Home / Technology / Beyond Nvidia: Google's Custom AI Silicon Revealed
Beyond Nvidia: Google's Custom AI Silicon Revealed
22 Apr
Summary
- Google announced two new custom AI chips for training and inference.
- These chips aim to reduce costs and improve efficiency for AI workloads.
- Google controls its AI stack end-to-end, a key competitive advantage.
In a strategic move to reduce reliance on third-party hardware, Google unveiled its eighth-generation Tensor Processing Units (TPUs) in late 2026. These custom silicon designs, shipping later in the year, are bifurcated into two specialized chips: TPU 8t for frontier model training and TPU 8i for agentic inference and real-time sampling.
This strategic split, decided upon in 2024, predates the industry's broader pivot to reasoning models and agents. Google's vertical integration of its AI stack, from energy to services, is highlighted as a key differentiator, enabling cost-per-token economics that rivals reportedly cannot match.
The TPU 8t offers a substantial generational leap in training performance, boasting increased EFlops, bandwidth, and networking speeds, with scalability potentially exceeding one million chips. The TPU 8i introduces architectural innovations, including a network topology optimized for low latency, crucial for agentic workloads, delivering a claimed five-fold improvement in real-time LLM sampling.
This development reframes cloud evaluations for enterprise buyers in 2026-2027. Teams focused on large-scale training will assess 8t availability and networking, while those serving agents will scrutinize 8i performance and HBM capacity. Google's self-reported benchmarks suggest significant gains, though independent evaluations are anticipated in the coming quarters.
