Alphabet (GOOGL.US) shares have surged approximately 31% in Q4, making it the tenth-best performer in the S&P 500. A key driver behind this rally is the growing attention on Google's in-house Tensor Processing Units (TPUs) and the impressive capabilities of its AI model Gemini 3, trained using these chips. Investors are increasingly optimistic that Google could unlock a near-trillion-dollar revenue stream by selling TPUs to third parties.

Google has been developing TPUs for nearly a decade. In April, the company unveiled its seventh-generation TPU, "Ironwood," which delivers twice the performance and energy efficiency of its predecessor, Trillium. Ironwood TPUs were used to train Gemini 3, currently the world's leading multimodal and reasoning model, excelling in AI tasks like agent-based programming.

While most large language models and AI workloads still rely on NVIDIA's (NVDA.US) GPUs, industry interest in Google's TPUs is growing. Reports suggest Meta (META.US) may spend billions on Google TPUs starting in 2027, while Anthropic has committed to using up to 1 million TPUs from 2026. Gemini 3's dominance in benchmarks—scoring 1501 on LMArena and 37.5% on Humanity's Last Exam—highlights TPUs' competitive edge over NVIDIA in AI computing.

Google's TPUs, as application-specific integrated circuits (ASICs), may lack the versatility of NVIDIA GPUs but offer advantages in cost, performance, and efficiency. Mark Iong of Homestead Advisers notes, "NVIDIA chips are pricier and harder to obtain. Google's TPUs present a compelling alternative, and the company leads in this space." However, it remains unclear how aggressively Google will pursue third-party sales.

DA Davidson's Gil Luria estimates that if Google captures 20% of the AI chip market, it could generate around $900 billion in revenue. Morgan Stanley's Brian Nowak sees early signs of a TPU sales strategy, projecting 5 million TPUs sold in 2027 (up 67% from prior estimates) and 7 million in 2028 (up 120%). Nowak's analysis suggests that selling 500,000 TPUs externally could boost Google Cloud revenue by $13 billion (11%) and EPS by 3% ($0.37) in 2027.

Despite Google's elevated valuation—trading at 27x forward earnings, its highest since 2021—analysts like Allen Bond of Jensen Investment Management remain bullish, citing TPUs as a credible revenue driver. "Google's AI progress is tangible, and its valuation remains reasonable given growth expectations," Bond said.

The debate over whether Google's TPUs threaten NVIDIA's dominance has sparked mixed reactions. While some analysts dismissed Meta's reported TPU purchase as a "zero-sum fallacy," NVIDIA countered claims of disruption, asserting its industry-leading position. Bank of America's Vivek Arya argues NVIDIA's GPUs remain irreplaceable, powering most LLMs across cloud platforms with unmatched efficiency.

Futurum Group's Daniel Newman emphasizes the trillion-dollar AI infrastructure market can accommodate multiple players, including AMD (AMD.US), Amazon (AMZN.US), and Qualcomm (QCOM.US). Wedbush and Mizuho analysts echo this view, noting NVIDIA's entrenched lead in AI hardware and software, bolstered by its CUDA platform's high barriers to entry.

In summary, Google's TPUs represent a significant growth opportunity, but NVIDIA's dominance in AI chips appears secure—for now.