Google's new chips are a shot at Nvidia — and a big hint at where AI goes next

Google unveiled two new AI chips for the first time split into separate training and inference lines: TPU 8t for frontier model training and TPU 8i for inference, both due later this year. The launch targets the growing inference market and could help Google chip away at Nvidia's dominance, while also supporting power efficiency and higher HBM capacity to address the "memory wall." Morgan Stanley estimated that sales of 500,000 TPU chips could add about $13 billion to Google's revenue in 2027.

Analysis

GOOGL is moving from a single-chip narrative to a productized silicon portfolio, which matters because it turns TPU into a margin lever rather than just a cost offset. If inference becomes the larger and faster-growing workload, Google can monetize the most durable part of the AI stack while also lowering its own compute bill, creating a double benefit that the market has not fully capitalized into cloud/AI estimates yet.

The second-order winner is likely Apple as a TPU customer if Google’s inference economics are materially better than buying scarce NVIDIA capacity for always-on workloads. More importantly, this broadens TPU adoption by improving framework compatibility, which can reduce customer switching friction and increase the probability that TPU becomes a default option for agent workloads over the next 12-24 months. That would pressure NVIDIA’s share not necessarily on peak-performance training, but on the much larger installed base of inference spend.

For NVDA, the threat is not immediate demand collapse; it is incremental share leakage in cloud inference where buying decisions are increasingly made on cost-per-token, power efficiency, and memory bandwidth rather than raw FLOPS. The market may underappreciate how quickly hyperscalers can reallocate workloads once software portability improves, especially if agents drive sustained inference utilization. The main risk to the bearish NVDA view is that frontier model training remains highly GPU-intensive longer than expected, delaying any meaningful erosion in NVIDIA’s growth trajectory.

The key contrarian point is that this is as much about Google Cloud differentiation as it is about semiconductors. If TPU 8i lands well, GOOGL can use custom silicon to improve cloud gross margins and win AI workloads from enterprises that want predictable inference pricing, while NVIDIA still benefits from an expanding total addressable market. The market may be overestimating winner-take-all dynamics and underestimating a more balanced outcome where hyperscalers capture more economic rent from inference without fully displacing NVIDIA.

Key Decisions for Investors

• Long GOOGL vs short NVDA in a 3-6 month pair trade: thesis is incremental cloud inference share capture and margin expansion at Google versus slower multiple support for NVIDIA if cloud customers start rebalancing spend. Use a 1:1 dollar-neutral structure; add on weakness in NVDA after cloud capex headlines.
• Buy GOOGL call spreads 6-12 months out to express TPU monetization upside with defined risk. Favor strikes around 10-15% above spot to capture a rerating if TPU adoption becomes visible in cloud revenue and margins.
• Trim or hedge NVDA exposure into any rally driven by training-cycle headlines; maintain core long if you believe training remains dominant, but size down by 20-30% as a hedge against inference share loss over the next 2-4 quarters.
• Watch AAPL as a secondary beneficiary: initiate a tactical long only if TPU adoption starts showing up in enterprise AI workloads, since Apple can benefit from cheaper external AI infrastructure without needing to own the full stack. Time horizon: 6-12 months.
• Avoid AMZN/MSFT directional shorts here; instead, monitor for evidence that cloud inference pricing is compressing. If Google proves it can offer materially better cost-per-token, then a relative long GOOGL / short basket of AMZN-MSFT cloud proxies becomes attractive over 6-9 months.