Keras Kinetic Brings Decorator-Based TPU Access to Python Functions, Threatening Modal's Core Differentiator

7. Keras Kinetic Brings Decorator-Based TPU Access to Python Functions, Threatening Modal's Core Differentiator

Keras announced Keras Kinetic at its community call, a new library that lets developers execute code on cloud TPUs and GPUs by attaching a Python decorator to a function. The mechanics mirror what Modal has built its business around: write local Python, decorate the function, and the infrastructure layer handles remote execution, environment setup, and hardware provisioning automatically. The critical distinction flagged by Keras creator François Chollet is TPU support, a hardware class that Modal does not currently serve and that Google controls the supply of almost exclusively through its cloud infrastructure.

This matters because Modal has positioned itself as the frictionless compute layer for ML engineers who want to avoid raw cloud configuration, and it has accumulated significant developer mindshare among exactly the Keras user base that Kinetic now directly targets. Google, through the Keras project, is effectively verticalizing: it owns the TPU hardware, the ML framework, and now the deployment abstraction layer sitting on top of both. Independent tooling companies like Modal, Banana, and RunPod compete on GPU access and developer experience, but none can offer TPU execution at all, which means Kinetic enters the market with a capability gap competitors cannot close without negotiating hardware access from Google itself. Keras users who previously chose Modal for its simplicity now have a first-party alternative that connects directly to the hardware their models were likely already tuned to run on.

The broader signal here is that ML framework maintainers are expanding their surface area downward into infrastructure tooling, a pattern already visible in Hugging Face's push into model serving and endpoint management. Google's play with Kinetic follows the same logic: the framework layer is only defensible long-term if it also owns the execution layer, because developers who can run elsewhere will eventually standardize on whichever stack reduces the most friction end to end. Kinetic, if it ships with solid reliability, compresses the ML development loop inside Google's ecosystem in a way that pure hardware offerings never could.

Source: https://twitter.com/fchollet/status/2040119594984284218