Google TPU

AI Hardware

Google's Tensor Processing Units (TPUs) are custom-designed application-specific integrated circuits (ASICs) developed internally to accelerate machine learning workloads.

Reviewed by 7wData
Share Email

On this page

Profile

Design and manufacture custom AI accelerator chips (TPUs) for training and inference, offered via Google Cloud and now directly sold to customers.

Google's Tensor Processing Units (TPUs) are custom-designed application-specific integrated circuits (ASICs) developed internally to accelerate machine learning workloads. First deployed in 2015 to power RankBrain for Search, TPUs have evolved through eight generations, with the latest, TPU 8t and TPU 8i, announced in April 2026. TPUs are the backbone of Google's own AI services—Gemini, Search, Photos, Maps—serving over 1 billion users.

In Q1 2026, Google Cloud revenue hit $20 billion, growing 63% year-over-year, driven by enterprise AI infrastructure and solutions. Alphabet's total revenue was $109.9 billion, with operating income of $39.7 billion and a headcount of 194,668. CEO Sundar Pichai announced on the Q1 2026 earnings call that Google had begun directly selling TPU chips to customers, a shift from its previous model of only leasing cloud access.

Alphabet raised its 2026 capex forecast to $180–$190 billion, up from $175–$185 billion, reflecting massive infrastructure investment. The cloud backlog nearly doubled quarter-on-quarter to over $460 billion. However, Nvidia CEO Jensen Huang claimed in November 2025 that Nvidia is 'a generation ahead' of Google's TPUs, and Alphabet continues to rely heavily on Nvidia GPUs alongside its own chips. The TPU line faces the strategic tension of competing with Nvidia while also being a customer.

Track Google TPU and 240+ vendors.

335k+ subscribers read the daily AI & data note. One email, both newsletters. Unsubscribe anytime.

Products by Google TPU

Who buys this

  • Large enterprises running large-scale AI training workloads (e.g., LLMs, recommendation systems)
  • AI-native startups and labs needing high-throughput inference for agentic AI and reasoning models
  • Google's own product teams (Search, Gemini, Photos, Maps, YouTube) using TPUs internally
  • Cloud-native organizations using Google Kubernetes Engine (GKE) with TPU-backed pods
  • Research institutions and organizations using JAX or PyTorch for frontier model development

Strengths and what to watch

Strengths

  • Deep vertical integration: TPUs are co-designed with Google's software stack (JAX, PyTorch, vLLM) and internal AI models, enabling tight optimization for Google's own massive workloads.
  • Aggressive generational performance gains: TPU 8t claims 2.7x performance-per-dollar improvement over Ironwood for large-scale training, and TPU 8i claims 80% better performance-per-dollar for inference on MoE models.
  • Massive financial commitment: Alphabet's 2026 capex of $180–$190 billion and a cloud backlog of $460 billion signal sustained investment in TPU capacity and infrastructure.

Watch for

  • Nvidia holds a commanding lead in the AI chip market; Jensen Huang publicly stated Nvidia is 'a generation ahead' of Google's TPUs, and Google remains a major Nvidia GPU customer, creating a strategic dependency.
  • Google's decision to sell TPUs directly is nascent—CFO Anat Ashkenazi said only a small percentage of revenue from direct sales will be recognized in 2026, with the bulk in 2027, so commercial traction is unproven.
  • Alphabet's cloud backlog of $460 billion is partly constrained by compute capacity, meaning TPU supply cannot keep up with demand, potentially capping growth and ceding ground to rivals.

Recent moves

Key Information

Industry
AI Hardware

Frequently Asked Questions

What is a Google TPU?

A Google TPU is a custom-designed ASIC chip that accelerates machine learning workloads. First deployed in 2015 for RankBrain, it now powers Google's AI services like Gemini and Search. The latest eighth-generation models are TPU 8t for training and TPU 8i for inference.

What are the latest Google TPU models and their performance?

Announced in April 2026, TPU 8t offers 2.7x better performance per dollar over Ironwood for large-scale training. TPU 8i delivers 80% better performance per dollar for inference on mixture-of-experts models. Both are eighth-generation accelerators designed for AI workloads.

Is Google selling TPU chips directly to customers?

Yes, in Q1 2026, CEO Sundar Pichai announced Google began directly selling TPU chips to customers, shifting from only leasing cloud access. However, CFO Anat Ashkenazi noted only a small percentage of revenue from direct sales will be recognized in 2026, with most in 2027.

How do Google TPUs compare to Nvidia GPUs?

Nvidia CEO Jensen Huang claimed in November 2025 that Nvidia is 'a generation ahead' of Google's TPUs. Alphabet remains a major Nvidia GPU customer while also developing its own chips, creating a strategic tension between competing with and relying on Nvidia.

What is driving Google Cloud's AI infrastructure growth?

Google Cloud revenue hit $20 billion in Q1 2026, growing 63% year-over-year, driven by enterprise AI infrastructure. Alphabet raised its 2026 capex forecast to $180–$190 billion, and the cloud backlog nearly doubled quarter-on-quarter to over $460 billion.

Who uses Google TPUs and for what workloads?

TPUs are used by Google's own product teams for Search, Gemini, Photos, and Maps, serving over 1 billion users. External customers include large enterprises for LLM training, AI startups for inference, and research institutions using JAX or PyTorch for frontier model development.

Sources

  1. cloud.google.com — Product portfolio, TPU 8i/8t specs, availability, and technical architecture claims
  2. www.reuters.com — Q1 2026 cloud revenue of $20B, 63% growth, direct TPU sales announcement, capex increase to $180-190B, backlog of $460B
  3. s206.q4cdn.com — Q1 2026 consolidated revenue $109.9B, operating income $39.7B, headcount 194,668, cloud segment operating income $6.6B
  4. abc.xyz — Q4 2025 annual revenues exceeded $400B, cloud run rate over $70B, 2026 capex guidance of $175-185B
  5. fortune.com — Nvidia CEO Jensen Huang's claim that Nvidia is 'a generation ahead' of Google's TPUs
  6. blog.google — CEO Sundar Pichai's Q1 2026 earnings remarks, including TPU direct sales and cloud backlog details