Google TPUs delivered Gemini 3 model challenging NVIDIA

NVIDIA now has a frontier model competitor with very deep pockets with Google making a real play for TPU general availability

Google’s latest TPU generation has created a hardware foundation that positions the Gemini 3 family as a direct challenger to NVIDIA’s top accelerators in both frontier-scale training and high-volume inference [1]. Ironwood TPUs introduce higher-density systolic arrays, larger unified memory pools, and a low-latency optical interconnect that allows very large models to maintain throughput during long-sequence reasoning workloads that traditionally strain GPU-based systems. Gemini 3 benefits from Google’s vertically integrated stack, which aligns model architecture, compiler optimizations, and TPU kernels within a single design space. This alignment reduces orchestration inefficiencies that arise when models must adapt to heterogeneous GPU environments. NVIDIA still holds a dominant market position through CUDA, its extensive developer ecosystem, and its global installation footprint, but the combination of Gemini 3 and general availability of next-generation TPUs now gives hyperscalers a credible alternative that changes the competitive equilibrium and shifts the long-term economics of scaling frontier models [2]. This makes NVQLink an even more important play for NVIDIA as these TPUs represent a legitimate threat to future infrastructure spending. Maybe we will see quantum systems mixed with NVQLink help NVIDIA defend against a full replacement cycle of older datacenter infrastructure by making existing GPU systems more capable, more efficient, and more scalable without requiring customers to adopt a new hardware ecosystem such as Google’s TPUs

Footnotes:
[1] Google. (2025, November 18). A new era of intelligence with Gemini 3: A note from our CEO. https://blog.google/products/gemini/gemini-3/#note-from-ceo
[2] Google Cloud. (2025, November 6). Announcing Ironwood TPUs general availability and new Axion-based VMs to power the age of inference. https://cloud.google.com/blog/products/compute/ironwood-tpus-and-new-axion-based-vms-for-your-ai-workloads