Google Cloud launches 8th-generation TPUs targeting Nvidia in AI chip market
Two chips cover training and inference workloads, with Google claiming speed and cost gains over prior generation.
Briefing
- •Google Cloud unveiled two eighth-generation TPUs designed for AI training and inference, its most direct challenge yet to Nvidia's dominance in cloud AI compute.
- •The new chips are faster and cheaper than their predecessors, according to Google, though specific performance benchmarks were not disclosed in available sources.
- •The dual-chip lineup is framed around the 'agentic era', suggesting optimisation for autonomous AI workloads that require both heavy training runs and rapid inference.
- •Google continues to offer Nvidia GPUs on its cloud platform alongside the new TPUs, indicating a parallel rather than replacement strategy for third-party silicon.
Why it matters
- →Google's eighth-generation TPU launch directly targets Nvidia's cloud AI compute revenue by offering a cheaper alternative on Google Cloud. However, Google simultaneously continues to sell Nvidia GPUs on the same platform, meaning the primary effect is margin pressure on Google's GPU-related cloud revenue mix rather than a displacement of Nvidia's hyperscaler orders. AWS and Azure remain Nvidia's dominant cloud distribution channels; Google Cloud's TPU push reduces Nvidia's share within GCP specifically, not across the hyperscaler complex.
- →Apple's leadership transition to a hardware engineer CEO, noted in the Tim Cook succession story, raises the probability that Apple accelerates its internal silicon roadmap for AI workloads. If Apple's custom AI chips advance faster, Google faces a two-front competitive response: TPUs competing against Nvidia in cloud, and Apple Silicon potentially reducing demand for both Nvidia GPUs and Google TPUs in on-device AI inference. This compounds Google's need to convert TPU economics into developer adoption before Apple's inference silicon locks in the edge compute layer.
- →Victory Giant's 50% IPO debut confirmed strong investor appetite for Nvidia supply chain exposure. Google's TPU challenge, if it gains traction at GCP, reduces Nvidia's long-term order trajectory for AI training chips and creates a secondary risk for PCB and substrate suppliers concentrated in the Nvidia customer book. The magnitude depends on adoption speed, but investors in Nvidia-adjacent hardware names should begin discounting a scenario where hyperscaler in-house silicon captures a larger fraction of training workloads over a 3-5 year horizon.
Historical Context
- 2016-2018
Google first deployed TPUs internally for its own workloads before offering them externally on Google Cloud. The internal deployment reduced Google's own Nvidia GPU consumption but did not materially dent Nvidia's hyperscaler revenue because AWS and Azure continued large-scale GPU purchases. The parallel strategy then mirrors the current dual-offering approach.
- 2023
AWS Trainium and Inferentia chips launched as Nvidia alternatives on AWS. Adoption remained limited relative to Nvidia GPU instances because developer tooling and software ecosystems favored CUDA. Google's TPU faces the same ecosystem lock-in problem, as most AI training frameworks are CUDA-optimized, constraining switching speed.
Connected Events
Victory Giant surges 50-60% in Hong Kong debut after $2.6bn IPO
Victory Giant's 50% Hong Kong IPO debut, driven by its status as a Nvidia PCB supplier, shows how concentrated investor exposure to the Nvidia supply chain has become. A credible TPU alternative gaining hyperscaler share introduces a valuation risk for that supply chain premium that the IPO pricing did not reflect.
Tim Cook steps down as Apple CEO after 15 years; John Ternus to succeed him
Tim Cook's departure and John Ternus's appointment as Apple CEO raised immediate questions about Apple's AI silicon acceleration. A hardware-focused CEO pursuing faster on-device AI inference directly competes with both Nvidia GPU deployments and Google TPU cloud inference for agentic AI workloads.
