TellWorldAI World Archive
Ask中文
News
Models
Organisations
People
Papers & concepts
Products & agents
Chips & robots
History
Topics
Ask中文

Home›Companies

OrganizationCompany

Google

Google, part of Alphabet, works on AI at several levels at once: it designs its own TPU chips, trains large language models such as PaLM and Gemini, and brought its research labs together as Google DeepMind in 2023.

Founded 1998Website about.google Era Statistical learning

2016 · 4 20062026
Events per year
14Models & products30Events4People4Investors & partners

What Google does in AI

Google is a technology company that builds AI systems at several levels at once. It designs the chips its models run on, trains large language models, and runs research organisations that develop new methods. That combination — hardware, models and labs under one roof — is what makes it a reference point when people trace how modern AI was built.

Custom silicon: the TPU

A TPU (tensor processing unit) is a chip built around the arithmetic neural networks do most: multiplying and accumulating large grids of numbers. In a blog post dated 19 May 2016, Google said it had been running TPUs inside its data centers for more than a year [1], which means the hardware was already in production before it was described publicly. The same post said TPUs powered AlphaGo in its matches against Go champion Lee Sedol [1], tying the chip programme to one of the most visible demonstrations of machine learning of that period.

Google introduced Ironwood on 9 April 2025 as its seventh-generation TPU and the first designed specifically for inference [2]. Inference is the phase in which a trained model answers requests, as opposed to training, where the model's parameters are adjusted; a chip tuned for inference is optimised for serving rather than for the long training runs that produce a model in the first place. Google said Ironwood scales to 9,216 chips per pod for a total of 42.5 exaflops [2], and that it offers 192 GB of memory per chip, six times that of the Trillium generation [2]. Memory capacity matters because large models have to keep weights and intermediate data close to the compute that uses them.

Language models: PaLM and Gemini

The PaLM paper describes a 540-billion-parameter dense Transformer language model, the Pathways Language Model, trained on 6,144 TPU v4 chips using Pathways [5]. Parameters are the learned numbers that shape a model's behaviour; a dense Transformer uses all of them for every token it processes, unlike architectures that activate only part of the network at a time. The paper reports that the 540B model outperformed the average human on BIG-bench [5], a large collection of benchmark tasks. That result is a statement about an aggregate score across many tasks, not about any one skill.

The Gemini Team submitted the Gemini report to arXiv on 19 December 2023. It describes a family of Ultra, Pro and Nano multimodal models [3] — multimodal meaning that the same model handles more than one kind of input rather than text alone — and reports that Gemini Ultra advances the state of the art on 30 of 32 benchmarks [3]. Benchmark counts of this kind depend on which benchmarks are selected and how each is scored, so the number is best read as a summary of a broad evaluation rather than a single measurement.

Organisation: Google DeepMind

Alphabet announced on 20 April 2023 that Google Brain and DeepMind would be combined into Google DeepMind [4], merging two groups that had grown up separately inside the same company. According to CNBC's report on the April 2023 merger, Demis Hassabis leads the combined unit and Jeff Dean became Google's chief scientist [4]. Consolidating the labs put model research and the surrounding research engineering under one leadership structure at a moment when large models were becoming the organisation's central product concern.

Where it sits

The pieces described here are connected. A chip programme that began with TPUs running quietly in data centers [1] later provided the 6,144 TPU v4 chips behind a 540-billion-parameter model [5], and was later extended with a generation designed for inference [2]. The model work runs from PaLM's single dense language model [5] to the multimodal Gemini family [3]. And the research effort was reorganised into a single unit [4]. Read together, they describe an organisation that treats chips, models and lab structure as one problem rather than three.

TechniquesFoundation models

ApplicationsConsumer apps

Ecosystem & regionCloudHardware & compute

20261

  1. PartnershipApple and Google announce Gemini-based collaboration for Apple's foundation models

    On 12 January 2026 Apple and Google announced a multi-year collaboration in which Google's Gemini models and cloud technology would underpin Apple's next generation of foundation models, while Apple said its privacy standards would stay unchanged.

20253

  1. AcquisitionPerplexity offers $34.5 billion for Google Chrome

    On 12 August 2025 the AI startup Perplexity made a $34.5 billion all-cash bid to buy Google's Chrome browser, presenting the sale as a remedy for Google's antitrust litigation and pledging to keep Chromium open source.

  2. PartnershipGoogle hires Windsurf's CEO in a $2.4 billion deal as OpenAI's acquisition falls apart

    In July 2025 OpenAI's $3 billion deal to acquire the AI coding startup Windsurf collapsed; the same day Google DeepMind hired Windsurf's CEO and top researchers, with Google paying $2.4 billion to license the startup's technology.

  3. UpdateLinux Foundation takes over the A2A protocol

    The Linux Foundation announced the Agent2Agent project on 23 June 2025, taking stewardship of Google's open protocol for secure communication between autonomous AI agents and placing it under vendor-neutral governance.

20242

  1. PartnershipGoogle signs a deal with Kairos Power for 500 MW of advanced nuclear power by 2035

    In October 2024 Google and Kairos Power agreed on a path to deploy 500 MW of advanced nuclear power in the United States by 2035, with a first plant by 2030 — the first US corporate deal covering multiple deployments of one advanced reactor design.

  2. IncidentGoogle pauses Gemini's people image generation after historically inaccurate images

    On 22 February 2024 Google paused Gemini's generation of images of people after users shared historically inaccurate pictures; Google blamed diversity tuning applied where no range was appropriate, and over-cautious refusals.

20233

  1. FundingHugging Face raises $235 million at a $4.5 billion valuation

    On 24 August 2023 Hugging Face raised a $235 million Series D led by Salesforce Ventures, valuing the open-model hub at $4.5 billion, more than double its 2022 valuation, with chipmakers and cloud providers among the investors.

  2. PolicyAnthropic, Google, Microsoft and OpenAI launch the Frontier Model Forum

    An industry body formed on 26 July 2023 by Anthropic, Google, Microsoft and OpenAI to work on the safe and responsible development of frontier AI models, defined as large-scale models that exceed the capabilities of the most advanced existing models.

  3. PolicySeven AI companies make voluntary commitments to the White House

    On 21 July 2023 President Biden convened seven leading AI companies at the White House and secured voluntary commitments covering eight measures, from watermarking AI-generated audio and video to independent safety testing before public release.

20221

  1. ReleaseGoogle publishes RT-1

    In December 2022, Google researchers posted RT-1, a Robotics Transformer trained on real robot data; they reported over 700 instructions at a 97% success rate and released the code.

20182

  1. PolicyGoogle publishes its AI principles

    On 7 June 2018 Google published a set of seven objectives to guide its AI research and products, together with four kinds of application it said it would not pursue.

  2. ReleaseGoogle announces Duplex at I/O 2018

    On 8 May 2018, at its I/O developer conference, Google announced Duplex, a system that holds natural-sounding phone conversations to complete narrow tasks such as scheduling certain appointments.

20171

  1. AcquisitionSoftBank agrees to acquire Boston Dynamics from Alphabet

    On 9 June 2017 SoftBank Group announced that a subsidiary had agreed to buy Boston Dynamics from Alphabet, together with the Japanese bipedal-robotics company Schaft. The terms were not disclosed.

20164

  1. MilestoneAlphabet spins its self-driving car project out as Waymo

    In December 2016 Alphabet spun its self-driving car project out into a separate company, Waymo, turning a long-running research effort into a business that supplies self-driving technology rather than cars.

  2. ReleaseGoogle launches GNMT in production for Chinese-to-English translation

    On 27 September 2016 Google Research announced GNMT, a neural machine translation system it put into production for Chinese-to-English on Google Translate, reporting large cuts in translation errors.

  3. ReleaseGoogle reveals the Tensor Processing Unit

    In May 2016 Google disclosed the Tensor Processing Unit, a custom chip built for machine learning that had already run in its data centers for over a year and powered RankBrain, Street View and AlphaGo.

  4. ReleaseGoogle introduces the Assistant and Google Home at I/O 2016

    At its I/O conference in May 2016, Google introduced the Google assistant, previewed the Google Home speaker that brings it into the house, and showed the Allo and Duo apps that carry it on phones.

20151

  1. ReleaseGoogle open-sources TensorFlow

    On 9 November 2015 Google released TensorFlow, its second-generation machine learning system, as open source under the Apache 2.0 licence, letting people outside Google build and train models with the same system Google used in its own products.

20141

  1. ReleaseGoogle unveils its own self-driving prototype without a steering wheel

    On 27 May 2014, Google published a design for a self-driving prototype with no steering wheel or pedals and said it planned to build about a hundred vehicles and test them with safety drivers.

20132

  1. AcquisitionGoogle acquires Boston Dynamics

    In December 2013 Google confirmed it had completed its acquisition of Boston Dynamics, the eighth robotics company it had bought in six months, as part of a robotics effort led by Andy Rubin.

  2. AcquisitionGoogle acquires DNNresearch

    In March 2013 the University of Toronto announced that Google had acquired DNNresearch, a start-up incorporated in 2012 by Geoffrey Hinton and two of his graduate students for work on deep neural networks.

20123

  1. ReleaseGoogle announces Google Now at I/O 2012

    Google announced Google Now at its I/O conference in June 2012, presenting it as an Android search experience that shows cards like weather, traffic and sports scores without the user searching. It arrived alongside Android 4.1 Jelly Bean and the Nexus 7 tablet.

  2. ReleaseGoogle launches the Knowledge Graph

    On 16 May 2012 Google announced the Knowledge Graph, an attempt to move search from matching strings to identifying entities; at launch it held more than 500 million objects.

  3. MilestoneNevada issues the first US autonomous vehicle licence

    On 7 May 2012 Nevada's DMV issued the first US licence for an autonomous vehicle, a Toyota Prius fitted with Google's self-driving technology, after tests on highways, Carson City streets and the Las Vegas Strip.

20111

  1. ReleaseGoogle acquires facial-recognition company PittPatt

    In July 2011 Google acquired Pittsburgh Pattern Recognition (PittPatt), a Carnegie Mellon spin-off whose software could find and track faces in photos and video, bringing facial-recognition technology into Google.

20102

  1. ReleaseGoogle publicly announces its self-driving car project

    In October 2010 Google announced on its corporate blog that it had developed technology for cars that drive themselves, a project led by Sebastian Thrun and framed around safety, time and emissions.

  2. HireGoogle staffs the self-driving project with DARPA Challenge veterans

    In 2010 Google staffed its self-driving car work with engineers who had led winning teams in the US government's DARPA autonomous vehicle races, and put paid drivers in the driver's seat to watch the cars in traffic.

20091

  1. MilestoneGoogle starts its self-driving car project

    On 17 January 2009 Google began developing self-driving cars in its Google X lab; the project, revealed in October 2010 after 140,000 miles of testing, was later renamed Waymo.

20081

  1. ReleaseGoogle launches Voice Search

    On 17 November 2008 Google added voice search to its Google Mobile App for iPhone, letting U.S. English users speak a web query instead of typing it.

20061

  1. ReleaseGoogle Translate launches with statistical machine translation

    On 28 April 2006 Google announced a statistical machine translation approach, and Google Translate launched as a service built on it — the start of free web translation for a mass audience.

Models and products over time

11
  1. 2012
    • May 16, 2012 Google Knowledge Graph Product

      A Google search feature announced in 2012 that models real-world entities and their relationships, so that search can work with "things, not strings".

    • Jun 27, 2012 Google Now Product

      Google Now is an Android feature, announced with Android 4.1 Jelly Bean in June 2012, that pushes weather, traffic, scores and other information to users as cards at the moment they need it, without a search query.

  2. 2015
    • 2015 Google TPU (first generation) Chip

      Google's first-generation Tensor Processing Unit is a custom chip built to speed up neural-network inference in its data centers. Deployed in 2015 and revealed in 2016, it showed that dedicated accelerators could beat CPUs and GPUs on cost-performance.

    • Nov 9, 2015 TensorFlow Product

      TensorFlow is Google's second-generation machine learning system, released as an open-source library under the Apache 2.0 license on 9 November 2015, succeeding the earlier DistBelief system.

  3. 2017
    • May 2017 Google TPU v2 Chip

      Google TPU v2 is the second-generation Tensor Processing Unit announced in May 2017. It added floating-point support and the bfloat16 format, and scaled to 256-chip pods rated at 11.5 petaFLOPS.

  4. 2018
    • 2018 Google TPU v3 Chip

      Google's third-generation Tensor Processing Unit, introduced in 2018, with doubled HBM capacity, higher peak throughput, lower power draw and liquid cooling for denser pods.

  5. 2020
    • 2020 Google TPU v4 Chip

      Google's TPU v4 is a machine-learning chip and supercomputer deployed since 2020, built into 4,096-chip pods with optically reconfigurable networking; PaLM and Gemini models were trained on it.

  6. 2022
    • Dec 13, 2022 RT-1 (Robotics Transformer) Model

      RT-1 is a transformer-based model for real-world robot control, trained on over 130,000 episodes collected with 13 robots and reported to perform more than 700 language instructions at a 97% success rate.

  7. 2023
    • Dec 2023 Google TPU v5p Chip

      Google TPU v5p is a 2023 TPU generation with 459 teraflops of peak BF16 and FP8 compute per chip, 95 GiB of HBM, and 8,960 chips per pod.

  8. 2024
    • Dec 2024 Google TPU v6e (Trillium) Chip

      Google TPU v6e, also called Trillium, is a Google Cloud AI accelerator whose published specifications cover per-chip BF16 and FP8 compute, high-bandwidth HBM, and pod-scale interconnects for large training jobs.

  9. 2025
    • 2025 Google TPU7x (Ironwood) Chip

      TPU7x (Ironwood) is Google's seventh-generation TPU: 192 GB of high-bandwidth memory per chip, FP8 matrix units, 9,216 chips per pod, and a programming model that exposes two devices per chip.

People4

Sundar Pichai2004–Jeff Dean1999–Kai-Fu Lee2005–2009Fei-Fei Li2017–2018

Includes3

Google DeepMindGoogle ResearchGoogle Brain–2023

Invested in1

Hugging Face

Contributed to7

Notes and sources

  1. Google supercharges machine learning tasks with TPU custom chiphttps://cloud.google.com/blog/products/ai-machine-learning/google-supercharges-machine-learning-tasks-with-custom-chip
  2. Ironwood: The first Google TPU for the age of inferencehttps://blog.google/products/google-cloud/ironwood-tpu-age-of-inference/
  3. Gemini: A Family of Highly Capable Multimodal Modelshttps://arxiv.org/abs/2312.11805
  4. Alphabet merges AI-focused groups DeepMind and Google Researchhttps://www.cnbc.com/2023/04/20/alphabet-merges-ai-focused-groups-deepmind-and-google-research.html
  5. PaLM: Scaling Language Modeling with Pathways (ar5iv render of arXiv 2204.02311)https://ar5iv.org/html/2204.02311

Cite this page: Google. TellWorld, https://www.tellworld.net/companies/google. Updated 2026-11-08