Unboxing the AI Stack
Organization · L02 · Language Models

Google

Google is the largest of the companies that make up Alphabet Inc.,1 whose SEC filings give its business address as Mountain View, California.2 It works on every layer of the AI stack: models, chips and data centers. Its AI lab, Google DeepMind, was formed on April 20, 2023, when DeepMind and the Brain team from Google Research were combined into a single unit.3 Google DeepMind introduced the Gemini model family on December 6, 2023.4 Google also designs its own accelerators. It announced the Tensor Processing Unit (TPU), a custom ASIC for machine learning, in May 2016, saying it had already run TPUs in its data centers for more than a year.5

Its seventh-generation TPU7x (Ironwood) carries 192 GB of HBM per chip at about 7.37 TB/s, vendor-stated, and became generally available on Google Cloud on March 31, 2026 (AI Accelerators With the Most HBM per Chip).67 In Epoch AI’s tracking, seven of the 20 largest AI data centers by estimated IT power in operation are Google’s, led by Pryor (North), Oklahoma, at an estimated 368 MW (Largest AI Data Centers by Estimated Power Draw).8 Google is one of eight US-headquartered companies named in a July 2026 export rule that eases licensing for AI chips sent to the UAE (How US AI Chip Export Controls Work).9

Among its models, Gemini 3.8 Flash scores 10.37% on ARC-AGI-3’s Semi-Private set on ARC Prize’s Standard harness, the highest of any model from outside OpenAI and Anthropic (Highest ARC-AGI-3 Scores, by Model),10 and Gemini 3 Flash resolved 75.8% of SWE-bench Verified tasks in the shared mini-SWE-agent v2.0.0 harness (SWE-bench Verified, Pro and What Replaced Them).11 Gemini 3.1 Flash-Lite lists $1.50 per 1M output tokens (Cheapest LLM APIs per 1M Output Tokens).12 Its Veo 3.1 video model on Vertex AI makes 4-, 6- or 8-second clips at up to 4K, with sound (How AI Video Generation Works).13 In robotics, Google DeepMind’s RT-2 paper of July 2023 named the vision-language-action category, and its Gemini Robotics 2 was announced on July 30, 2026 (How Vision-Language-Action Robot Models Work).1415

Illustration: Google sits in the Language Models layer of the stack, and also makes chips and runs data centers.

Key facts

Parent companyAlphabet Inc.1
Business address (Alphabet Inc.)Mountain View, California2
AI labGoogle DeepMind, formed April 20, 20233
Model familyGemini, since December 6, 20234
First TPU announcedMay 20165
TPU7x (Ironwood) HBM per chip192 GB Vendor-stated6
Largest site in Epoch AI’s top 20, IT power in operation368 MW (Google Pryor North) Estimate8

As of . Google’s TPU7x page heads its table column in GiB (192 GiB is about 206 GB); data center power is Epoch AI’s estimate. How figures are labeled.

Appears in

Sources

  1. Alphabet, G is for Google (Larry Page’s 2015 letter on abc.xyz), read via the r.jina.ai reader (abc.xyz returns 403 to our tools), retrieved 2026-10-13.
  2. Alphabet Inc., SEC EDGAR submissions record (business address), retrieved 2026-10-13.
  3. Google DeepMind, Announcing Google DeepMind, April 20, 2023, retrieved 2026-10-13.
  4. Google, Introducing Gemini: our largest and most capable AI model, December 6, 2023, retrieved 2026-10-13.
  5. Google Cloud Blog, Google supercharges machine learning tasks with TPU custom chip, May 2016, retrieved 2026-10-13.
  6. Google Cloud, TPU7x (Ironwood) documentation, retrieved 2026-10-13.
  7. Google Cloud, Cloud TPU release notes, retrieved 2026-10-13.
  8. Epoch AI, Frontier Data Centers (hub and directory, including Google Pryor (North)), CC BY 4.0, retrieved 2026-10-13.
  9. Bureau of Industry and Security, “Enhanced Favorable Treatment for the United Arab Emirates Under the Export Administration Regulations,” final rule, Federal Register, July 14, 2026, retrieved 2026-10-13.
  10. ARC Prize, ARC-AGI-3 Verified leaderboard data (v3.json, generated 2026-10-05), retrieved 2026-10-13.
  11. SWE-bench team, SWE-bench Leaderboards (Bash Only view, mini-SWE-agent v2.0.0 runs), retrieved 2026-10-13.
  12. Google, Gemini Developer API pricing, retrieved 2026-10-13.
  13. Google Cloud, Veo 3.1 model page (Vertex AI), retrieved 2026-10-13.
  14. Google DeepMind (Brohan et al.), RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control, arXiv 2307.15818 (v1 July 28, 2023), retrieved 2026-10-13.
  15. Google DeepMind, Gemini Robotics 2 brings whole body intelligence to robots, July 30, 2026, retrieved 2026-10-13.