Google unveils two new TPUs designed for the "agentic era"

Google unveiled its eighth-generation TPU lineup, split into TPU 8t for training and TPU 8i for inference, positioning the chips for the AI "agent era." The TPU 8t pods reportedly contain 9,600 chips, 2 petabytes of shared high-bandwidth memory, and up to 121 FP4 EFlops per pod, nearly 3x Ironwood's training compute ceiling. The announcement reinforces Google's custom-chip strategy versus Nvidia dependency, but it is primarily a product/technology update with limited near-term market impact.

Analysis

Google is signaling a strategic shift from selling “more compute” to selling a differentiated AI stack optimized for throughput economics. That matters because the bottleneck for frontier AI is increasingly not raw FLOPs but delivered training time, memory bandwidth, and power efficiency; if Google can credibly compress time-to-train, it can pull workloads away from Nvidia-centric stacks in the highest-value workloads first. The first-order beneficiary is GOOGL, but the second-order winner is likely Google Cloud’s attach rate: custom silicon is sticky, and once a customer builds around TPU economics, migration costs rise sharply.

For NVDA, this is not an immediate demand cliff, but it is a margin and mix risk over 12-24 months. Hyperscalers do not need to displace Nvidia broadly to matter; they only need to carve out the most cost-sensitive frontier training and inference workloads to slow share gains and weaken pricing power at the top end. The market may still be underestimating how quickly internal chips move from “experimental” to “default” when the ROI is measured in weeks saved per model cycle.

The more interesting second-order effect is supply chain inflation shifting from accelerators into memory and advanced packaging. A TPU architecture built around very large shared memory pools should keep pressure on HBM, interconnect, and foundry capacity even if Nvidia unit share stalls, so semicap and memory bottlenecks remain intact. The contrarian view is that this is bullish for GOOGL but not necessarily bearish enough for NVDA yet: Nvidia’s software moat and ecosystem still dominate most enterprise deployments, and the real risk is a slower growth rate, not a collapse.

Catalyst timing matters: near term, this is mostly a narrative and valuation event; the operating data that matters will arrive over the next 2-3 quarters via cloud margin mix, TPU utilization, and customer adoption disclosures. If Google starts showing higher cloud gross margin with faster AI revenue growth, the market will likely re-rate GOOGL as a hardware-enabled platform rather than a software-adjacent cloud laggard.

Key Decisions for Investors

• Long GOOGL vs. NVDA as a 6-12 month pair trade: Google’s custom silicon should support cloud margin expansion and AI workload retention, while NVDA faces incremental pricing-pressure risk; target a 10-15% relative return with a stop if Nvidia re-accelerates enterprise share gains.
• Buy GOOGL calls or call spreads into the next earnings cycle: the market is likely to reward any evidence of TPU-driven cloud margin inflection; favor 3-6 month maturities to capture a rerating before adoption data fully shows up.
• Trim/hedge NVDA upside via short-dated call spreads rather than outright shorts: the near-term downside is limited by AI demand still outrunning supply, but a modestly negative mix shift could cap multiple expansion over the next 1-2 quarters.
• Overweight HBM and advanced packaging exposure on any weakness: TPU scale-up keeps memory and packaging demand elevated even if accelerator share broadens; this is the cleaner second-order trade than shorting hardware outright.