Read that Google Cloud has reportedly been turning down customer deals because it doesn't have enough compute internally, and I immediately wanted to know what they're actually building to fix it. Turns out it's a chip codenamed Frozen V2, and the backstory is genuinely interesting. An earlier version of this idea, led by DeepMind's chief scientist, Jeff Dean, wanted to burn Gemini's actual model weights directly into the silicon. Got scrapped, because the chip would've become useless the moment Gemini updated to a new version. Frozen V2 fixes that by hardwiring only the architecture instead, the stable underlying blueprint, while leaving the weights updatable through normal loading. The reported result is 6 to 10x more tokens processed per watt compared to Google's current TPUs! Targeted for deployment as early as 2028, running alongside the existing TPU lineup rather than replacing it. Feels like the clearest sign yet that the biggest labs are starting to treat the model and the chip as one design problem instead of two separate ones!