Google's Cloud TPUs, hardware accelerators custom-built to speed up TensorFlow ML workloads, are now available in beta, starting at $6.50 per Cloud TPU per hour
By John Barrus, Product Manager for Cloud TPUs, Google Cloud and Zak Stone, Product Manager for TensorFlow and Cloud TPUs, Google Brain Team
Context & Ripple Effects
Google is converting an internal research program into a commercial product. Nine months earlier, the TensorFlow Research Cloud put 1,000 Cloud TPUs into researchers' hands for free, while the second-generation TPU chips promised up to 180 teraflops on Google Compute Engine. Today's beta at $6.50 per TPU per hour is the first time anyone outside Google can simply rent that silicon.
First-order effects
- TensorFlow users gain on-demand access to Google's custom accelerators without buying hardware, with pricing set at $6.50 per Cloud TPU per hour.
- Google starts selling its own ML silicon as a cloud service, turning what was an internal cost center for its Brain team into billable Google Cloud revenue.
Second-order effects
- Rival clouds face pressure to match a first-party accelerator offering rather than reselling commodity GPUs, since Google controls both the chip design and the framework it accelerates.
- The hourly price point gives enterprises a concrete benchmark for accelerator costs, anchoring how buyers compare ML compute across providers.
Third-order effects
- If the cadence holds — from this beta through the fifth-generation TPUs' performance-per-dollar gains to the eighth-generation split between training and inference SKUs — custom accelerators harden into a permanent cloud product line with distinct tiers, not a niche experiment.
- Cloud competition shifts toward who designs their own silicon, because owning the chip-to-framework stack lets a provider tune both cost and performance in ways GPU resellers cannot.
The trend: Custom AI accelerators are becoming rentable cloud infrastructure, with each TPU generation widening the gap between first-party silicon vendors and clouds that only host commodity GPUs.