AI Chips · Brand guide

Google TPU: The Complete Guide

Google's own AI chip, and until 2026 you could only rent it. What Google's Tensor Processing Units are, how they differ from GPUs, who builds them, the official prices, the Anthropic deal, and the facts buried in Alphabet's and Broadcom's filings. Three reading levels, every number sourced.

Stuck at any point? Ask an AI about this page →
Google TPU 16 sections · 3 levels 64 linked sources Checked September 2026
How to read this page. Every brand in our AI Chips section uses the same 16 sections in the same order. Each section starts Simple, then goes Deeper, then Expert: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where a company has not published something, we say so rather than estimate.

1.At a glance

SimpleStart here

A TPU (Tensor Processing Unit) is Google's own AI chip, built only for the maths AI needs 1. Google has used them inside its own services since 2015 2 and rents them to other companies through Google Cloud 3. The current chip is Ironwood, the seventh generation 4; the eighth generation, TPU 8t for training and TPU 8i for running models, was announced in April 2026 5.

What it isA chip designed only for AI maths, not a general graphics chip 1
Latest chipIronwood (TPU7x); TPU 8t and 8i "coming soon" 3
Who builds itGoogle designs; Broadcom co-develops and supplies 6
How you get oneRent on Google Cloud; since 2026, a few customers can have TPUs in their own data centres 7
DeeperThe detail

TPUs are Google's answer to NVIDIA. Anthropic has agreed to use up to one million of them 8, and Google Cloud's revenue grew 82% in the June 2026 quarter to $24.8 billion 9. Google publishes a price per chip per hour, which NVIDIA does not do for its data-centre chips: Ironwood lists at $12.00 an hour on demand 10.

ExpertFor specialists

Alphabet's revenue in the June 2026 quarter was $119.8 billion, up 24%; Google Cloud earned $8.8 billion of operating profit on $24.8 billion of revenue, and Alphabet spent $44.9 billion on capital projects in the quarter 9. Google Cloud's backlog of signed business reached $514 billion 11. Alphabet guided 2026 capital spending of $175 to $185 billion in February 12 and raised it to $195 to $205 billion in July 13.

2.Company card

SimpleStart here

TPUs are made by Google, part of Alphabet. Google was founded in 1998 4; Alphabet became its parent company in October 2015 14. Sundar Pichai is chief executive of both 11.

Legal nameAlphabet Inc. 4
Headquarters1600 Amphitheatre Parkway, Mountain View, California 94043, USA 4 (map)
Chief ExecutiveSundar Pichai, CEO of Google and Alphabet 11
Google Cloud CEOThomas Kurian, since November 2018 15
Employees190,820 (31 December 2025) 4
StockGOOGL and GOOG, Nasdaq Global Select Market 4
Investor siteabc.xyz/investor
DeeperThe detail

Alphabet's annual report describes the chips as "our own custom-built Tensor Processing Units (TPUs), such as Ironwood, our seventh-generation TPU" 4. TPUs sit inside Google Cloud, which reports as its own segment 9. Alphabet had 198,933 employees at 30 June 2026 9.

ExpertFor specialists

Alphabet's fiscal 2025 revenue was $402.8 billion and its capital spending $91.4 billion 12. Google Cloud's fourth-quarter 2025 revenue was $17.7 billion with $5.3 billion of operating income 12; by the second quarter of 2026 that had risen to $24.8 billion and $8.8 billion 16.

3.The story

SimpleStart here
  • 2015 First TPU starts running inside Google's data centres, for running models only 2.
  • 2016 Google reveals the TPU publicly; it had helped power AlphaGo 17.
  • 2017 TPU v2 can also train models; Google starts renting TPUs 18.
  • 2018 TPU v3 needs liquid cooling, a first for Google's data centres 19.
  • 2021 TPU v4: one pod passes an exaflop 20.
  • 2023 TPU v5e and v5p 21 22.
  • 2024 Trillium (TPU v6e) 23.
  • 2025 Ironwood (TPU7x), built for running AI models 24.
  • 2026 Eighth generation split into two chips, TPU 8t and TPU 8i 5.
DeeperThe detail

Google built the first TPU for its own needs. When it revealed the chip in May 2016, it said TPUs had been running in its data centres "for more than a year", powering RankBrain search, Street View and AlphaGo 17. The first TPU only ran trained models; TPU v2 in 2017 could train them too, and Google offered 1,000 of them free to researchers 18. From TPU v4, Google linked thousands of chips with optical switches that can rewire the network 25.

ExpertFor specialists

Google's own paper on the first TPU reported 15 to 30 times the speed of contemporary GPUs and CPUs on its workloads, and 30 to 80 times the performance per watt 2. Its 2023 paper on TPU v4 described a 4,096-chip machine in service since 2020, 1.2 to 1.7 times faster than NVIDIA's A100 while using 1.3 to 1.9 times less power, on Google's comparison 25. Ironwood was announced in April 2025 24; Google's release notes list it in preview from 24 November 2025 and generally available from 31 March 2026 26.

4.The lineup

SimpleStart here
NameWhat it isAvailable
TPU v5eLow-cost chip for training and running mid-size modelsNov 2023 21
TPU v5pLarge-scale training chip; 8,960 per podDec 2023 22
Trillium (TPU v6e)Sixth generationDec 2024 26
Ironwood (TPU7x)Seventh generation, built for running models; 9,216 per pod2025–26 24 26
TPU 8tEighth-generation training chipComing soon 3
TPU 8iEighth-generation chip for running models and AI agentsComing soon 3
DeeperThe detail

The eighth generation is the first time Google has split a TPU generation into two chips: 8t for training and 8i for inference (running models) 5. Google said both would be generally available "later in 2026" 5; its TPU page still shows them as "Coming Soon" 3. Broadcom said in September 2026 that it had begun production shipments of TPU 8i 27.

ExpertFor specialists

Ironwood has two TensorCores and four SparseCores per chip 28; Trillium has one TensorCore 29. Google's documentation says TensorFlow is not supported on Ironwood; JAX and PyTorch are 28. TPU 8i pods hold up to 1,152 chips in a new "Boardfly" layout that cuts the longest path between chips from 16 hops to 7 30.

5.How it’s made

SimpleStart here

Google designs TPUs, but it does not say who manufactures them 4.

  1. Design: Google, with Google DeepMind for the eighth generation 5.
  2. Co-development and supply: Broadcom, under a long-term agreement to "develop and supply" future TPUs through up to 2031 6.
  3. Chip factory: not disclosed by Google. Broadcom says about 95% of the wafers its manufacturers produced in 2025 came from TSMC 31, and press reports name TSMC 32.
  4. Memory: Ironwood carries 192 GiB of HBM memory 28; Google does not name the supplier.
  5. Racks: Google builds its own systems, with Arm-based Axion processors and liquid cooling for TPU 8 5.
DeeperThe detail

Broadcom's filing of 6 April 2026 describes two agreements with Google running to 2031: one to develop and supply custom TPUs for future generations through up to 2031, and one to supply networking and other parts for next-generation AI racks 6. On its September 2026 call Broadcom said it had delivered Ironwood "in high volume to both Anthropic and Google" 27.

Other partners appear only in press reports, not confirmed by Google: MediaTek designing TPU 8i 32 33, and Intel making or packaging millions of TPUs through 2028 34.

ExpertFor specialists

Google is also adding suppliers for other custom chips: filings by Marvell, reported in August 2026, show Google holding a warrant for about 59 million Marvell shares that vests as Google buys up to $120 billion of Marvell products 35. Ironwood's lifetime emissions are 23% from manufacturing and 77% from running it on Google's market-based accounting (8% manufacturing on a location basis), and it is 3.7 times more carbon-efficient than TPU v5p 36.

6.Where it’s made: factories and addresses

SimpleStart here

Google does not disclose which factory makes TPUs, so there is no factory address to give. What Google does publish is where you can rent TPUs 37:

Google Cloud regionLocationTPUs offered
us-central1Iowa, USA 10Ironwood, Trillium, v5p, v5e 37
us-east5Columbus, Ohio, USA 10Trillium, v5p 37
us-east1South Carolina, USA 10Trillium 37
us-south1Dallas, Texas, USA 10Trillium, v5e 37
europe-west4NetherlandsTrillium, v5p, v5e 37
asia-northeast1Tokyo, JapanTrillium 37
southamerica-west1Santiago, ChileTrillium 37

Google publishes data centre cities but not street addresses 38.

DeeperThe detail

Ironwood is offered only in the Iowa region 37. Google has operated in Iowa since 2007 with a $600 million first investment in Council Bluffs, has invested more than $20 billion in the state, and announced $7 billion more for Council Bluffs and a new Cedar Rapids campus 38. No TPU of any generation is offered in Google Cloud's Mumbai or Delhi regions 37.

ExpertFor specialists

TPUs are now also leaving Google's buildings. Google delivered its first TPU systems to customers' own data centres in the June 2026 quarter 13, and Blackstone is putting $5 billion into a joint venture with Google that offers TPU capacity from new US data centres, with 500 MW online in 2027 39. Google's data centres ran at an average power usage effectiveness of 1.09 in 2025 40.

7.How it works

SimpleStart here

A TPU is built around a matrix unit: a grid of tens of thousands of small multiply-and-add circuits that pass numbers to their neighbours in a steady rhythm, called a systolic array 1. Because AI models are mostly huge grids of numbers multiplied together, this one job covers most of the work.

DeeperThe detail

Each TPU core has matrix units, a vector unit for steps like softmax, and a scalar unit for control 1. On Trillium and Ironwood each matrix unit is 256 by 256 multiply-accumulators; earlier generations used 128 by 128 1. SparseCores handle "embeddings", the lookup tables used in recommendation models 1. Chips link directly to each other through a fast inter-chip interconnect: inside a cube of 64 chips, each chip connects to six neighbours 41.

ExpertFor specialists

Optical circuit switches join 144 cubes into a 9,216-chip Ironwood superpod with 1.77 petabytes of shared memory, and can reroute around failed parts 41. Google's 2023 paper said the optical switches cost under 5% of the system and use under 3% of its power 25. Software runs through the XLA compiler: JAX was "co-designed with the TPU architecture", and TorchTPU, announced in April 2026, runs PyTorch natively and is in preview with select customers 41 42 43.

8.Specs explained

SimpleStart here

Google's own figures, per chip. Google calls these "peak" figures and does not label them dense or sparse 28; see our How AI chips work guide.

ChipAI computeMemoryMemory speedChips per pod
TPU v5e197 TFLOPS BF1616 GB800 GiB/s256 44
TPU v5p459 TFLOPS BF1695 GiB2,765 GB/s8,960 45
Trillium (v6e)918 TFLOPS BF1632 GB1,638 GB/s256 29
Ironwood (TPU7x)4,614 TFLOPS FP8; 2,307 BF16192 GiB HBM7,380 GB/s9,216 28 41
TPU 8t12.6 PFLOPS FP4216 GB6,528 GB/s9,600 30 5
TPU 8i10.1 PFLOPS FP4288 GB8,601 GB/sUp to 1,152 30
DeeperThe detail
PodChipsComputeShared memory
Ironwood superpod9,21642.5 exaflops FP81.77 PB 41
TPU 8t superpod9,600121 exaflops2 PB 5

Google says Ironwood has twice the performance per watt of Trillium 24 and TPU 8t and 8i up to twice that of Ironwood 30. Those are Google's own claims.

ExpertFor specialists

Compare like with like: TPU 8 figures are FP4, Ironwood's headline is FP8, and older chips are quoted in BF16, so each step up in the table partly reflects a smaller number format 30 28. Ironwood's chips link at 1,200 GB/s each way and have 100 Gb/s each to the data-centre network 28. TPU 8i has 384 MB of on-chip memory, three times the previous generation 30.

9.Official pricing

SimpleStart here

Google publishes an hourly price per TPU chip. These are Google Cloud's list prices in US dollars 10:

ChipOn demand1-year commitment3-year commitmentRegion
Ironwood (TPU7x)$12.00$8.40$5.40Iowa 10
Trillium (v6e)$2.70$1.89$1.22South Carolina, Ohio 10
TPU v5p$4.20$2.94$1.89Ohio, South Carolina 10
TPU v5e$1.20$0.84$0.54Iowa and others 10
TPU 8t / TPU 8iNot published by the company 10

All prices are per chip per hour. Commitments work out at 30% off for one year and 55% off for three years, by our arithmetic from Google's table.

DeeperThe detail

Two cheaper ways to book: "Flex-start", where you wait for capacity, lists Ironwood at $6.00 an hour and Trillium at $1.35; "Calendar mode" bookings list Ironwood at $8.40 10. Spot (interruptible) prices are dynamic and can change up to once every 30 days 10. Google does not publish a price for buying TPU hardware outright 46.

ExpertFor specialists

For comparison, Ironwood's $12.00 on-demand rate sits close to what Microsoft Azure and Google Cloud list per NVIDIA H100 or H200 GPU (roughly $10.60 to $12.29 an hour; see our NVIDIA guide), while Ironwood carries 192 GiB of memory per chip 28. No TPU is offered in India's Google Cloud regions, so there is no rupee price 37.

10.The software lock-in

SimpleStart here

Google's advantage is owning the whole stack: its own chip, its own network, its own data centres and its own AI models 5. The software that runs on TPUs, JAX and the XLA compiler, comes from Google too 41. TPU 8 supports JAX, PyTorch, vLLM and SGLang 5.

DeeperThe detail

The long-standing weakness was PyTorch, the framework most AI developers use. Google addressed it in two steps: in October 2025 the popular vLLM serving tool got a TPU backend that handles JAX and PyTorch through one path 47, and in April 2026 Google announced TorchTPU, which runs PyTorch natively on TPUs 42. Pichai said on the July 2026 call that the software stack enables "workload portability across GPUs and TPUs" 11.

ExpertFor specialists

Scale is the other moat. Google's "Virgo" network can link more than 134,000 TPU 8t chips in a single fabric 30, and up to a million chips in a single logical cluster 5. Google claims TPU 8t delivers up to 2.7 times Ironwood's training price-performance 30.

11.Who buys it and why

SimpleStart here

The biggest outside user is Anthropic. In October 2025 it agreed to use up to one million TPUs, with "well over a gigawatt" of capacity coming online in 2026, a deal worth "tens of billions of dollars" 8. Other named Ironwood customers include Lightricks and Essential AI 48.

DeeperThe detail
  • Anthropic (April 2026): "multiple gigawatts" more of next-generation TPU capacity from 2027, with Google and Broadcom, mostly in the US 49; Broadcom's filing puts it at about 3.5 gigawatts, contingent on Anthropic's commercial success 6.
  • Apple: its 2024 paper said Apple trained its server model on 8,192 TPU v4 chips and its on-device model on 2,048 TPU v5p chips 50.
  • Safe Superintelligence: using TPUs for research, announced April 2025 51.
  • OpenAI: said in June 2025 it was "in early testing" with some TPUs but had no plans to deploy them at scale 52.
  • Meta: press reports in February 2026 of a multibillion-dollar deal to rent TPUs; neither company confirmed it 53.
ExpertFor specialists

Google now sells TPU systems, not just rental time. Pichai said in April 2026 that Google would "deliver TPUs to a select group of customers in their own data centers" 7; the first revenue from these sales came in the June 2026 quarter, with most expected in 2027 13. Broadcom said it expects Anthropic to deploy 5 gigawatts of TPU 8i in 2027 and to become Broadcom's largest custom-chip customer that year 27.

12.Hidden in plain sight

SimpleStart here

Public facts buried in the filings and releases. Each links to the exact document.

  1. Google's signed-but-unbilled cloud business nearly doubled in one quarter, to $462 billion at 31 March 2026 54, and reached $514 billion by June 11.
  2. The TPU deal is in Broadcom's filings, not Google's: a long-term agreement to develop and supply Google's TPUs through up to 2031 6.
  3. Alphabet has $232.7 billion of long-term purchase commitments for infrastructure, energy and content 54.
  4. Capital spending guidance was raised twice in 2026, from $175–185 billion in February, to $180–190 billion in April, to $195–205 billion in July 12 13.
  5. Google delivered TPUs to customers' own data centres for the first time in the June 2026 quarter 13.
DeeperThe detail
  1. $75.6 billion of data centre leases have been signed but not started, commencing between 2026 and 2031 54.
  2. Few suppliers: Alphabet says supply of "specialized AI chips, is limited to a small number of qualified suppliers" 4.
  3. Ironwood is not supported by TensorFlow, Google's own older AI framework 28.
  4. Broadcom's AI revenue was $16.7 billion in one quarter, up 221% 55.
  5. Electricity demand rose 37% in 2025, and supply-chain emissions 25%, driven by AI hardware 56.
ExpertFor specialists
  1. Water use reached 10.9 billion gallons in 2025, up 34%, as reported from Google's environmental report 57.
  2. Manufacturing is almost a quarter of Ironwood's lifetime carbon (23%, on market-based accounting) 36.
  3. No TPUs in India: none of Google's TPU generations is offered in its Mumbai or Delhi regions 37.
  4. Commitment discounts are uniform: exactly 30% off for one year and 55% for three, across generations, by our arithmetic from Google's prices 10.
  5. The first TPU paper claimed 30 to 80 times better performance per watt than the chips of its day 2.

13.Weak spots

SimpleStart here

Limits, in Google's own words and documents:

  • Few suppliers for specialised AI chips, some parts from a single source 4.
  • Not enough capacity: "demand is outpacing capacity" (July 2026) 13.
  • Hard to get outside Google Cloud: hardware sales go only to "a select group of customers" 7.
  • Few regions: Ironwood only in Iowa; no TPUs in India 37.
DeeperThe detail

Google itself decides who gets TPUs first. On the July 2026 call Pichai said the first priority is allocating what Google needs "to compete at the frontier" 13. In February 2026 he said he expected Google to "go through the year in a supply constrained way" 58.

ExpertFor specialists

Dependence runs both ways: Broadcom names the loss of a custom-chip customer as a risk, and its top five customers made up about 40% of its revenue in fiscal 2025 31. Google has not confirmed who makes its chips, so the foundry and memory risk cannot be assessed from its filings 4.

14.India angle

SimpleStart here

Google is building its first AI hub in India at Visakhapatnam, Andhra Pradesh: about $15 billion over 2026 to 2030, with a gigawatt-scale data centre campus, built with AdaniConneX and Airtel 59 60 61. Ground was broken on 28 April 2026 at Tarluvada 62.

DeeperThe detail

The Visakhapatnam site is planned at nearly 1 gigawatt across three campuses, built by AdaniConneX and Nxtra by Airtel 62. Google Cloud already has two Indian regions, Mumbai and Delhi NCR (opened July 2021) 63, but offers no TPUs in either 37.

ExpertFor specialists

No official Google release we found says TPUs will be installed at Visakhapatnam 62. The IndiaAI Mission said in March 2026 it had onboarded more than 38,000 GPUs; its release does not mention Google or TPUs 64.

15.What’s next

SimpleStart here

Only what Google and its partners have announced:

  • TPU 8t and TPU 8i: generally available "later in 2026" 5.
  • TPU 8i shipping: Broadcom began production shipments in its quarter to August 2026 27.
  • More TPUs outside Google Cloud: hardware deliveries to customers and the Blackstone TPU cloud, 500 MW in 2027 13 39.
DeeperThe detail

Broadcom plans to deliver "multi-tens of billions of dollars of TPUs annually over the next several years" 27, and Alphabet expects its capital spending "to increase significantly in 2027" 13.

ExpertFor specialists

The test is supply. Anthropic's roadmap, as described by Broadcom, runs from 1 gigawatt of Ironwood in 2026 to 5 gigawatts of TPU 8i in 2027 and 10 more in 2028 27; whether TSMC, Intel or others make those chips has not been confirmed by Google 32 34.

16.Sources

64 sources, all checked September 2026. Official = the company or organisation’s own page; Filing = a document filed with a regulator; Press = news coverage, used only where no official source exists and labelled in the text.

  1. Google Cloud: TPU system architectureOfficial
  2. Jouppi et al. (Google), ISCA 2017: In-Datacenter Performance Analysis of a Tensor Processing UnitPaper
  3. Google Cloud: Cloud TPU product pageOfficial
  4. Alphabet: Form 10-K, fiscal year 2025Filing
  5. Google (The Keyword): Eighth-generation TPUs for the agentic era (22 Apr 2026)Official
  6. Broadcom: Form 8-K, long-term TPU agreement with Google (6 Apr 2026)Filing
  7. Google (The Keyword): CEO remarks, Alphabet Q1 2026 earnings (29 Apr 2026)Official
  8. Anthropic: Expanding our use of Google Cloud TPUs and services (23 Oct 2025)Official
  9. Alphabet Investor Relations: Alphabet second-quarter 2026 results (22 Jul 2026)Official
  10. Google Cloud: Cloud TPU pricingOfficial
  11. Google (The Keyword): CEO remarks, Alphabet Q2 2026 earnings (22 Jul 2026)Official
  12. Alphabet Investor Relations: Alphabet fourth-quarter and fiscal year 2025 results (4 Feb 2026)Official
  13. Investing.com (press): Alphabet Q2 2026 earnings call transcriptPress
  14. Alphabet: Form 10-K, fiscal year 2015 (Alphabet becomes successor to Google)Filing
  15. Google Cloud Press Corner: Thomas Kurian, CEO, Google CloudOfficial
  16. Alphabet (SEC exhibit): Second-quarter 2026 results, Form 8-K exhibit 99.1Filing
  17. Google Cloud Blog: Google supercharges machine learning tasks with TPU custom chip (19 May 2016)Official
  18. Google (The Keyword): Google Cloud offers TPUs for machine learning (17 May 2017)Official
  19. HPCwire (press): Google I/O 2018: TPU 3.0 delivers 100 petaflops, requires liquid coolingPress
  20. Google (The Keyword): Google I/O 2021 announcements (18 May 2021)Official
  21. Google Cloud Blog: Announcing Cloud TPU v5e (29 Aug 2023)Official
  22. Google Cloud Blog: Introducing Cloud TPU v5p and AI Hypercomputer (6 Dec 2023)Official
  23. Google Cloud Blog: Introducing Trillium, sixth-generation TPUs (15 May 2024)Official
  24. Google (The Keyword): Ironwood: the first Google TPU for the age of inference (9 Apr 2025)Official
  25. Jouppi et al. (Google), ISCA 2023: TPU v4: An Optically Reconfigurable Supercomputer for Machine LearningPaper
  26. Google Cloud: Cloud TPU release notesOfficial
  27. Investing.com (press): Broadcom Q3 fiscal 2026 earnings call transcriptPress
  28. Google Cloud: TPU7x (Ironwood) documentationOfficial
  29. Google Cloud: TPU v6e (Trillium) documentationOfficial
  30. Google Cloud Blog: TPU 8t and TPU 8i technical deep diveOfficial
  31. Broadcom: Form 10-K, fiscal year 2025Filing
  32. Tom's Hardware (press): Google splits its TPU into two chips for the first time (27 Apr 2026)Press
  33. TrendForce (press): Marvell, AMD reportedly shake up Google TPU race (20 Aug 2026)Press
  34. Tech Times (press): Google's reported Intel TPU order (10 Jun 2026)Press
  35. EE Times (press): Google's Marvell deal shows custom silicon spreading beyond the TPU (28 Aug 2026)Press
  36. Google Cloud Blog: Ironwood TPUs deliver 3.7x carbon efficiency gains (6 Apr 2026)Official
  37. Google Cloud: Cloud TPU regions and zonesOfficial
  38. Google Data Centers: Council Bluffs, IowaOfficial
  39. Blackstone: Blackstone announces joint venture with Google to create new TPU cloud (18 May 2026)Official
  40. Google Data Centers: Operating sustainably (fleet PUE)Official
  41. Google Cloud Blog: Inside the Ironwood TPU codesigned AI stackOfficial
  42. Google Developers Blog: TorchTPU: running PyTorch natively on TPUs (7 Apr 2026)Official
  43. Google Cloud Blog: AI infrastructure at Next '26 (22 Apr 2026)Official
  44. Google Cloud: TPU v5e documentationOfficial
  45. Google Cloud: TPU v5p documentationOfficial
  46. DatacenterDynamics (press): Google to sell TPUs to a select group of customers for their data centersPress
  47. vLLM project blog: vLLM TPU: a new unified backend (16 Oct 2025)Official
  48. Google Cloud Blog: Ironwood TPUs and new Axion-based VMs (6 Nov 2025)Official
  49. Anthropic: Google and Broadcom partnership for compute (6 Apr 2026)Official
  50. Tom's Hardware (press), on Apple's paper arXiv 2407.21075: Apple uses thousands of Google TPUs for its AI models (Jul 2024)Press
  51. DatacenterDynamics (press): Safe Superintelligence to use Google's TPU chips (10 Apr 2025)Press
  52. Yahoo Finance / Reuters (press): OpenAI says no plan to use Google's TPUs at scale (30 Jun 2025)Press
  53. SiliconANGLE (press, citing The Information): Google and Meta reportedly strike multibillion-dollar AI chip deal (26 Feb 2026)Press
  54. Alphabet: Form 10-Q, quarter ended 31 March 2026Filing
  55. Broadcom Investor Relations: Broadcom third-quarter fiscal 2026 results (2 Sep 2026)Official
  56. Google (The Keyword): 2026 Environmental ReportOfficial
  57. Axios (press): Google's AI boom drives emissions and power use (30 Jun 2026)Press
  58. Fortune (press): Alphabet AI spending and supply constraints (4 Feb 2026)Press
  59. Google India (The Keyword): Our first AI hub in India, powered by a $15 billion investment (14 Oct 2025)Official
  60. Bharti Airtel: Airtel partners with Google on India's first mega AI hub in Visakhapatnam (Oct 2025)Official
  61. Adani Group: Adani and Google partner on Visakhapatnam data centre campus (Oct 2025)Official
  62. Google Cloud Press Corner: Google breaks ground on India AI hub (28 Apr 2026)Official
  63. Google Cloud Blog: Google Cloud region in Delhi NCR is now open (15 Jul 2021)Official
  64. Press Information Bureau, Government of India: IndiaAI Mission compute update (25 Mar 2026)Government

On this page

1. At a glance2. Company card3. The story4. The lineup5. How it’s made6. Where it’s made: factories and addresses7. How it works8. Specs explained9. Official pricing10. The software lock-in11. Who buys it and why12. Hidden in plain sight13. Weak spots14. India angle15. What’s next16. Sources

Quick facts

HeadquartersMountain View, California
Founded1998 (Google)
CEOSundar Pichai
Makes its own chips?Designs them; partners build them
Main chip partnerBroadcom
Latest quarter Google Cloud revenue$24.8 billion

AI Chips

All chip brandsAmazon TrainiumAMDAppleCerebrasGroqHuawei AscendIntelMeta MTIAMicrosoft MaiaNVIDIAAI memory (HBM) and why it is scarceExport rules and the US–China chip contestHow a chip is made: sand to serverHow AI chips workIndia’s semiconductor pushHow to read a chip spec sheetToday’s AI news

In the news

Newest stories mentioning Google TPU, straight from the companies’ own newsrooms and the wider press. Links open the original.

All of today’s AI news →

Ask an AI about this page

Opens your assistant with this page as the source, and a question rather than a summary. It will ask what you are building before it answers.

ChatGPTClaudeGeminiPerplexityGrok

Nothing is sent from here. The link carries only this page’s title and address.