Google will pay SpaceX $920M a month for AI compute: at a rival's data centers

Good morning 👋 The AI race just turned into a real-estate race. Google is about to pay SpaceX nearly a billion dollars a month to rent compute sitting inside a rival's data centers.

In today's issue:

• 🔭 Google to pay SpaceX ~$920M/month for compute at xAI data centers
• 🧠 ChatGPT gets a new memory, Nemotron 3 Ultra hits AWS, GPT-Rosalind levels up
• 🔬 NVIDIA's physical-AI "agent skills" at CVPR
• 🏛️ OpenAI floats a federal framework for frontier AI
• 🛠️ Pressure-test your AI infra dependency + 5 trending tools
• 🔗 One more worth a click

🔭 THE ONE THING

💸 Google will pay SpaceX $920M a month for AI compute: at a rival's data centers

What happened: Google agreed to pay SpaceX about $920 million per month from October 2026 through June 2029 (roughly $32 billion over the deal) for access to approximately 110,000 NVIDIA GPUs, plus CPUs, memory, and related gear (TechCrunch, CNBC). The catch: that compute sits in data centers tied to xAI (the Grok maker SpaceX absorbed in an all-stock merger earlier this year), and a SpaceX SEC filing confirms the terms. Access ramps up at a reduced fee through September, hits full rate in October, and carries 90-day cancellation clauses after Dec 31, 2026.

Why it matters: This is the clearest sign yet of how tight frontier compute has gotten. Google, a company that designs its own TPUs, is renting ~110,000 GPUs from infrastructure owned by a direct competitor, and calling it "bridge capacity" for surging Gemini Enterprise demand. It mirrors Anthropic's ~$1.25B/month lease of SpaceX's Colossus 1 site near Memphis (Google's slice is about half that compute). The race is now as much about who controls power, GPUs, and data-center real estate as who has the best model. SpaceX has quietly become one of the biggest compute landlords in the game.

The takeaway: If you build on Gemini or Claude, your app's reliability now rides on month-to-month compute leases between giants. Keep a multi-provider fallback and watch capacity, not just price.

🧠 MODELS & RELEASES

• 🧠 OpenAI began rolling out "Dreaming" (V3), a memory overhaul that synthesizes context across years of chats in the background instead of a manual saved-list. Plus/Pro in the US first, free tier to follow, enabled by a ~5x cut in serving cost. link
• ⚙️ NVIDIA made Nemotron 3 Ultra, its 550B-param (55B active) hybrid Transformer-Mamba MoE reasoning model, one-click deployable on Amazon SageMaker JumpStart, claiming 5x faster inference and up to 30% lower cost for agentic workloads. link
• 🧬 OpenAI expanded GPT-Rosalind with stronger biological reasoning, medicinal-chemistry, and genomics capabilities aimed at life-sciences research workflows. link

🔬 RESEARCH HIGHLIGHTS

• NVIDIA at CVPR: physical-AI "agent skills." New building blocks that wrap a full workflow around robotics, autonomous-vehicle, and vision models. The bet: the bottleneck in physical AI isn't stronger models but the tooling around them. link

🏛️ POLICY & REGULATION

• OpenAI published a blueprint for democratic governance of frontier AI, proposing a U.S. federal framework for safety, resilience, and national security. Note it's a company *proposal*, not enacted policy; it signals where the labs want regulation to land. link

🛠️ TOOLS & HOW-TO

Pressure-test your AI infra dependency in 10 minutes

1. List every AI feature you ship and the exact model + provider each one calls.

2. For each, write the fallback: if that provider doubled its price or capped capacity tomorrow, what's plan B: an alternate model, self-hosting, or graceful degrade?

3. Estimate monthly token spend per feature (requests/day × avg tokens × price) and flag your top 2 cost drivers.

4. Add a kill-switch / env flag so you can swap providers per feature without a redeploy.

5. Run the prompt below on your architecture notes to surface single points of failure.

Prompt: You are a skeptical infrastructure analyst. Here is my AI stack: [paste features + providers + rough usage]. List the 3 biggest single points of failure, the cheapest mitigation for each, and the one dependency that would hurt most if its price doubled.

5 trending right now

• 🪨 caveman: a Claude Code skill that cuts ~65% of tokens by stripping prose down to terse instructions. link
• 🧠 MemPalace: an open-source, benchmark-topping memory system for AI agents, and it's free. link
• ✨ Taste-Skill: gives your AI "good taste," steering it away from generic, boring slop. link
• 🎬 HyperFrames: write HTML, render video; a rendering layer built for agents (from HeyGen). link
• 🔭 Crucix: a personal intelligence agent that watches multiple data sources and pings you when something changes. link

🔗 QUICK LINKS

• OpenAI showed how Wasmer used Codex with GPT-5.5 to build a Node.js runtime for the edge, reporting 10x–20x faster development and shipping in weeks instead of months. link

That's a wrap. Build on. — inovAIte

How was today's issue? 😍 / 🙂 / 😴