Google Cloud will finally add GPU services in early 2017
Dan Richman / GeekWire :
Context & Ripple Effects
By late 2016, Google Cloud Platform had been rounding out its enterprise stack piece by piece — three months earlier it took Cloud SQL, Bigtable, and Datastore out of beta — but its compute menu still lacked the one ingredient machine-learning customers were asking for: GPUs. Committing to GPU services for early 2017 closes that gap and puts Google on equal footing for training workloads rather than just serving and storage.
What makes the timing notable is what Google was building alongside it: within months it would argue its own TPUs beat GPU/CPU combinations by 15x to 30x on ML tasks (its April 2017 performance claims), so the GPU offering functions as both a competitive necessity and a foil for Google's custom silicon story.
First-order effects
- Machine-learning teams that had to run training jobs elsewhere can now keep their data and models on Google Cloud, removing the main reason to split workloads across providers.
Second-order effects
- GPU availability forces Google to sharpen its differentiation argument around TPUs — the 15x-to-30x speedup claims arrive three months after this announcement — while price becomes the lever on the GPU side itself, later formalized in Preemptible GPUs at a 50% discount.
Third-order effects
- This is the moment accelerator capacity becomes a first-class cloud product rather than a niche add-on — the lineage runs straight to today's H100-powered A3 supercomputer VMs, by which point GPU/accelerator supply underpins Google Cloud's growth and contracted backlog.
The trend: Cloud platforms are evolving from general-purpose compute vendors into AI compute utilities, with accelerator supply — bought, custom-built, or both — becoming the core product.