Gemini 3 — and the custom chips that power it — is a wake up call for AI investors (www.cnbc.com)

Google’s surprise launch of Gemini 3 — a reasoning-focused large language model reportedly trained entirely on Google’s custom TPUs co-designed with Broadcom — has rattled markets and reignited debate over whether bespoke silicon can erode NVIDIA’s GPU dominance. The news, plus a report that Meta is considering TPUs for its data centers by 2027, knocked NVIDIA shares lower and lifted Alphabet, Broadcom and Meta chatter among investors. Gemini 3’s performance leapfrogged current OpenAI offerings and highlights that major cloud and AI players can justify huge upfront R&D and manufacturing costs to achieve large, recurring cost savings on at-scale LLM training and inference. Technically, TPUs are ASICs optimized for specific ML workloads and can be far more efficient and cheaper for massive internal use than general-purpose GPUs. But they carry trade-offs: vendor lock‑in, the need to port years of CUDA-based code, limited availability across clouds (no TPU equivalents on AWS/Azure), and huge development investment. That preserves NVIDIA’s edge for rented cloud compute and flexible, cross-platform workloads thanks to CUDA, broad adoption, and software ecosystem. For the AI/ML community this means more heterogeneity — specialized stacks for hyperscalers, GPUs for broad developer portability — and continued competition between model and hardware suppliers. Investors should note the strategic implications, but it’s too early to declare a paradigm shift away from NVIDIA.