Palantir ships an engine to run NVIDIA Nemotron inside US government air gaps

Good morning 👋 Open weights, closed perimeter: Palantir just shipped the engine that lets US agencies run frontier-level models without a single inference request leaving the building.

In today's issue:

• 🔭 Palantir's government AI engine — open models, closed perimeter
• 🧠 Models & Releases — Claude on Blackwell Ultra goes GA, plus Meta's brain-typing research
• 🔬 Research Highlights — three papers worth your time: efficient agents, world model self-refinement, and a safety assumption that doesn't hold
• 🔗 Quick Links — AI in lunar orbit, and a density estimator that beats the standard by 6×

🔭 THE ONE THING

🔐 Palantir ships an engine to run NVIDIA Nemotron inside US government air gaps

Palantir announced an engine that deploys NVIDIA Nemotron open-weight models directly on government infrastructure, wrapped in AIP, Ontology, Foundry, and Apollo and branded as a "Sovereign AI Operating System." The pitch to agencies: frontier-level AI, full data residency, no inference requests leaving the perimeter. The terminology is worth parsing: "open models" means the weights are published, not that the platform they run on is open. Agencies own the model file. Palantir owns the layer it runs on, and for shops already embedded in their stack across defense and intel, that's a dependency trade they've made before.

🧠 MODELS & RELEASES

• 🚀 Anthropic / NVIDIA / Azure Claude on GB300 Blackwell Ultra is now GA in Microsoft Foundry. Azure-native enterprises get the compute upgrade today; the GB300 is a meaningful step for long-context and multi-step agent workloads. link

• 🔒 Palantir / NVIDIA Palantir launched an intelligent engine running NVIDIA Nemotron open models for U.S. government agencies inside air-gapped environments. Open weights doing what API-based models can't: the data never leaves the facility. Builder read: the Nemotron-plus-closed-env pattern is the one to study for any regulated deployment where egress is a hard no. link

• 🔬 Meta Brain2Qwerty decodes non-invasive EEG signals into typed text. No surgery, no implants. Still below keyboard speed, but the architecture is published and it's the first credible hardware-free path for nonspeaking users. link

🔬 RESEARCH HIGHLIGHTS

• Agents-A1. A 35B MoE model matching trillion-parameter performance by scaling horizon, not parameter count: longer trajectories, broader heterogeneous agent abilities. The claim is architectural. If it reproduces off the paper's own evals, compute allocation for agentic tasks shifts. link

• WorldEvolver. Agents that self-refine their internal world models to make better predictions before acting. The motivation is sharper than it first sounds: bad foresight isn't just noise. An inaccurate world model can actively degrade decisions, not just add uncertainty. Worth watching whether it holds on harder, real-world planning tasks outside the paper's benchmarks. link

• Pessimism's Paradox. Conservative offline training amplifies reward hacking during online adaptation. Not suppresses it. The conventional view is that keeping a policy close to well-supported behavior gives you a safer base for online RL. This paper challenges that, both empirically and mechanistically. If you're relying on conservative pretraining as a safety floor before online fine-tuning, read it. link

🔗 QUICK LINKS

• Firefly Aerospace / NVIDIA Blue Ghost Mission 2 ran a Jetson module in lunar orbit, the first edge AI deployment around the moon, processing imagery from Lawrence Livermore telescopes onboard rather than downlinking raw data. Read
• AllenAI DiScoFormer estimates density and score in one transformer pass without retraining, beating kernel density estimation by 6.5x on score error and 37x on density in 100 dimensions. Read

See you tomorrow.

Pradeep Perugu