Anthropic flags Chinese open model risk while geothermal quietly solves AI's power problem

What Anthropic actually found in Z.ai's release

Anthropic's Frontier Red Team tested Z.ai's GLM-4-9B, released in late September, and found it could autonomously discover and exploit CVEs at a rate that surprised even their researchers. The model doesn't just suggest exploit code , it chains reconnaissance, payload generation, and post-exploitation steps without human prompting. That's a meaningful step beyond what we saw with Llama 3.1 405B's cyber evals six months ago. The weights are downloadable from Hugging Face right now.

Why the 9B parameter size changes the threat model

Everyone focuses on frontier models, but GLM-4-9B runs on a single A100 or even high-end consumer GPUs with quantization. That puts automated vulnerability hunting in reach of anyone with $2K in hardware and an internet connection. We've tracked similar capability jumps before , when Stable Diffusion dropped to 2B parameters, content moderation became effectively impossible. The same dynamic applies here: capability density per parameter is improving faster than our ability to gate access.

The open weights debate misses the deployment layer

Anthropic argues for staged release and access controls, but the model is already public. The real leverage point isn't weight distribution , it's inference infrastructure. Cloud providers could implement output filtering for known exploit patterns at the API layer, similar to how they block CSAM. Nobody's doing this systematically yet. The conversation stays stuck on 'should we release' while the operational question 'how do we monitor inference at scale' goes unanswered.

Geothermal's 90% capacity factor is the quiet breakthrough

While everyone argues about model weights, Fervo Energy's Utah project just hit 90% capacity factor over 30 days , matching nuclear baseload without the decade-long permitting. Their enhanced geothermal systems use horizontal drilling from oil and gas, not new physics. Google's already buying 115 MW from them for Nevada data centers. This matters because AI training runs need guaranteed 24/7 power; solar-plus-storage tops out around 40% capacity factor in practice.

Why this beats SMRs for the next five years

Small modular reactors get the hype, but NuScale's first commercial deployment slipped to 2030. Fervo's drilling rigs are operating today using existing supply chains. The levelized cost hits $60-80/MWh , competitive with gas peaker plants and half the cost of lithium-ion storage for firm power. For a 100 MW data center campus, that's $50M/year savings over battery-backed renewables. The economics work at 10 MW scale, not just gigawatt plants.

The second-order effect nobody's pricing in

Cheap baseload geothermal changes where compute clusters get built. Right now, AI companies cluster in Virginia and Santa Clara for grid access. With geothermal viable in Nevada, Utah, Idaho, and parts of Texas, we'll see training infrastructure follow power instead of fiber. That redistributes the $500B data center buildout away from constrained coastal grids. The fiber follows , it's cheaper to trench conduit than upgrade substations.

What this signals for Q1 2025 procurement cycles

Hyperscalers are already rewriting RFPs to require firm renewable power with >80% capacity factor. Microsoft's 10 GW pipeline, Meta's 2 GW Louisiana campus , they're all baking in geothermal now. The vendors who can deliver drilled, tested wells by 2026 will lock in decade-long PPAs. We're watching three drilling services companies that just raised Series B rounds specifically for this pivot. The capital allocation shifted in September.

The uncomfortable parallel to 2022's diffusion moment

When Stable Diffusion open-sourced, the industry spent months debating artist compensation while the real shift , generative video, 3D, protein folding , happened in the open weights ecosystem. Same pattern here: we're litigating GLM-4-9B's cyber risk while Chinese labs iterate on 32B and 72B variants trained on the same infrastructure. The capability ceiling moves every 90 days. Policy responds in years. That gap is where the actual risk lives.