Sovereign Escape Velocity: Ownership w Open Models — Gus Martins, & Ian Ballantyne, Google DeepMind

TL;DR

• 1

  Gemma 4 is optimized for ownership, not just leaderboard bragging rights: Gus Martins frames Gemini as Google's strongest hosted model, but says Gemma exists for cases where you need to run locally, customize heavily, or keep proprietary data inside your own infrastructure.

• 2

  The small models are built for phones and edge devices: Gemma 4 includes E2B and E4B models with text, vision, and audio input plus text output, and Gus says they can do thinking, coding, and function calling directly on a Pixel phone or similar hardware.

• 3

  The larger models punch above their weight on hardware cost: The 26B mixture-of-experts model activates roughly 4B parameters at a time, and the 31B dense model can run on one GPU, compared with competitors that Gus says may need roughly 200 GB of memory across four or five GPUs.

• 4

  Apache 2.0 is a strategic shift for enterprise and sovereign users: DeepMind moved away from a custom Gemma license because custom terms can trigger long procurement reviews, while Apache 2.0 makes it easier for legal teams and public institutions to approve deployment.

• 5

  Sovereignty already has real examples, not just theory: Gus cites Ukraine using Gemma in public services, a Bulgarian national LLM based on Gemma 2, and a Brazilian Portuguese model based on Gemma 3, while noting fine-tuning is getting harder because the base multilingual performance is already strong.

• 6

  Agentic workloads change the economics in favor of local models: Ian Ballantyne argues that coding and other high-token tasks become more attractive to run on owned hardware, where the tradeoff shifts from API token spend to GPU, NPU, and energy costs.