Google challenges Nvidia with new chips to speed up AI

Google is set to unveil a new generation of TPUs focused on AI inference, strengthening its challenge to Nvidia in a fast-growing semiconductor segment. Demand for TPUs is already rising, with Anthropic expanding to as many as 1 million TPUs, Meta signing a multibillion-dollar cloud deal, and Broadcom-linked capacity planned at about 3.5 gigawatts starting in 2027. The news supports Google’s AI platform and cloud strategy, though it also raises supply-allocation and execution risks as the market shifts from training to inference.

Analysis

The key shift is that AI compute is moving from a capex race in training to a throughput race in inference, and that is structurally worse for the GPU monopoly narrative. Inference is more latency-sensitive, more price-elastic, and more routable across architectures, which means customers will increasingly benchmark on $/token and response time rather than raw FLOPS. That creates a multi-vendor market where Google’s vertically integrated stack can win share even if it never displaces Nvidia in frontier training.

The second-order effect is that Google is not just selling chips; it is weaponizing application insight from Search, Gemini, and cloud telemetry into silicon. That feedback loop can compress product cycles on the software side, but the harder-to-spot impact is on customer lock-in: once models are optimized for TPU pods, networking, and scheduling, switching costs rise materially. The clearest beneficiary is AVGO, which sits behind the custom silicon/networking buildout regardless of which model layer wins.

For NVDA, the near-term risk is not a sudden demand collapse but a margin mix problem as hyperscalers reallocate incremental inference workloads away from general-purpose GPUs. That pressure can show up over 2-4 quarters first in cloud procurement behavior, then in pricing concessions, then in investor expectations for growth durability. The most vulnerable segment is not AI training spend, but inference-heavy agents and consumer query workloads where cost per request matters most.

Contrarian read: the market may be underestimating how supply-constrained Google’s TPU franchise still is. If capacity is the bottleneck, the near-term winner may not be Google’s revenue but its strategic position as customers accept multi-year commitments to secure access. That argues for selective exposure to the ecosystem rather than a blanket short on Nvidia; the trade is a relative-value rotation, not a full replacement thesis.

Key Decisions for Investors

• Long GOOGL / short NVDA pair trade over 3-6 months: thesis is inference share gains and TPU adoption reduce incremental GPU attach in cloud workloads; target 10-15% relative outperformance if Google sustains customer conversion.
• Add AVGO on pullbacks for 6-12 months: Broadcom benefits from both TPU-related networking and custom silicon content regardless of which AI model architecture wins; better risk/reward than pure-application names.
• Trim NVDA into strength, especially on rallies tied to training spend headlines; use 3-6 month horizon and watch for inference-focused customer commentary from hyperscalers as the catalyst for multiple compression.
• Long META vs short IT (or keep IT underweight) over 6-9 months: META is an early TPU testing beneficiary with optionality on inference cost reduction, while IT faces slower fundamental pass-through from AI infrastructure shifts.
• If available, buy GOOGL call spreads 6-9 months out rather than outright calls: upside is tied to TPU adoption and cloud monetization, while downside is cushioned by core ad/search cash flow.