Google

Updated October 2026

Google is the maker of Gemini and the only AI company with a full owned stack — frontier models from Google DeepMind, its own TPU chips, Google Cloud, and distribution through Search, Android, Chrome and Workspace. In August 2026 it restructured its AI leadership: Demis Hassabis stepped back, Sergey Brin returned from retirement to drive commercialisation, and Google's compute business is now outgrowing its model business.

Website →API →Pricing →Status →

Google is the maker of Gemini and the only AI company that owns its entire stack: frontier models from Google DeepMind, its own TPU chips, Google Cloud to serve them, and unrivalled distribution through Search, Android, Chrome, YouTube and Workspace. Google invented the transformer architecture behind modern AI in 2017, was caught flat-footed by ChatGPT in 2022, and has since staged a comeback — the Gemini app passed 900 million monthly active users by Google I/O in May 2026, up from 400 million a year earlier (TechCrunch).

The 2026 story has split in two, and the halves point in opposite directions.

The infrastructure business is winning. Google Cloud grew 82% to $24.8 billion in Q2 2026, backlog reached $514 billion, Alphabet raised full-year capex guidance to $195–205 billion, and Google began shipping TPU systems into customers’ own data centres — becoming a chip vendor, not just a cloud. Rival Anthropic alone contracted for up to $40 billion of Google compute and as many as a million Ironwood chips.

The model business is struggling. Gemini 3.5 Pro has missed three release targets since Sundar Pichai promised it “within a month” at I/O in May. On 30 September 2026 Google announced Gemini 4 Argon, which Artificial Analysis scores at 53 on its Intelligence Index, level with OpenAI’s GPT-6 Astra, and on 13 October it became generally available in the Gemini app for Google AI Pro and Ultra subscribers and on the Gemini API. In the first week of August 2026 Google restructured its AI leadership outright: Demis Hassabis stepped back from running Google DeepMind, Sergey Brin returned from retirement to drive AI commercialisation, and Jeff Dean left after 27 years — the culmination of a talent exodus that has been running all year. Alphabet’s shares fell roughly 5% on the news.

The uncomfortable thread connecting the two: Google is selling so much TPU capacity to rivals that its own researchers report queuing for compute — and departing staff cite that as a reason for leaving.

Quick facts

History and founding

Google was founded in 1998 by Larry Page and Sergey Brin. Its researchers published “Attention Is All You Need” in 2017, inventing the transformer — the architecture behind virtually every modern large language model, including those of its rivals. Despite that head start, OpenAI’s November 2022 launch of ChatGPT caught Google off-guard; the company declared an internal “code red,” and a botched 2023 demo of its Bard chatbot wiped roughly $100 billion off the stock in a day.

The turnaround began with structure. In April 2023 Google merged DeepMind and Google Brain into a single division, Google DeepMind, under Demis Hassabis, and focused it on the Gemini family launched that December. By 2026 Google had reorganised further, folding the Gemini product team into DeepMind to tighten the research-to-product loop.

Who owns Gemini? Google DeepMind and the corporate structure

The ownership chain is straightforward, and worth stating plainly because it is frequently confused:

Alphabet Inc. is the publicly traded holding company (Nasdaq: GOOGL/GOOG). Google LLC is its main subsidiary. Google DeepMind is Google’s AI division, and it builds Gemini. So Gemini is owned by Google, which is owned by Alphabet — DeepMind is not a separate company and has not been since 2023, when Google merged the UK-based DeepMind it acquired in 2014 with its internal Google Brain team into one division. There is no separate DeepMind stock, and you cannot invest in Gemini directly; the only way to own any of it is Alphabet shares.

The reporting lines changed in August 2026. Following the restructure (see below), day-to-day control of Google DeepMind sits with Koray Kavukcuoglu as senior vice president, reporting directly to Sundar Pichai, CEO of both Alphabet and Google. Sergey Brin, Google’s co-founder, oversees AI development and commercialisation. Demis Hassabis — DeepMind’s co-founder and a 2024 Nobel laureate in Chemistry for AlphaFold — moved from CEO to chairman of Google DeepMind and chief scientist of Alphabet, a role focused on long-horizon research rather than shipping.

That reshuffle also moved DeepMind’s centre of gravity from London to Mountain View, where Kavukcuoglu is based — a meaningful change for a division that had operated with unusual independence from its parent since the acquisition.

The leadership shake-up

In the first week of August 2026, Google restructured the group that builds its AI — losing a CEO, a chief scientist, a Google Brain co-founder and, earlier in the year, a Gemini co-lead. Alphabet’s shares fell about 5% on the announcement, wiping an estimated $160–200 billion in market value.

What changed:

• Demis Hassabis stepped down as CEO of Google DeepMind (5 August 2026), becoming chairman of DeepMind and Alphabet’s first chief scientist.
• Sergey Brin came out of retirement to lead AI development and product commercialisation. The Financial Times reported on 7 August that the reshuffle was designed to accelerate Gemini monetisation and close the competitive gap — a deliberate shift from Hassabis’s research-first instincts toward Brin’s commercial ones.
• Koray Kavukcuoglu took over day-to-day operations, relocating from London to Mountain View and reporting to Pichai.
• Jeff Dean, Google’s chief scientist, and Sanjay Ghemawat left after 27 years to found a startup, Discovery Loop, joined by DeepMind’s Oriol Vinyals and Quoc Le.

The exodus started well before August. In January, reinforcement-learning pioneer David Silver left to found Ineffable Intelligence. In a single week in June, Gemini co-lead Noam Shazeer went to OpenAI and Nobel laureate John Jumper — AlphaFold’s co-inventor — went to Anthropic, along with AlphaFold veterans Jonas Adler and Alexander Pritzel (Axios).

Why people left. Fortune’s reconstruction points to four compounding pressures: repeated Gemini 3.5 Pro delays, 60-hour weeks and burnout, anger over the Pentagon contract (below), and a loss of DeepMind’s independence as power moved to Mountain View. Separately, researchers report competing internally for TPU capacity that Google is selling to external customers — including the rivals hiring them.

This is the most significant leadership change at any frontier lab in 2026, and its effect on Google’s model roadmap will not be visible for several quarters.

The full-stack advantage: models, chips and cloud

Google’s defining edge is that it owns every layer of the AI stack, which no other frontier lab does:

• Models — the Gemini family (and open-weight Gemma), built by Google DeepMind.
• Chips — Google’s own Tensor Processing Units (TPUs). The seventh-generation Ironwood is built for large-scale inference, an eighth generation is previewed on TSMC’s 2nm process, and Google projects roughly 4.3 million TPUs shipped in 2026, rising sharply thereafter — making it the only credible alternative to Nvidia at scale.
• Cloud — Google Cloud and Vertex AI to train, host and sell the models.
• Distribution — Search (billions of users), Android, Chrome, YouTube, Pixel and Workspace.

This vertical integration lets Google control cost and supply end to end, and it is why rivals now buy Google’s compute (see the Anthropic deal below).

Models and the Gemini line

Google ships Gemini on a Pro/Flash cadence, plus a Deep Think reasoning mode and the open-weight Gemma line. Live benchmarks for each model render in the table below this page.

The generally available flagship is still Gemini 3.1 Pro: Gemini 4 Argon has not reached the public API or the Gemini app.

The Gemini 3.5 Pro delays

Gemini 3.5 Pro — targeting a 2M-token context window, Deep Think reasoning and frontier multimodal — was announced at Google I/O on 19 May 2026, when Sundar Pichai asked for “until next month”. It has since missed three targets: the original June date, a mid-July date, and it remained unreleased in early August. Google has not released it.

The reasons are technical and specific. Bloomberg reported Google was working to improve the model’s capabilities particularly in coding, with testers citing hallucinations and inconsistent outputs; Fortune’s account has engineers blaming a strategic failure to prioritise coding ability from summer 2025 onward, at exactly the moment agentic coding became the industry’s competitive battleground. When the Bloomberg report landed on 16 July, Alphabet fell 4.4% in a day — roughly $200 billion.

Google announced Gemini 4 Argon on 30 September 2026 without having shipped 3.5 Pro. The Decoder reports that 3.5 Pro was skipped; Google has not announced a cancellation.

Where Google’s models actually stand

Google’s strongest generally available model is Gemini 3.8 Flash (2 September 2026), its fourth Flash-tier release in under four months — a Flash-class workhorse rather than a flagship, and the cheapest model near the frontier at $0.75/$3.75 introductory pricing. It is the one place Google is clearly winning on price: Artificial Analysis measures $1.24 per completed index task, the lowest near the frontier. The structural problem stands on capability: it scores 41 on the re-based v4.3 Intelligence Index (59 at launch, before the 7 September re-basing), behind models from Anthropic, OpenAI, xAI and several Chinese labs. Gemini 4 Argon, announced on 30 September 2026 and generally available since 13 October, scores 53 on the same index, level with OpenAI’s GPT-6 Astra.

Google also ships Gemini 3 Deep Think (extended reasoning), the Gemma 4 open-weight models, and dedicated media models — Veo for video, Imagen for images, and the Nano Banana image models, where it remains genuinely competitive: Nano Banana Pro (Gemini 3 Pro Image, announced 20 November 2025) and Nano Banana 2 (Gemini 3.1 Flash Image), which became the default image model across the Gemini app, AI Mode and Lens on 26 February 2026 and reached enterprise general availability with Pro on 28 May 2026. Nano Banana 2.1 followed on the Gemini API on 6 October 2026 at $0.0336 per 1K image, about half Nano Banana 2’s price, and is on paid Google AI plans in the Gemini app. Gemini’s durable strengths are very large context windows and native multimodality. See best AI models for where the lineup ranks; vendor numbers are a ceiling and standardised leaderboards a floor.

Products and ecosystem

• Gemini app — the consumer assistant (web, Android, iOS), 900M+ monthly users, with free, Google AI Pro ($19.99/mo) and Google AI Ultra ($99.99/mo, cut from $249.99) tiers.
• AI in Search — AI Overviews reach about 2 billion monthly users, and AI Mode (a conversational, multimodal search experience) reached tens of millions of daily users — described by Google as its biggest Search change in 25 years.
• Gemini in Workspace — built into Gmail, Docs, Sheets, Meet and the rest of Workspace.
• Vertex AI and the Gemini API / Google AI Studio — the enterprise and developer platforms.
• Agents and tools — Project Astra (live multimodal assistant), Project Mariner (web agent), and NotebookLM.
• On-device and hardware — Gemini Nano on Pixel and Android, and deep Pixel integration.

Business and financials

Alphabet’s Q2 2026 results (reported 22 July) showed roughly 24% revenue growth, but the story was infrastructure at both ends of the income statement. Google Cloud grew 82% year on year to $24.8 billion — an acceleration from Q1’s ~63% — and backlog jumped $50 billion in a single quarter to $514 billion, with just over half expected to convert within 24 months (CNBC).

The cost side is why the shares fell on the print. Q2 capital expenditure hit a record $44.9 billion, and Alphabet raised full-year guidance to $195–205 billion from $180–190 billion, warning that capex will increase materially again in 2027. Roughly 60% of the spend goes to servers — mostly AI accelerators, its own TPUs and Nvidia GPUs. Search advertising remains the profit engine funding all of it.

Alphabet’s Q1 2026 comparison point: revenue of $109.9 billion, up 22%, with net income up about 81% (CNBC).

The pivot: from model maker to arms dealer

The most consequential strategic shift of 2026 is not a model — it is that Google started selling the picks and shovels, and that business is now growing faster than the one it was meant to support.

Google became a chip vendor. In Q2 2026 Google Cloud began delivering TPU systems into customers’ own data centres and recognising the revenue — a different business from renting cloud capacity, and one that puts Google in direct competition with Nvidia rather than merely reducing its dependence on it. Management expects only a small share of contracted TPU-system revenue to land in 2026, with the vast majority arriving in 2027, so the reported numbers understate the commitment already signed.

It sells to its own rivals. Google agreed in April 2026 to invest up to $40 billion in Anthropic in cash and compute, locking in around 5 gigawatts of TPU capacity and access to as many as one million seventh-generation Ironwood chips, and backstopping Anthropic’s lease payments — effectively underwriting a ~$35 billion chip commitment (CNBC, Bloomberg). A Broadcom-mediated supply line adds a further 3.5GW from 2027 on top of 1GW in 2026. Meta signed a separate TPU deal earlier in the year. Google now profits whether its own models win or its competitors’ do.

And that is the tension. Google is bringing well over a gigawatt of new AI compute online in 2026, yet its own DeepMind researchers report jockeying for access to chips being sold externally — Hassabis has acknowledged researchers “need a lot of chips to be able to experiment on new ideas at a big enough scale” (TNW). Pichai’s stated allocation priority is frontier AGI development first, but the fabric is not big enough for frontier research, external sales and production serving simultaneously. Departing researchers cite tightening internal compute access as a reason for leaving — which is how an infrastructure win becomes a modelling problem.

Whether this is a deliberate repositioning or an emergent consequence of demand, the financial gravity is clear: the compute business is compounding at 82% while the model business misses deadlines.

Leadership

As restructured in August 2026:

• Sundar Pichai — CEO of Alphabet and Google.
• Sergey Brin — co-founder; returned from retirement to lead AI development and commercialisation.
• Koray Kavukcuoglu — SVP running Google DeepMind day to day, reporting directly to Pichai.
• Demis Hassabis — chairman of Google DeepMind and chief scientist of Alphabet; co-founder and 2024 Nobel laureate (Chemistry, for AlphaFold). Stepped back from the CEO role on 5 August 2026.
• Ruth Porat — President and Chief Investment Officer (former CFO).

Departed in 2026: Jeff Dean (chief scientist) and Sanjay Ghemawat to found Discovery Loop with Oriol Vinyals and Quoc Le; Noam Shazeer (Gemini co-lead) to OpenAI; John Jumper (Nobel laureate) to Anthropic; David Silver to found Ineffable Intelligence.

Competition and market position

Google competes with OpenAI and Anthropic at the model frontier, with Microsoft in enterprise AI, and with Amazon and Microsoft in cloud. Its unique position is the full stack: it is the only rival that makes its own frontier models and its own chips and runs a hyperscale cloud and owns mass-market distribution. That lets it absorb AI into products billions already use — Search, Android, Workspace — rather than having to win new audiences.

On models, that position has weakened materially. Gemini remains strongest on long context and multimodal breadth, and Gemini 4 Argon, generally available since 13 October, matches GPT-6 Astra on the Artificial Analysis index while trailing Anthropic’s Claude Opus 5.5 and Sonnet 5.5. Google’s pressure points are rebuilding a research bench after a year of senior departures, converting enormous usage into rivals’ per-user monetisation, and the antitrust overhang.

The strategic question for 2027 is whether Google is a frontier lab that also sells compute, or a compute business that also ships models. Its financials increasingly suggest the second.

Controversies

• The Pentagon contract and the union bid. In April 2026 Google signed a Pentagon deal allowing the Department of Defense to run Gemini on classified networks. More than 580 Google employees, including senior DeepMind researchers, signed an open letter opposing it over autonomous-weapons and mass-surveillance concerns; AI-safety researcher Alex Turner resigned citing the deal, and research scientist Andreas Kirsch criticised it publicly. A union bid launched in May 2026 — the first formal unionisation attempt at a frontier AI lab. The dispute is a named factor in the 2026 talent exodus.
• Antitrust. Google faces the most aggressive US antitrust action since Microsoft in the 1990s. In the search-monopoly case it must end exclusive default-search agreements by mid-2026, and remedies have been extended to its generative-AI products to stop search-era tactics carrying into AI; a separate ad-tech case and a Chrome-divestiture fight are also live.
• AI Overviews accuracy. AI-generated answers in Search have drawn criticism for confident errors and for diverting traffic from publishers.
• Image-generation history. Google’s 2024 Gemini image generator was pulled after producing historically inaccurate images, a reputational episode it has worked to move past.
• Energy and capex. The ~$180–190 billion capex and the data-centre energy footprint draw scrutiny over cost and sustainability.

Recent developments

• Gemini 4 Argon announced (30 September 2026) — Google’s first new frontier model since Gemini 3.1 Pro, rolling out to Fairwind Program cyber defenders first. Artificial Analysis scores it 53 on its Intelligence Index, level with GPT-6 Astra, at an introductory $2/$10 per million tokens.
• AI leadership restructured (5–7 August 2026) — Hassabis stepped back to chairman and Alphabet chief scientist, Sergey Brin returned from retirement to drive commercialisation, Kavukcuoglu took day-to-day control, and Jeff Dean left after 27 years. Alphabet shares fell about 5%.
• Gemini 3.5 Pro missed a third target and remains unreleased; Gemini 4 entered pretraining. A Bloomberg report on the delay knocked 4.4% off Alphabet on 16 July.
• Q2 2026 earnings (22 July) — Google Cloud up 82% to $24.8 billion, backlog $514 billion, record $44.9 billion quarterly capex, and full-year capex guidance raised to $195–205 billion.
• Google began shipping TPU systems into customers’ own data centres, recognising chip-sale revenue for the first time — with most contracted revenue landing in 2027.
• Talent exodus accelerated — Shazeer to OpenAI, Jumper and two AlphaFold veterans to Anthropic (June), Dean and Ghemawat to Discovery Loop (August).
• Gemini 3.6 Flash and two sibling Flash models shipped in July, standing in for the delayed Pro.
• Pentagon contract backlash — 580+ employees signed an open letter; a union drive launched in May.
• $40 billion Anthropic deal (April) committed up to 5GW of TPU capacity and a million Ironwood chips to a direct model rival.
• Antitrust remedies began taking effect, including the end of exclusive default-search deals and AI-product provisions.

Where Google excels

• Full-stack control. Owns models, TPUs, cloud and distribution — no rival matches all four.
• Distribution. Search, Android, Chrome, YouTube and Workspace put Gemini in front of billions.
• Long context and multimodal. Gemini leads on very large context windows and native multimodality, with Veo and Imagen for media.
• Compute economics. Its own TPUs lower cost and reduce Nvidia dependence — and are now sold to rivals.

Where Google falls short

• Frontier position. Google’s strongest generally available model ranks behind Anthropic, OpenAI, xAI, Meta and several Chinese labs on independent intelligence benchmarks; Gemini 4 Argon, which scores level with GPT-6 Astra, is restricted to the Fairwind Program.
• Shipping cadence. Gemini 3.5 Pro missed three targets after May and was never released; Gemini 4 Argon was announced on 30 September 2026 with no public release date.
• Talent retention. A CEO, a chief scientist, a Gemini co-lead and a Nobel laureate all left in 2026, several to direct competitors.
• Coding. The specific gap behind the delays: Google under-prioritised coding ability from 2025, and Gemini trails badly on agentic coding benchmarks just as those became the industry’s main battleground.
• Internal compute contention. Researchers compete for TPU capacity being sold to external customers, including rivals.
• Monetisation per user. Huge reach has not yet converted to rivals’ per-user revenue, and AI risks cannibalising lucrative search ads.
• Legal overhang. Antitrust remedies now reaching AI products are a structural risk no rival faces to the same degree.

Developer resources

Google’s developer stack centres on the Gemini API via Google AI Studio (fast prototyping, free tier) and Vertex AI (enterprise deployment, tuning, governance), spanning the Gemini Pro/Flash models, Gemma open weights, and the Veo and Imagen media models. Models run on Google’s own TPUs, and Gemini is embedded across Workspace, Android (Gemini Nano) and Firebase. Pricing is on the Gemini API pricing page; Cloud status is at status.cloud.google.com.

Frequently asked questions

Who owns Gemini?

Alphabet Inc. owns it. Gemini is built by Google DeepMind, which is the AI division of Google LLC, which is a subsidiary of Alphabet (Nasdaq: GOOGL/GOOG). DeepMind stopped being a separate company in 2023, when Google merged the UK lab it acquired in 2014 with its internal Google Brain team. There is no separate DeepMind or Gemini stock — Alphabet shares are the only way to own any of it.

Is Gemini made by Google or DeepMind?

Both, because they are the same company. Google DeepMind is Google’s AI division and it builds the Gemini models. It was formed in 2023 by merging DeepMind and Google Brain, and sits inside Alphabet alongside the rest of Google.

Who runs Google DeepMind now?

Since the August 2026 restructure, Koray Kavukcuoglu runs Google DeepMind day to day as senior vice president, reporting directly to Sundar Pichai. Demis Hassabis, who had been CEO, stepped back on 5 August 2026 to become chairman of Google DeepMind and chief scientist of Alphabet, focusing on long-term research. Co-founder Sergey Brin returned from retirement to oversee AI development and commercialisation.

Why did Sergey Brin come back to Google?

Brin returned from retirement in August 2026 to lead Google’s AI development and product commercialisation, as part of a restructure the Financial Times reported was designed to accelerate Gemini monetisation and close the gap with competitors. The move shifts leadership emphasis from Demis Hassabis’s research-first approach toward commercial execution, after a year in which Gemini 3.5 Pro missed three release targets and Google fell behind rivals on independent benchmarks.

What is Google’s latest AI model?

Gemini 4 Argon, announced on 30 September 2026. It is rolling out first to cyber defenders in Google’s Fairwind Program, with paid API customers and Google AI Ultra subscribers to follow and no date given. The generally available flagship is still Gemini 3.1 Pro, and Gemini 3.8 Flash (2 September 2026) is the newest and strongest model on general release. Gemini 3.5 Pro, announced at Google I/O in May 2026, missed three release targets and was never released.

Is Google behind in AI?

On generally available models, yes. Google’s strongest generally available model ranks behind Anthropic, OpenAI, xAI, Meta and several Chinese labs on independent intelligence benchmarks, Gemini 3.5 Pro missed three targets and was never released, and it lost a DeepMind CEO, a chief scientist, a Gemini co-lead and a Nobel laureate during 2026. On infrastructure it is arguably ahead of everyone: Google Cloud grew 82% year on year in Q2 2026, it is the only credible alternative to Nvidia at scale with its own TPUs, and rivals including Anthropic and Meta buy its compute. Gemini 4 Argon, announced on 30 September 2026, scores level with OpenAI’s GPT-6 Astra on the Artificial Analysis Intelligence Index and five points behind Claude Opus 5.5, and has been generally available since 13 October.

Does Google make its own AI chips?

Yes. Google designs its own Tensor Processing Units (TPUs); the seventh-generation Ironwood is built for inference and an eighth generation is previewed on TSMC’s 2nm process. As of Q2 2026 Google does more than rent them out — it has begun shipping TPU systems into customers’ own data centres and recognising chip-sale revenue, putting it in direct competition with Nvidia. Rivals buy heavily: Anthropic contracted for up to 5 gigawatts and as many as a million Ironwood chips, and Meta signed a separate deal.

Is Gemini 3.5 Pro cancelled?

Google has not announced a cancellation, but it was never released. It missed three release targets after Sundar Pichai promised it “within a month” at Google I/O on 19 May 2026, with coding performance, hallucinations and inconsistent outputs reported as the reasons. On 30 September 2026 Google announced Gemini 4 Argon instead, and The Decoder reports that 3.5 Pro was skipped.

How many people use Gemini?

The Gemini app passed 900 million monthly active users by May 2026. Separately, AI Overviews in Google Search reach around 2 billion monthly users, and AI Mode reached tens of millions of daily users.

Why did Google invest in Anthropic?

In April 2026 Google committed up to $40 billion to Anthropic in cash and compute, locking in about 5 gigawatts of TPU capacity. It spreads Google’s AI bets and fills its data centres, even though Anthropic competes with Gemini — Google profits whether its own models or Anthropic’s win.

Is Gemini better than ChatGPT or Claude?

For most work, no. Gemini still leads on very large context windows and multimodal breadth, and its Workspace, Android and Search integration is unmatched — but Google’s strongest generally available model ranks behind both Claude and ChatGPT on independent intelligence benchmarks, and well behind Claude on coding specifically. Gemini’s free tier remains one of the most generous ways to use a capable model. See the best AI models ranking for the current standings.

Models

Apps

• Gemini