How Google put together the pieces for its AI comeback (www.cnbc.com)

Google has quietly stitched together hardware, models and cloud distribution to stage an AI comeback: this month’s debut of Gemini 3 and the Ironwood TPU (Google’s 7th‑gen ASIC) have analysts excited. Gemini 3 — faster, “less prompting” and more accurate than prior releases, plus new image features (Nano Banana/Nano Banana Pro) — was rolled out quickly after Gemini 2.5 and is being embedded across consumer apps and enterprise services. Ironwood, which Google says is up to ~30× more power‑efficient than its 2018 TPU, lets customers run and scale very large, data‑intensive models and underpins recent multi‑billion dollar deals. Those moves helped Google Cloud deliver a strong quarter (first $100B quarter for Alphabet) and lifted Alphabet shares amid heavy market cap interest. The significance is twofold: vertical integration (models optimized for Google’s TPUs and trained on massive YouTube/video/text datasets) gives Google a practical edge in image/video generation and enterprise deployment, while its cloud+ASIC stack makes it a credible challenger to Nvidia’s GPU dominance. But experts caution the field remains fiercely contested — OpenAI, Anthropic and others continue rapid model updates (GPT‑5 tweaks, Opus 4.5), Nvidia still controls >90% of AI chips, and scaling costs are huge (companies projecting collective capex >$380B). In short, Google has closed important gaps, but sustained leadership will depend on capacity expansion, continued model quality improvements and winning large enterprise workloads.