Google's TPU Fine-Tuning Stack Gets Easier, Tightening the Keras-JAX-Gemma Flywheel

4. Google's TPU Fine-Tuning Stack Gets Easier, Tightening the Keras-JAX-Gemma Flywheel

François Chollet, the creator of Keras and a principal figure at Google DeepMind, published a tutorial demonstrating fine-tuning of Gemma on TPU v5 hardware using a stack combining Kinetic, Keras, and JAX. The post, shared directly from Chollet's account, frames the combination as the most accessible path to fully leveraging TPUs at scale, a claim that carries weight given his authorship of Keras and his proximity to Google's infrastructure decisions.

The significance here is competitive lock-in. Google's TPU hardware has historically been difficult to utilize efficiently without deep infrastructure expertise, giving NVIDIA's GPU ecosystem a usability advantage even when TPU raw performance was comparable or superior. By lowering the friction through an opinionated, well-documented stack, Google makes it easier for researchers and fine-tuning practitioners to stay inside the Google Cloud ecosystem rather than defaulting to CUDA-based workflows on AWS or Azure. Gemma model developers and open-source practitioners who want enterprise-scale TPU throughput are the clear beneficiaries. NVIDIA and the broader CUDA tooling ecosystem are the relative losers if this friction reduction actually holds at production scale.

This is also a signal of Keras's continued strategic repositioning. Since Keras 3 relaunched as a true multi-backend framework, Google has been quietly reinforcing the Keras-plus-JAX path as the canonical on-ramp to its own hardware. Tutorials like this one are infrastructure marketing: they establish default workflows that compound over time, as teams that learn Gemma on TPUs via Keras are unlikely to re-platform. The Kinetic addition is worth watching as a potential new layer in that stack.

Source: https://twitter.com/fchollet/status/2040822483511841057