outside the benchmarks the authors picked is the open question. arxiv
• Multiplayer world models Most "world models" treat every other agent as scenery baked into the environment. This one is the first to condition on multiple agents' action streams directly, so it can attribute a physics outcome to who actually did what. Early-stage, narrow domain, but it's the first real step toward simulators for multi-agent robotics and games instead of solo demos. arxiv
• How Much is Left? Turns out LLMs linearly encode how many tokens they have left to generate, in the hidden state, before a single word of the answer exists. That's a tidy explanation for why models ramble or truncate in oddly consistent patterns, and a real hook for anyone trying to build length control that doesn't rely on cutting text off after the fact. arxiv

🏛️ POLICY & REGULATION

• Illinois signed SB 315, the Artificial Intelligence Safety Measures Act, the first state law to require independent third-party audits of frontier AI safety practices. It binds developers of the largest AI systems starting January 1, 2027, and adds mandatory incident disclosure plus whistleblower protections for employees who flag problems. Anthropic publicly backed the bill, which says something about who expects to clear the bar without much trouble. No penalty structure has been published yet, so the real test is what happens the first time a lab misses a disclosure. Governor's announcement

🛠️ TRY THIS

Score data-center jobs before you read them

TeraWulf's 20-year, $19B Anthropic lease means real hiring, for infra and AI platform roles that follow the compute into Kentucky. career-ops runs the search locally, in your coding CLI, so you're not tab-hopping ten job boards by hand.

1. Clone santifer/career-ops; it runs inside Claude Code, Gemini, Codex, or OpenCode.

2. Point a scan at data center infrastructure, power engineering, and AI platform roles tied to the TeraWulf build-out.

3. Let it grade each listing A-F and skip anything under a C.

4. Tailor your CV to the A and B matches, then let it track what you've applied to.

Prompt: Scan job boards for data center infrastructure and AI platform engineering roles connected to TeraWulf's Kentucky build-out and Anthropic's compute lease. Score each listing A-F on fit and flag any requiring relocation.

🔗 QUICK LINKS

• AWS Nova can now be taught to unlearn narrow slices of content on purpose, via a reverse-DPO technique that trims over-blocking without dragging down general quality. AWS
• AWS A second Nova post chains Meta's SAM 3 with Textract to auto-redact faces and text in images. Handy, if you trust a pipeline to decide what counts as PII. AWS
• Hugging Face Kernels finally got a security model worth the name: Sigstore-signed builds, Nix reproducibility, and an explicit opt-in before untrusted code runs. Hugging Face
• NVIDIA Tallied this year's ICML accepted papers and found open frontier models doing most of the citing. Convenient framing from the company selling the GPUs under them, but the count checks out. NVIDIA
• Proxima Fusion Raised €411M at a €2.4B valuation, with Google and RWE among the backers, to build Alpha, its stellarator demonstrator going up near Munich. The AI power crunch is now funding fusion rounds directly. BusinessWire
• AWS Added a one-click jump from any Hugging Face model page into SageMaker Studio, skipping the usual download-and-reupload routine. AWS

See you tomorrow.

Pradeep Perugu