Google DeepMind

Google DeepMind is the combined AI research division of Alphabet, formed in 2023 by merging the original DeepMind (founded 2010, acquired by Google in 2014) with Google Brain. It develops the Gemini model family, foundational scientific AI such as AlphaFold, robotics models, and the Antigravity coding agent. It authored the Chinchilla paper, described as one of the two most important algorithmic innovations in AI.

Snapshot

Usage

Revenue and financials (Alphabet / Google Cloud)

Compute and capital

Models

Overview and competitive position

Google DeepMind develops Gemini as both a frontier model family and a set of consumer and enterprise products, alongside scientific and robotics research. According to the guide to AI in the agentic era, Gemini's chatbot website is "much less capable" than ChatGPT or Claude in harness and tool use despite comparable model quality (A Guide to Which AI to Use in the Agentic Era).

On the model frontier, Gemini 3 Pro took what Interconnects called a "definitive benchmark lead" in November 2025 but was, in the same source's words, "hailed as a false king"; within two months it had "effectively no impact at the frontier of coding agents" (Source: interconnects.ai). The division is also noted for strength in scientific AI (AlphaFold) and multimodal capabilities.

In a May 18, 2026 Big Technology essay, Alex Kantrowitz and Marty Swant argued that Google entered its May 19 I/O developer conference needing a credible answer to Anthropic's Claude Code and OpenAI's Codex, even as Google Cloud's 63% revenue growth last quarter and Alphabet's 136% one-year stock surge had insulated the company from immediate pressure; the essay framed the open question as whether Antigravity would graduate into a credible Claude Code / Codex peer or remain structurally behind for another quarter (Source: bigtechnology.com). On June 25, 2026 The Information reported that Google had begun reorganizing its recently launched AI coding strike team in an effort to close the gap with Anthropic, which it described as the most lucrative AI application category (Source: theinformation.com). See Coding AI Market Map.

Models and consumer products

Google's consumer assistant line began with Bard, announced by Sundar Pichai on February 6, 2023 as "an experimental conversational AI service, powered by LaMDA" and opened to trusted testers, then released to US and UK users on March 21, 2023. Both releases ran on "a lightweight and optimized version of LaMDA," a choice Google explained by compute cost: the smaller model "requires significantly less computing power, enabling us to scale to more users, allowing for more feedback." The February post bundled two further commitments with the announcement — that AI features distilling "complex information and multiple perspectives" would begin rolling out in Search, and that a Generative Language API would open to developers the following month — and named Google Cloud partnerships with Cohere, C3.ai and Anthropic, the last "just announced last week." The March general-access post stated Bard's limitations directly, including that models "can provide inaccurate, misleading or false information while presenting it confidently," and illustrated the point with an error in Google's own screenshot, where Bard gave the wrong scientific name for the ZZ plant (Bard launch announcements (Google, February–March 2023)). See Sycophancy and Hallucination.

The Gemini model family spans frontier (Pro), fast (Flash), reasoning (Thinking, Deep Think), and video (Omni Flash) tiers; benchmark and release details are in the Snapshot above. Gemini 3.5 Flash launched at I/O 2026 as the new default model in AI Mode worldwide, coding- and agentic-tuned. Search VP Elizabeth Reid described it as offering "sustained frontier performance for agents and coding." Early external measurements varied: Box CEO Aaron Levie reported roughly 20% gains over Gemini 3 Flash, while Wharton's Ethan Mollick said it was capable but "not as powerful as a full frontier model" (Source: nlp.elvissaravia.com). Gemini 3.5 Pro, slated at I/O for June 2026, had not shipped by mid-July: on July 16, 2026 Bloomberg reported the model was months behind schedule as Google worked to improve its coding performance, and Alphabet shares closed down about 4% on the report (Source: bloomberg.com; cnbc.com). Reuters reported on July 21 that the delay and investor concern over AI spending were the dominant questions heading into Alphabet's upcoming earnings report (Source: reuters.com). On July 23, 2026, current and former DeepMind employees told Axios the slippage stemmed partly from morale problems and departures — including Noam Shazeer's move to OpenAI and Nobel laureate John Jumper's to Anthropic, with some resignations tied to Google's April 2026 Pentagon deal (Source: axios.com). On August 6, 2026, SemiAnalysis reported that Google had silently cancelled Gemini 3.5 Pro and was promoting Gemini 4 in its place; the same analysis put Gemini annual recurring revenue at $12 billion in the second quarter of 2026 on its Tokenomics Model, and stated that Google Cloud already rents hundreds of thousands of TPUs to Anthropic, with many hundreds of thousands more committed to Anthropic and Meta over the next six quarters. The remainder of the analysis is paywalled and the cancellation is SemiAnalysis's reporting rather than a Google announcement (Source: newsletter.semianalysis.com).

Consumer and tooling products include:

• Gemini (gemini.google.com): consumer chatbot, bundled with Google nano banana (image generation), Veo 3.1 (video), and Guided Learning (tutoring) (Source: A Guide to Which AI to Use in the Agentic Era). In late April 2026 the Gemini chatbot added the ability to generate and edit native files (documents, spreadsheets, slides) directly in chat, framed as the consumer counterpart to ChatGPT's agentic file-editing push. Google announced on April 29, 2026 that the Gemini app can generate PDF, Word, Excel, Google Docs, Sheets, and Markdown files directly in chat for all users globally, producing structured-document outputs without leaving the chat surface and narrowing the gap with Workspace Intelligence and Microsoft Copilot (Source: blog.google).
• Antigravity: coding agent similar to Claude Code and OpenAI Codex, featuring an "Inbox" concept for asynchronous agent management (Source: oneusefulthing.org). At I/O 2026 Google released Antigravity 2.0 as its updated coding assistant, positioned as its answer to whether Antigravity could become a credible Claude Code / Codex peer.
• Gemini Notebook (formerly NotebookLM): knowledge-base tool for document analysis, podcasts, and mind maps (Source: A Guide to Which AI to Use in the Agentic Era). Google renamed NotebookLM to Gemini Notebook on July 16, 2026, citing more than 30 million users and 600,000 organizations, and gave every notebook a secure cloud computer that writes and executes code for data analysis — live for AI Ultra tiers and rolling out to Pro users (Source: blog.google).
• Google AI Plus subscription: On June 8, 2026 Google cut the monthly price of its Google AI Plus tier from $7.99 to $4.99 while doubling included storage from 200GB to 400GB, extending to U.S. consumers a budget price war that had begun in markets such as India and undercutting rivals on price as Anthropic — reporting a ~$47B run rate — had not introduced localized or budget pricing (Source: techcrunch.com). See Inference Economics and Token Pricing.
• Gemini Spark: an always-on, proactive Agentic AI personal agent that automates everyday tasks such as inbox summaries and event planning; cloud-hosted and pitched against OpenAI's OpenClaw. First announced at I/O 2026, Spark's rollout began May 30, 2026, running continuously rather than only on demand. Hands-on reviews published May 30 called the assistant useful in practice while questioning why Google split it into a separate product rather than folding the proactive capabilities into the main Gemini app (Source: techcrunch.com). On July 1, 2026 Google made Spark available on macOS as a beta for Google AI Ultra subscribers in the U.S., adding real-time tracking and integrations with applications including Google Tasks and Keep, positioning it against Microsoft's Copilot and Anthropic's Claude desktop app (Source: techcrunch.com).
• Gemini Omni / Video Remix in Google Photos: Google rolled out Gemini Omni, a model for creating and editing video through conversational prompts, powering a Video Remix option in Google Photos announced July 8, 2026 and promoted to Google One subscribers on July 14 (Source: socialmediatoday.com).
• Personalized image generation (Nano Banana): On June 29, 2026 Google made the Gemini app's Nano Banana–powered personalized image generation free for all eligible U.S. users (ages 13+, with editing limited to 18+), removing a paywall that had restricted the feature to Plus, Pro, and Ultra subscribers since April; the feature draws on Gmail, Google Photos, YouTube, and Search through the opt-in Personal Intelligence framework, and Google introduced a lighter Nano Banana 2 Lite model alongside it (Source: techcrunch.com; blog.google).

Nano Banana 2 in Google Earth and its withdrawal

Google paused the Nano Banana 2 image-generation feature in Google Earth on Friday, July 31, 2026, less than 48 hours after launching it, saying it had seen "people sharing screenshots of generated imagery that appear to violate our policies" and that it was rolling the feature back "while we work on implementing stronger guardrails" (Source: blog.google).

The withdrawal followed a BBC Verify report in which researcher Henk van Ess used the feature to generate a non-existent Iranian nuclear plant, a US–Mexico border refugee camp, and a Gaza hospital with an adjacent bomb crater. Van Ess said Google had allowed "invented" imagery to be "welded to genuine coordinates, drawn on genuine imagery" — the objection being to the pairing of synthetic content with an authoritative geospatial base layer rather than to image generation as such. Google said its AI outputs carry invisible watermarks checkable through Gemini or Lens; BBC Verify found those checks generally effective but circumventable, and found that some external AI-detection tools failed to flag the imagery (Source: bbc.com). See AI Content Provenance and Synthetic Media / Deepfakes.

• COSMO (experimental): on May 1, 2026 Google inadvertently published an experimental on-device assistant called COSMO to the Play Store before pulling it; the 1.13 GB app ran Gemini Nano locally and shipped 14 skills, including deep research, browser agents, and conversation summaries, positioned against Apple Intelligence and OpenAI's Codex agents (Source: 9to5google.com).

Google I/O 2026

Google's I/O 2026 spanned May 19 (a developer keynote in Mountain View) through a May 22, 2026 platform-and-AGI keynote, in which Pichai delivered 24 announcements in 35 minutes and named 2030 as Google's AGI target year — Google's first explicit company-wide AGI deadline. The 2030 framing aligns with Hassabis's April 30, 2026 "AGI by 2030" video-interview position, moving from an individual to a CEO-level institutional commitment. As reported AGI-timelines context, Anthropic and OpenAI have not declared comparable institutional AGI dates (Source: the-ai-corner.com). The May 22 keynote confirmed the updated Gemini 3.5 Flash benchmark of 1656 Elo on GDPval-AA alongside the 76.2% Terminal-Bench 2.1 figure, and reiterated that Gemini 3.5 Pro ships in June 2026.

Launches across the I/O week included Gemini 3.5 Flash and Gemini Omni Flash (covered above), the Gemini Spark agent, Antigravity 2.0, and:

• Managed Agents in the Gemini API — agents that run in isolated Linux sandboxes configured via AGENTS.md / SKILL.md files, paralleling Anthropic's managed-agents architecture and skill-file conventions.
• AI-agent-driven Search overhaul — background "information agents" that monitor the web for users, expanded agentic booking and calling, and generative-UI coding inside Search results. Search executive Liz Reid down-ranked the "ten blue links" interface in favor of AI-driven interactive experiences ("AI Search").
• Google + Blackstone "neocloud" company — disclosed May 19, with $5B Blackstone equity and Blackstone as majority owner; Google's vehicle for selling AI-infrastructure capacity to non-Google customers at neocloud scale.

Pichai said in the keynote that Gemini reached 900 million monthly users (about 2× year-over-year) and that AI Overviews in Search reached 2.5 billion monthly users (see Snapshot). On pricing, Google cut the AI Ultra plan to $200/month (from $250) and added a $100 developer tier, a repricing against OpenAI's and Anthropic's premium and developer tiers.

The day after I/O, on May 20, Google began testing AI-generated ad formats inside Search results and AI Mode: a "conversational discovery" format builds ads from a brand's own website text and imagery, while "direct offers" ads expand from AI Mode into standard Search. Search ads VP Shashi Thakur said the move follows OpenAI, which has tested ads in ChatGPT since February (Sources: reuters.com; techcrunch.com; blog.google; wsj.com).

Google's AI marketing itself drew criticism in July 2026: a 250th-anniversary commercial called "Declaration," aired July 4, imagined the Founding Fathers drafting the Declaration of Independence with help from Google Workspace AI, and prompted public pushback, including a July 5 assessment that the ad "should make Americans of every persuasion" uneasy (Source: techcrunch.com; theverge.com).

Enterprise and Workspace products

On April 22, 2026 Google responded to OpenAI's Workspace Agents launch with a Gemini Enterprise Agent Platform adding cross-app orchestration and durable long-running agents (Source: blog.google). Workspace Intelligence adds meeting summaries and proactive suggestions across Docs, Sheets, and Slides (Source: workspace.google.com), and Chrome was positioned as an "AI co-worker" offering inline summarization and task execution (Source: blog.google). Google Cloud CEO Thomas Kurian framed the next wave of cloud demand as the "agentic moment" in an April 24, 2026 Stratechery interview (Source: stratechery.com).

Sundar Pichai said 75% of new code at Google is now AI-generated, up from roughly 30% a year earlier (Source: theinformation.com). See AI Coding Agents, AI Deskilling.

Google reported that Gemini 3.6 Flash scores 49% on DeepSWE against 37% for Gemini 3.5 Flash, and 63.9% on MLE Bench against 49.7%, with the knowledge cutoff moved from January 2025 to March 2026. A limited-access Gemini 3.5 Flash Cyber pilot, offered to governments and trusted partners, powers the CodeMender vulnerability-repair system (Source: 9to5google.com).

Scientific AI, robotics, and research

The division's first widely cited result predates the merger: AlphaGo, published in Nature on 27 January 2016 by David Silver, Aja Huang and colleagues at what was then Google DeepMind, combined value and policy networks trained by supervised learning from human expert games and reinforcement learning from self-play with Monte Carlo tree search, reporting a 99.8% win rate against other Go programs and a 5–0 defeat of the human European champion — the first computer victory over a human professional at full-sized Go, "a feat previously thought to be at least a decade away" (Mastering the Game of Go with Deep Neural Networks and Tree Search (Silver et al., Nature, January 2016)).

Google DeepMind develops foundational scientific AI, with AlphaFold its best-known system. The dedicated AlphaFold team was dissolved over the year to July 2026 and most of the researchers behind the original paper reassigned, in an account that became public on July 29, 2026 and whose personnel moves DeepMind confirmed. Nearly a quarter of the original paper's authors who were full-time at DeepMind have left; those remaining moved to Gemini-related work, life sciences, fusion energy, or Alphabet's Isomorphic Labs. Pushmeet Kohli, who leads AI for Science, said the nine-year focus on concrete goals for individual grand challenges has since evolved. The AlphaFold Database, AlphaFold Server and academic AlphaFold 3 remain available (Source: xenospectrum.com). On May 25, 2026 it disclosed that its AlphaProof Nexus agent autonomously solved nine open Erdős problems — including two unsolved for 56 years — at a cost of a few hundred dollars per problem, by pairing an LLM with the Lean proof assistant to generate machine-verified proofs across combinatorics and graph theory. The system also proved 44 open conjectures from the Online Encyclopedia of Integer Sequences (OEIS). The disclosure came one day after OpenAI announced its own Erdős-conjecture disproof, the first head-to-head between the two labs on machine-verified open-mathematics-problem solving. The Lean-pairing pattern (LLM proposes, formal verifier checks) is described as a template for autonomous-research deployments, with verification cost identified as the binding rate-limiter on autonomous mathematics rather than proof-search cost (Source: arxiv.org). See AI-for-science.

In robotics, Google DeepMind released Gemini Robotics-ER 1.6 on April 27, 2026 — a reasoning-first robotics model that achieves 93% success on instrument-reading tasks via agentic vision combining visual reasoning with code execution, developed through partnership with Boston Dynamics. It is described as the first frontier-lab robotics-VLA combining advanced reasoning with explicit code-execution tool use, positioning Google DeepMind in embodied-AI competition with Figure AI, Skild Ai, and AEON Robotics (Source: deepmind.google).

On July 30, 2026 Google DeepMind introduced Gemini Robotics 2, comprising a vision-language-action model, the Gemini Robotics ER 2 embodied-reasoning model, and Gemini Robotics On-Device 2. Controlling Apptronik's Apollo 2 humanoid fitted with Inspire hands, it reported success rates of 68.4% picking up objects from a table, 76.3% from a shelf and 45.7% from the floor; multi-finger tasks using SharpaWave hands ranged from 32% to 92%. DeepMind said On-Device 2 adapts to new bi-arm embodiments with typically fewer than 200 examples, and introduced ASIMOV-Agentic, a benchmark for agentic safety orchestration (Source: deepmind.google).

In healthcare research, Google DeepMind disclosed on May 1, 2026 an effort pairing Gemini with Project Astra into a multimodal "AI Co-Clinician" that handles telehealth calls and guides live physical exams; in tests it matched primary-care doctors on 68 of 140 consultation skills (Source: deepmind.google). See Healthcare — AI Deployment, AI Mental Health and Psychological Harm.

On AI consciousness, Google DeepMind researcher Alexander Lerchner published a paper (April 27, 2026) arguing that LLMs are structurally incapable of consciousness, framing them as embodiment-lacking simulators. The paper, held in the wiki as The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness — Alexander Lerchner (Google DeepMind, March 2026), is positioned as a structural counterpoint to Bostrom's computational-functionalism argument, with the two anchoring opposing ontological positions in the AI welfare and consciousness debate (Source: 404media.co). See AI Welfare / Model Welfare / Moral Patienthood.

Compute infrastructure and TPUs

On the April 29, 2026 earnings call, confirmed April 30, Pichai said Google will sell its custom Tensor Processing Unit (TPU) chips to customers for deployment inside customers' own data centers, with Boston Dynamics confirmed as a current TPU customer. CFO Anat Ashkenazi said TPU hardware sales already contributed to Google Cloud's $460B backlog and will produce some revenue in 2026 and more in 2027. A Google Cloud spokesperson clarified that Pichai's earnings reference to Thinking Machines Lab and Hudson River Trading as TPU customers was incorrect — they are Google Cloud customers but not TPU customers. The shift converts an internal-Google compute advantage into a chip-vendor revenue line competing directly with Nvidia (Source: theinformation.com). See Semiconductor Supply Chain, Nvidia & TSMC — AI Compute Infrastructure. Google unveiled the TPU 8 generation on April 22, 2026 (Snapshot).

Google is designing a server chip internally dubbed "Frozen v2," slated for 2028, that integrates the Gemini model blueprint directly into silicon and could be six to ten times more power-efficient than its current AI chips measured in tokens per unit of power; the plan became public on July 20, 2026 and lifted Alphabet shares about 3% (Source: techcrunch.com; theinformation.com; theinformation.com).

Google is in talks with Samsung to manufacture part of its next-generation "Icefish" TPU using Samsung's 2-nanometer process — with TSMC making the main compute die and MediaTek on design, and mass production possible as early as 2028 — per a June 11, 2026 report that follows June 8 reporting on Google–Intel foundry talks, signaling an effort to diversify TPU manufacturing beyond a single foundry (Source: reuters.com). See Semiconductor Supply Chain.

On June 18, 2026 the Wall Street Journal reported that Alphabet had provided a $3.2 billion financial guarantee for Lake Mariner, an AI data-center cluster on the southern shore of Lake Ontario in western New York whose developers will rent computing power from thousands of Google's processors to Anthropic. The financing strategy was described as mirroring the approach Nvidia has used to stoke demand for its own GPUs, with Google additionally moving to sell its TPUs to "select" external customers and weighing increased financial support for data-center partners — broadening what observers called the most formidable challenge yet to Nvidia's hold on AI chips (Source: wsj.com; wsj.com). See Circular Financing in AI, Nvidia & TSMC — AI Compute Infrastructure.

In June 2026, Meta signed a multibillion-dollar deal to rent Google TPUs for its data centers from 2027 and is weighing a large equity raise on the back of it, extending Google's TPU-as-a-service push beyond Google Cloud's own workloads (Sources: finance.yahoo.com; cnbc.com). Separately, Google agreed to pay SpaceX roughly $920M/month from October 2026 through June 2029 for about 110,000 Nvidia GPUs at xAI data centers to backstop Gemini Enterprise demand — a hyperscaler renting frontier compute from a model-layer rival (Source: cnbc.com); tracked in full on xAI. See Circular Financing in AI.

The TPU sales campaign extended to "neoclouds" — specialist AI-cloud providers that have historically been Nvidia's most loyal customers. Per reporting made public July 13, 2026, Google has pitched TPUs to neoclouds including Nscale (whose largest preferred shareholder late in 2025 was Nvidia), offered to financially backstop TPU data centers, and some Google managers project TPUs could take 10 percent of Nvidia's AI server chip business (Source: The Information, "Behind Google's Ground War to Lure Nvidia's Most Loyal Customers," July 13, 2026). See Nvidia & TSMC — AI Compute Infrastructure, Circular Financing in AI.

On the energy side, Google announced during the week of July 13, 2026 its largest solar-and-battery purchase to date: the Steel River Energy Center in Arkansas, adding 1 gigawatt of solar and 1.9 gigawatt-hours of battery storage in its first two phases (backed by $3.5 billion in secured financing) and scaling to about 1.8 gigawatts by 2029, which would make it the largest solar facility in the United States (Source: techcrunch.com). Researcher estimates published July 16, 2026 put Google's electricity consumption up more than 140% between 2021 and 2025 (Source: axios.com). See AI Data Centers.

Google has also contracted for network capacity between data centers. Verizon CEO Dan Schulman disclosed on the company's July 24, 2026 post-earnings call a dark-fiber agreement with Google worth more than $1 billion, connecting Google data centers and separate from the companies' existing AI, cloud, and 5G work. Schulman said further deals expected by year end are "worth multiple billions of dollars in revenue over the next several years" (Source: blockspace.media).

Funding and Alphabet financials

Alphabet reported Q1 2026 revenue up 22% year-over-year to $109.9B, with Google Cloud revenue up 63% to $20B and the Google Cloud backlog roughly doubling to $460B; Google Cloud is now about 18% of Alphabet's business (Sources: s206.q4cdn.com; theinformation.com). Follow-on reporting (May 4) showed net income up 81% year-over-year and 2026 capex raised to $180–190B from a prior $170B guide. Anthropic's February $380B funding round drove a $73.6B remeasurement of Google's non-marketable equity stake, contributing materially to the net-income beat. Fortune's Eva Roytburg had reported on April 30 that roughly half of Google's "blowout AI profits" derived from its Anthropic equity stake rather than core operations, a structural dependence tying Alphabet's reported AI-margin numbers to Anthropic's continued growth (Sources: fortune.com; stratechery.com; wsj.com). See Anthropic, AI Bubble Debate.

Alphabet's stock rose 10% on April 30, 2026 and 34% over April, its best month since 2004, following the Q1 results and the TPU-sales announcement (Source: cnbc.com).

Alphabet said on June 1, 2026 it plans to raise up to $80 billion in equity to fund its AI ambitions, structured as $30B in underwritten public offerings (split between mandatory convertible preferred stock and common stock), a $40B at-the-market program expected to begin in Q3 2026, and a $10B private investment from Berkshire Hathaway, which had been building an Alphabet stake since Q3 2025. Proceeds are earmarked for "capital expenditures to scale AI infrastructure and global compute," extending an infrastructure-financing push that had already made Alphabet the first modern company to issue a 100-year bond. Pichai framed the spend as asymmetric, saying "the risk of under-investing is dramatically greater than the risk of over-investing," as major hyperscalers head toward $750B+ combined capex this year (Sources: axios.com; reuters.com; cnbc.com). On June 3 Alphabet sold $45B of equity in an oversubscribed offering (Berkshire Hathaway taking $10B), with another $40B planned the next quarter for a record $85B total, earmarked for 2026 AI capital spending of $180–190B. Goldman Sachs' international head described the sale as putting Wall Street in "unprecedented territory" (Sources: techcrunch.com; cnbc.com). See AI Bubble vs. Buildout — Synthesis, AI Bubble Debate.

Partnerships and distribution

Google committed up to $40 billion to Anthropic on April 24, 2026 at a $350B valuation — $10B immediate, up to $30B contingent on performance milestones, plus 5 gigawatts of compute online in 2027 (Source: bloomberg.com). The Information reported on May 5 that Anthropic has committed approximately $200 billion to Google's cloud and chips over five years — more than 40% of Google's recently disclosed cloud revenue backlog — under the previously announced 5-gigawatt deal, making Anthropic the single largest external customer of TPU capacity and the bulk of Google's reported cloud-backlog growth (Source: theinformation.com). See Clawed.

On model distribution, The Information reported May 6 that Alphabet is in talks with Blackstone, KKR, and EQT to give the firms' portfolio companies direct access to Google's AI models, mirroring the Anthropic and OpenAI $1.5B / $10B-tier PE consulting JVs but without a separate consulting business; the deals follow a Google Cloud–Vista Equity Partners distribution partnership in April 2026 (Source: theinformation.com).

Apple remained on track as of June 4, 2026 to launch its overhauled Siri in September, running partly on Google Cloud's Nvidia Blackwell B200 chips with hardware-based confidential compute and falling back to Gemini models for cloud queries, expanding Google's frontier-model distribution into Apple's installed base (Sources: theinformation.com; macrumors.com). Meta's TPU rental deal and the SpaceX/xAI compute purchase are covered under Compute infrastructure above.

Distribution has also run into capacity limits. The Financial Times reported on July 4, 2026 that Google moved to cap Meta's use of its Gemini models as surging AI demand made computing power the industry's scarcest commodity (Source: ft.com). On the data-supply side, reporting on July 22, 2026 that Reddit might cut Google's access to its content for AI training sent Reddit shares down 8.3% (Source: cnbc.com). See Training Data Walls.

General Motors announced on April 30, 2026 that it will deploy Google Gemini into approximately 4 million vehicles, one of the largest single AI-OEM partnerships announced to date (Source: thenextweb.com). See Embodied AI vs AGI (China's Race Redirect), Transportation — AI Deployment.

Military and government contracting

Google has expanded its military and government AI footprint while facing internal opposition. On April 22, 2026 it finalized a classified Pentagon AI deal reportedly worth more than $400M over three years (Source: washingtonpost.com). Google then formally signed an agreement with the Department of Defense allowing the Pentagon to use its AI models on classified work for "any lawful government purpose" — the same language used in OpenAI's earlier Pentagon deal. More than 600 Google employees signed an April 27 letter demanding Pichai reject the deal; the agreement was nonetheless executed on April 28, 2026. Google proposed contract language excluding lethal autonomous weapons and mass surveillance, but civil-liberties lawyers said the language is insufficient to prevent those uses in practice. The April 28 DoD disclosure confirmed Google's latest model is now on GenAI.mil alongside the existing 100,000 AI agents users have built (Sources: theinformation.com; defenseone.com).

Transformer News reported on April 30, 2026 that the "any lawful governmental purpose" agreement blindsided Google's own AI staff: the 600-plus employee letter to Pichai protested the negotiations, and more than 100 employees separately urged Google chief scientist Jeff Dean to refuse any deal crossing red lines on domestic mass surveillance and autonomous weapons. Institute for Law and AI senior researcher Charlie Bullock noted on X that the contract's "should not be used for" language imposes no enforceable obligation on the DoD, and that Google appears obligated to remove technical safeguards if they impede a "lawful purpose." DeepMind research scientists Andreas Kirsch and Alex Turner publicly criticized the agreement, with Kirsch saying he felt "incredibly ashamed" to be a Google DeepMind senior research scientist (Source: transformernews.ai). By July 16, 2026 Turner had resigned from Google DeepMind over the Pentagon deal, which he said violates Google's 2018 pledge against lethal autonomous weapons (Source: controlai.news). Turner, an AI safety researcher, published his reasons on July 15, 2026; he had organized a petition signed by more than 250 DeepMind employees and submitted a 25-page alternative framework to Demis Hassabis requiring human control over targeting (Source: storyboard18.com). In a first-person essay published July 21, 2026 in Transformer, Turner — who writes that he left the company on June 9, 2026 — recounted the internal campaign in detail: he states that the Pentagon pressed AI companies to hand over models for "all lawful use," that Anthropic refused and was labeled "a supply chain risk" (see Anthropic v. United States (Pentagon ban challenge)), that OpenAI signed a deal analysts say contains wide loopholes, and that Google signed a classified deal reportedly carrying even fewer restrictions, without binding concessions to ethics-focused employees (I tried to stop Google DeepMind's Pentagon deal. Then I quit. (Alex Turner, July 2026)). The essay adds that Turner sent his 25-page proposal to Hassabis, who routed it to senior policy staff, and that after an initial review by junior staff the senior staff did not respond; that he learned of the signing at 11:45 pm through a Signal group because "Google never announced the deal internally"; and that he sought unsuccessfully to persuade Dean — whom he identifies as holding the most leverage, and who had tweeted against lethal autonomous weapons and domestic mass surveillance and co-signed an amicus brief backing Anthropic — to threaten resignation. Turner's stated conclusion is that pledge-based and seat-at-the-table governance failed at Google's clearest red line, and that "we need structures: binding contracts, independent auditors, and review bodies that cannot be quietly dissolved" and eventually legislation. He also charges that the 2025 principles revision is irreconcilable with Hassabis's subsequent statement that "nothing's changed about our principles," and that external advocates including Stuart Russell and signatories of the Global Call for AI Red Lines stayed publicly silent when the counterparty was the US government.

On the same day it signed the classified deal, Google withdrew from a $100M Pentagon voice-controlled drone-swarm contest, citing an internal ethics review, after having advanced through earlier rounds. The split posture — taking the classified-AI deal while exiting an autonomous-targeting-adjacent procurement — reflects labs drawing internal lines on military procurement category by category (Source: bloomberg.com). See Autonomous Weapons, DOD — Department of Defense (AI Deployer).

Pentagon Chief Digital and AI Officer Cameron Stanley confirmed on April 28, 2026 that the DoD is expanding its use of Gemini, saying "overreliance on one vendor is never a good thing"; the move reads against the Anthropic supply-chain-risk designation and the administration's willingness to onboard frontier-lab models bilaterally rather than through a single vendor (Source: cnbc.com; see Anthropic). Pentagon-contract employee backlash continued through April 30, with Google publicly defending the work (Source: theinformation.com).

On May 1, 2026 Google was named one of eight AI vendors signed by the U.S. Department of War to deploy frontier AI on classified networks, alongside SpaceX, OpenAI, Nvidia, Reflection, Microsoft, AWS, and Oracle; the deal extends Google's military AI footprint to direct classified-network frontier-model deployment (Source: bloomberg.com). The classified-networks deal was clarified to specifically cover IL6 and IL7 networks. During the week of April 27, Google employees sent an internal letter urging leadership to refuse classified workloads, arguing that "the only way to guarantee that Google does not become associated with such harms is to reject any classified workloads," building on the existing 600-plus employee petition (Source: war.gov). See DOD — Department of Defense (AI Deployer).

Safety approach

Google DeepMind maintains a Frontier Safety Framework with tiered capability thresholds (Source: Managing Advanced Cyber Risks in Frontier AI Frameworks). Its cyber uplift threshold is defined as providing "sufficient uplift with high impact cyber attacks for additional expected harm at severe scale." Demis Hassabis acknowledged at Davos that closing the AI self-improvement loop is something all labs are actively working on (Source: The Shape of the Thing).

On April 30, 2026 in a video interview, Hassabis said he expects AGI by 2030 and characterized reasoning and persistent memory as the principal missing capabilities, with current AI agents "in early innings" (Source: youtu.be). See AGI Timelines.

On May 5, 2026 Google joined Microsoft and xAI in agreeing to give the U.S. Center for AI Standards and Innovation (CAISI) early access to evaluate frontier models for national-security risks before public release, joining OpenAI and Anthropic (Source: bloomberg.com). See AI Pre-Release Vetting, NIST CAISI (Center for AI Standards and Innovation).

On June 20, 2026 Google DeepMind released an AI Control Roadmap, setting out a research agenda for constraining advanced agents even when their alignment cannot be assumed (Source: nlp.elvissaravia.com). See AI Control.

In a report published July 20, 2026, Google DeepMind and Isomorphic Labs detailed their joint AI biosecurity work — an area congressional researchers say is largely uncovered by regulation — and called for "stringent measures" including a federal frontier AI safety framework (Source: insideaipolicy.com). See AI Biosecurity.

On 10 August 2026 researchers disclosed that the encrypted reasoning blocks returned by the Gemini API are interchangeable across sessions, users and models, with the paper's July 2026 compatibility matrix finding any Gemini model's traces replayable into any other. The authors used Gemini Robotics 1.6 as the decoder, chosen because it processes traces from both the 2.5 and 3.x series, and demonstrated recovery of hidden reasoning from stronger Gemini models in plaintext. Google acknowledged the disclosure and the researchers report they were subsequently unable to launch the same attacks (Stealing Reasoning Traces from Proprietary LLM APIs).

Labor and workforce

Google DeepMind's UK staff voted in April 2026 to seek joint recognition by the Communication Workers Union and Unite, the Guardian reported on May 5. The recognition letter, formally delivered to management on May 5, would represent at least 1,000 workers and make DeepMind the first frontier AI lab to unionize, with workers citing the Pentagon's classified-AI deal, the Iran war, the Anthropic feud, and Project Nimbus as drivers (Source: theguardian.com). See AI Labor Disruption.

On the broader jobs debate, Google SVP James Manyika argued in a May 19 Platformer interview that doomer forecasts of mass white-collar job loss are wrong, noting that under 10% of occupations have 90%-or-more automatable tasks, and offered to bet against anyone predicting 50% job loss within two years. His position pairs with Acemoglu's May 11 three-signals interview as a prominent skeptic anchor in the May 2026 jobs-apocalypse debate (Source: platformer.news).

On July 23, 2026 Google published ATLAS, an analysis of 14.65 million de-identified Gemini interactions from April 2026: AI use touched occupations representing 90% of U.S. employment, but the average worker used it for only 21% of tasks; under 10% of interactions aimed at automating non-routine cognitive work; 86% of conversations were non-work; and a 1% increase in an occupation's median earnings was associated with 2.68% higher AI use (Source: axios.com). See AI Labor Disruption.

Regulation

In late June 2026 Google published a policy paper proposing FARO, an industry-composed, federally overseen frontier AI regulatory organization intended to complement the administration's NSA pre-release review process — a process many observers have described as a "de facto licensing regime" — while separating frontier-model oversight from application-layer rules; the same paper restated Google's position that AI training on copyrighted works is fair use, with legal accountability focused on outputs (Source: insideaipolicy.com; digitalmusicnews.com). See AI Pre-Release Vetting.

On July 14, 2026 CEO Demis Hassabis published "A Framework for Frontier AI and the Dawning of a New Age," proposing a U.S. AI standards body modeled on the financial industry's FINRA: frontier labs would initially share models voluntarily up to 30 days before release for safety testing of dangerous cyber, biological, and deception capabilities, with passage later becoming mandatory for U.S.-market deployment of all frontier-class models, open or closed, regardless of country of origin (A Framework for Frontier AI and the Dawning of a New Age (Hassabis, July 2026); Source: axios.com). Hassabis said he had briefed the Trump administration, fellow lab leaders, and European officials, wanted the industry-funded body — with a majority-independent board — operational "before year-end," and called the administration's June 2026 export-control freeze of Anthropic's Mythos and Fable models "a bit of a wake-up call" (Source: axios.com; theverge.com). By July 17, 2026 the proposal — described as a Frontier AI Standards Body that would audit frontier models for cybersecurity and biological-threat capabilities before release — had drawn public endorsements from Sam Altman and Elon Musk, and praise from Anthropic's Jack Clark, ahead of planned meetings in Washington; it followed similar recent proposals from Altman and Dario Amodei (A Framework for Frontier AI and the Dawning of a New Age (Hassabis, July 2026); Source: platformer.news). Support continued to build through July 21, 2026, with endorsements from Mustafa Suleyman, Satya Nadella, and Jack Dorsey, while Dario Amodei said he favors an FAA-style federal agency instead (A Framework for Frontier AI and the Dawning of a New Age (Hassabis, July 2026); Source: axios.com). See AI Pre-Release Vetting and Demis Hassabis.

On July 14, 2026, publishers Hachette Book Group, Elsevier, and Cengage Learning, with novelist Scott Turow, were reported to have filed a class-action copyright suit against Google in the Southern District of New York (filed July 10), alleging Gemini models were trained on millions of copied books and journal articles — including works submitted under the earlier Google Books settlement — and citing an internal Google document acknowledging "$10Bs-$100Bs" in potential fines (Source: publishersweekly.com; thewrap.com). See Hachette et al. v. Google (Gemini training data).

The UK Competition and Markets Authority ruled on June 3, 2026 that Google must let publishers opt out of AI Search features — including AI Overviews — without losing regular search placement. The ruling responds to publisher complaints that AI Overviews cannibalize referral traffic while leaving sites no way to withhold content from AI summaries short of de-indexing entirely (Source: theverge.com).

In a separate liability ruling, the Regional Court of Munich held on May 28, 2026 (case no. 26 O 869/26, made public June 9, 2026) that Google is directly liable for false claims in its AI Overviews, treating the AI-generated summaries as Google's own statements rather than protected third-party search results, rejecting Google's argument that users could verify the answers themselves, and ordering Google to pay 80% of the legal costs after its AI falsely tied two Munich publishers to scams and subscription traps (Source: the-decoder.com). Gary Marcus argued in a June 10, 2026 essay that the ruling's logic may carry to US law because Section 230 shields platforms only for third-party speech, not for what their own software generates — a reading he said AI-law scholar Ryan Calo also hinted at the same day (Source: garymarcus.substack.com). See AI in Journalism and Media. German media regulation reached AI search directly on July 14, 2026, when Germany's Commission for Licensing and Supervision issued its first rulings holding AI search engines and chatbots — including Google's AI Overviews and Perplexity — subject to German media law (Source: reuters.com).

Under the EU Digital Markets Act, the European Commission ordered Google to share anonymized search data with rivals beginning January 2027, a decision reported July 21, 2026 alongside continuing European liability pressure on AI Overviews (Source: aigovernancelead.substack.com). On July 23, 2026 the Commission fined Google €890 million in two DMA non-compliance decisions — €460 million for self-preferencing its own shopping, hotels, transport, and sports services in Search results, and €430 million for restricting app developers from steering users to cheaper offers outside Google Play — and ordered Google to end the non-compliance, with press reports indicating search-results changes within 60 days (Source: digital-markets-act.ec.europa.eu; cnbc.com; searchengineland.com). In a separate DMA specification decision dated July 16, 2026, the Commission required Google to open 11 Android features — including hotword invocation, screen automation, and on-device Gemini Nano models — to third-party AI services free of charge, starting with Android 18 by August 2027 (Source: digital-markets-act.ec.europa.eu).

In a non-AI antitrust matter bearing on Google's platform position, the European Court of Justice on July 2, 2026 upheld Google's €4.1 billion ($4.67 billion) Android antitrust fine, dismissing the company's final appeal in a case dating to the European Commission's 2018 penalty (reduced from €4.34 billion in 2022); Google said the judgment "fails to recognize our significant investment to ensure Android remains open, interoperable and free" (Source: cnbc.com). A Swedish court a day earlier ordered Google to pay about $1.5 billion in damages to Klarna-owned price-comparison service PriceRunner (Source: reuters.com).

Key people

• Koray Kavukcuoglu (senior vice president leading Google DeepMind, from August 2026): previously the unit's chief technology officer and Google's chief AI architect. Reports to Sundar Pichai and oversees Gemini model development.
• Demis Hassabis (chair of Google DeepMind and chief scientist of Alphabet, from August 2026; chief executive of the unit 2023–2026): co-founder of the original DeepMind; Nobel Prize in Chemistry 2024 (AlphaFold).
• John Jumper: co-architect of AlphaFold 2; co-laureate of the 2024 Nobel Prize in Chemistry. Announced on June 19, 2026 that he would leave Google DeepMind after nearly nine years to join Anthropic. See John Jumper.
• Jeff Dean: Google chief scientist until August 5, 2026, when he left after 27 years to co-found Discovery Loop.

August 2026 leadership change

Sundar Pichai told Google DeepMind teams on August 5, 2026 that Hassabis would move to a strategic role as chair of Google DeepMind and chief scientist of Alphabet, with Kavukcuoglu taking over daily operations as senior vice president rather than holding a stand-alone chief executive title. Hassabis wrote that "it's critical that we collectively get the next steps right to ensure this all goes well for humanity," said he would spend more time at Isomorphic Labs, and added that he has "always believed the No.1 application of AI should be to improve human health." He named the unreleased Gemini 4 in describing the company's progress (Source: blog.google; fortune.com).

Jeff Dean left the same day, together with Google senior fellow Sanjay Ghemawat, Google DeepMind senior research scientist Oriol Vinyals, and Google Brain founding member Quoc Le, to found Discovery Loop, a public benefit corporation aimed at automating experimental research loops. Alphabet is a founding investor and supplies the startup's first-year compute (Source: techcrunch.com). Alphabet shares fell around 5% on the combined announcements, which landed while Gemini 3.5 Pro was months behind its original June launch target (Source: fortune.com).

Reuters published a fuller account of the sequence on August 12, 2026. It reported that Sergey Brin had in recent months urged key Google AI staff to go all in on Gemini and had pushed resource allocation toward recursive self-improvement, according to people familiar with his remarks, and that by August 2026 Google had delayed the new flagship Gemini version by two months after internal testing showed it continued to lag rivals in areas such as coding. The account records Google's stock falling 4% on the day of the August 5 overhaul, the departures of Gemini's original technical co-leads Jeff Dean and Oriol Vinyals, and an August 6 all-hands at which the company disclosed that several nontechnical teams were moving out of DeepMind into corporate Google (Source: reuters.com). The two-month delay is reported against the flagship model then in development; it is not stated whether this is the same slippage as the Gemini 3.5 Pro schedule described above, which SemiAnalysis reported on August 6 as a cancellation in favour of Gemini 4. The share-move figure of 4% here and the "around 5%" recorded above come from different accounts of the same day and are left unreconciled. See Recursive Self-Improvement (RSI).

A Fortune account published August 10, 2026 and sourced to three current and one former DeepMind employee assembles the schedule slippage and the departures into a single sequence. On its telling, Gemini 3.5 Pro — unveiled at I/O in May 2026 with a promise to ship "next month" — missed that June target and a subsequent mid-July date and entered August unreleased, while Google shipped a set of smaller Flash models and said it had begun training Gemini 4. Micah Hill-Smith, cofounder and chief executive of Artificial Analysis, said Gemini 3.6 Flash, then the strongest model Google had shipped, sat behind models from Anthropic, OpenAI, one or two leading Chinese labs, xAI and Meta on his firm's intelligence index. The account records that AlphaFold researchers John Jumper, Jonas Adler and Alexander Pritzel left for Anthropic, and that Gemini co-lead Noam Shazeer went to OpenAI in a single week in June. A Google spokesperson disputed that power was shifting to Mountain View, said Gemini 3.5 Pro was in testing, and acknowledged the company needed to improve on coding and agentic capabilities (Source: fortune.com).

Hill-Smith's ranking is stated as of the article's August 10 publication. Google released Gemini 3.7 Flash on August 13, 2026, three days later, so the comparison describes the position before that release rather than the current one. The Fortune account also describes Gemini 3.5 Pro as unreleased and in testing, where SemiAnalysis had reported on August 6 that it was silently cancelled in favour of Gemini 4; the Google spokesperson's "in testing" and the cancellation report are not reconciled by any source.

Tim O'Reilly, founder of O'Reilly Media, offered a reading of the reorganisation on August 8, 2026 in which it is a bet on diffusion over frontier leadership rather than a retreat, comparing it to Westinghouse's alternating-current bet against Edison and drawing on Jeff Ding's Technology and the Rise of Great Powers. He cited Alphabet's $24.8 billion of cloud revenue in the latest quarter, up 82 percent year over year against 37 percent at AWS and 43 percent at Microsoft's Azure and other cloud services, while noting the figures are not strictly comparable (Source: asimovaddendum.substack.com). The reading is an outside interpretation; neither Alphabet nor the departing researchers have characterised the move in those terms.

The departures extend a pattern the unit has recorded since mid-2026. Jumper's departure was reported days after Gemini co-lead Noam Shazeer left Google DeepMind for OpenAI, framed in coverage as an intensifying talent war among frontier labs as Anthropic and OpenAI approached public listings (Source: cnbc.com).

AGI and superintelligence positions

Two of the lab’s senior figures anchor its published position on the transition beyond human-level AI. Shane Legg, co-founder and Chief AGI Scientist, is the originator with Marcus Hutter of the Legg-Hutter measure of universal intelligence, which supplies the formal grounding for the AGI, ASI, and Universal AI definitions in From AGI to ASI (June 2026). Allan Dafoe, director of frontier safety and governance and previously founder of the Centre for the Governance of AI, is a co-author of the same report, whose treatment of deliberate slowdown as one of six bottlenecks records the governance-side judgment that "economic and political pressures, and international race-dynamics may override slowdown pressures, particularly in light of lacking global coordination and effective global oversight and enforcement."

Spinoffs and subsidiaries

• Isomorphic Labs (London, founded 2021): AI-for-drug-discovery biotech. Hassabis is CEO of both companies. Isomorphic holds commercial rights to AlphaFold 3 outputs and operates the proprietary IsoDDE drug-design engine. First clinical trials of AI-designed drugs were described as imminent by Max Jaderberg at WIRED Health on April 16, 2026 (Source: wired.com).

Related pages

Key sources and related concepts: A Guide to Which AI to Use in the Agentic Era, Training Compute-Optimal Large Language Models (Chinchilla), Managing Advanced Cyber Risks in Frontier AI Frameworks, Stanford HAI AI Index Report 2026, and (Source: oneusefulthing.org).