Google claims its 4th-gen TPU-based supercomputers, used for AI training, are up to 1.7x faster and 1.9x more power-efficient than Nvidia's A100 systems
Alphabet Inc's (GOOGL.O) Google on Tuesday released new details about the supercomputers it uses to train its artificial intelligence models …
Context & Ripple Effects
Google has been benchmarking its TPUs against merchant GPUs since 2017, when it claimed its first-generation chips were 15x to 30x faster than a standard GPU/CPU combination with far better performance-per-watt. The new disclosure extends that argument to the training-supercomputer scale — the exact workload the AI boom has made Nvidia's most valuable market.
What makes the claim strategically loaded is Google's own behavior in the same window: weeks after publishing these TPU numbers, Google Cloud launched an A3 supercomputer VM built on Nvidia's H100 GPUs, and by August it promised the 5th-generation TPUs would double training performance per dollar versus this 4th-gen generation. Google is simultaneously attacking Nvidia's benchmarks and reselling its chips.
First-order effects
- Google Cloud gains a marketing weapon against A100-based cloud rivals: published numbers arguing its in-house silicon trains models up to 1.7x faster at 1.9x better energy efficiency, aimed directly at customers choosing where to run large training jobs.
- Nvidia faces a first-party challenger that controls both the chip design and the cloud it runs on — a competitor whose benchmark claims don't depend on buying Nvidia inventory.
Second-order effects
- Google's dual-track hardware strategy — selling H100-based A3 VMs while promoting TPUs — pressures Nvidia to keep its cloud partners convinced that merchant GPUs, not custom silicon, remain the safer bet for third-party clouds.
- Efficiency-per-dollar claims from Google set a pricing benchmark that other clouds renting A100 capacity must answer, squeezing the margins of GPU rental as the default training option.
Third-order effects
- If each TPU generation keeps compounding the per-dollar gains — as the 5th-gen promise and the later 8th-generation TPU launch suggest — the long-run structure is hyperscalers vertically integrating their AI stacks and treating merchant GPUs as a hedge rather than the foundation.
- The 1.9x power-efficiency figure points to energy becoming the binding constraint on AI training scale, favoring whoever owns silicon optimized for performance-per-watt over whoever assembles the fastest discrete accelerators.
The trend: Hyperscalers are shifting AI training infrastructure from merchant GPUs toward vertically integrated custom silicon, with performance-per-watt and per-dollar as the competitive axis.