Google

Google is Alphabet’s largest business, combining Search, consumer platforms and Google Cloud with AI research and the Gemini model family. Its AI work spans model architectures, custom Tensor Processing Units (TPUs), developer tools and deployment in everyday products. Google combined its Brain team with DeepMind in 2023; by June 2026, Sundar Pichai reported that Gemini powered all 13 Google products with more than a billion users.[1][11][12]

In April 2023, Google brought the Brain team from Google Research and DeepMind together as Google DeepMind. Demis Hassabis became its CEO, while Jeff Dean became Google’s Chief Scientist, serving Google Research and Google DeepMind. Pichai presented the combination as a way to accelerate development of more capable AI systems. Google Research continued work across computer science, including algorithms, privacy, security and quantum computing.[11]

On August 5, 2026, Google announced that Koray Kavukcuoglu would become Senior Vice President of Google DeepMind, overseeing Gemini model development, frontier AI research, and Gemini app and developer teams while retaining his Chief AI Architect role. Hassabis would hand over day-to-day operations and become Google DeepMind Chair and Alphabet Chief Scientist, continuing to lead Isomorphic Labs and advise on models and research.[18]

In the same August 5 announcement, Google said Jeff Dean and Sanjay Ghemawat were launching an independent public benefit corporation focused on machine learning, science and engineering. Google would remain a founding investor and Cloud partner and collaborate on a research framework for ML systems and infrastructure.[18]

These organizational changes followed decades of work on search systems, neural-network infrastructure and language models. Google’s earlier history supplies the technical context: PageRank addressed ranking, TPUs addressed computation, and the Transformer and BERT changed how models represented language.[5][7][6][9][11]

Larry Page and Sergey Brin met at Stanford in 1995 and partnered the following year on Backrub, a search engine that used links to determine the importance of web pages. They renamed it Google, a play on “googol,” the number one followed by 100 zeros. In August 1998, Sun co-founder Andy Bechtolsheim wrote them a $100,000 check. The newly incorporated company moved from Stanford dorm rooms into a Menlo Park garage owned by Susan Wojcicki, later a Google employee and YouTube CEO.[2]

Brin and Page’s 1998 search-engine paper described a prototype combining text and hyperlinks. PageRank weighted links recursively: a link from an important page contributed more than one from a less important page, with each page’s contribution distributed across its outgoing links. Anchor text was associated with the destination page as well as the page containing the link, supplying descriptions that the destination itself might lack.[5][4]

Ranking was part of a working retrieval system. Distributed crawlers fetched pages, a storage server compressed them into a repository, and indexing and sorting produced the structures used to answer queries alongside PageRank. Brin and Page reported a full-text and hyperlink database of at least 24 million pages. This was the historical prototype, rather than a description of present-day Search.[5][3]

Google’s first Tensor Processing Unit grew out of a capacity concern. Its technical account recalls a 2013 projection that growing neural-network computation could require doubling its data centers. Google reports designing, building and deploying the processor in 15 months. Packaging it in an accelerator card that fit a hard-drive slot and connected over PCIe allowed installation in existing servers.[7]

The first TPU accelerated inference—the execution of trained models. It used 8-bit integer arithmetic and a systolic array that passed data between neighboring arithmetic units, reusing inputs for matrix multiplication while reducing memory access. This traded general-purpose flexibility for efficiency. The second generation, announced in May 2017, added training as well as inference, high-speed connections for TPU pods, and planned access through Google Cloud.[7][8]

Attention Is All You Need, by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser and Illia Polosukhin, introduced the Transformer in 2017. Gomez’s contribution was performed at Google Brain and Polosukhin’s at Google Research. The encoder-decoder architecture replaced recurrence and convolutions with attention: representations could draw on relevant positions elsewhere in a sequence without a recurrent chain of computations. Multi-head attention let the model attend to different positions and representation subspaces in parallel. The team demonstrated more parallelizable training on translation tasks and applied the architecture to parsing.[6][9]

Research breakthroughs did not automatically settle how to allocate computing resources. In a 2024 interview, former Google language-model lead David Luan recalled a Brain Credit Marketplace in which researchers received credits for compute. He said assembling a giant job required roughly 19 or 20 colleagues to leave their allocations unused, making large efforts difficult to organize from below. Luan attributed OpenAI’s advantage in scaling GPT-3 partly to concentrated resources and focus; this was his historical assessment.[16]

With BERT’s November 2018 open-source release, Jacob Devlin and Ming-Wei Chang explained how to pretrain contextual representations, then fine-tune them for a task. BERT predicted masked words using context on both sides and also learned whether one sentence followed another. Its representation of “bank” could therefore reflect both preceding words and a following word such as “account.” BERT built on earlier contextual-representation work; released TensorFlow code and pretrained models let developers adapt those representations using labeled examples without repeating base-model pretraining.[9]

In October 2019, Google applied BERT to Search ranking and featured snippets, using Cloud TPUs to serve some models. Google reported initial ranking coverage of one in ten US English searches. Its rollout account illustrated both an improvement and a remaining failure:[10]

• Travel direction. For “2019 brazil traveler to usa need a visa,” Google reported that BERT recognized the importance of “to,” returning more relevant results for a Brazilian traveling to the United States instead of results about US citizens traveling to Brazil.[10]
• Incorrect geography interpretation. For “what state is south of Nebraska,” Google reported that BERT’s best guess was a community called “South Nebraska.” Contextual representations still misinterpreted this straightforward question.[10]

Grounding with Google Search connects Gemini to current web content. With the search tool enabled, the model can generate queries, process results and return text with citation annotations linking answer segments to source URLs. Those annotations expose supporting sources; they do not by themselves establish that an answer is correct.[14]

Function calling lets Gemini request an application-defined operation with structured arguments; application code executes custom functions and returns their results. Google’s March 17, 2026 tooling update allowed built-in and custom tools in one request and preserved context across tool calls. Its headphone example paired Search with a check_inventory declaration, then printed requested names, arguments and call IDs. Completing that lookup required an inventory implementation and a returned function result. Supported tool combinations depend on the model and pairing.[13][15][14]

The Interactions API was generally available by June 2026 as a unified interface for models and agents; the older generateContent API remained supported. Interactions record execution steps and optionally retain conversation history for continuation through previous_interaction_id. History continuation does not preserve configuration: applications must resupply tools, system instructions and generation settings.[17]ⁿ¹

Google’s business extends beyond models. Google Services includes Search, YouTube, Android, Chrome, Gmail, Maps and Pixel devices, earning revenue primarily from advertising, alongside subscriptions, app sales and devices. Google Cloud earns consumption-based fees and subscriptions through infrastructure, Google Cloud Platform and Workspace. Vertex AI provides access to Google, third-party and open models, while Gemini supports enterprise agents and Workspace applications.[1]ⁿ²

At Alphabet’s June 2026 investor presentation, Pichai reported more than 2.5 billion monthly users for AI Overviews, more than one billion for AI Mode, and more than 900 million for the Gemini app. In the August 5, 2026 leadership announcement, he reported that the Gemini app had reached more than 950 million monthly users. Pichai also reported in June that more than 8.5 million developers were building with Google’s models each month. These figures describe different products and populations and cannot be added into a count of unique users.[12][18]

In the June presentation, Alphabet management reported 2025 consolidated revenue above $400 billion, including more than $60 billion from YouTube advertising and subscriptions. CFO Anat Ashkenazi reported Google Cloud revenue of $20 billion in the first quarter of 2026 and a quarter-end backlog of $462 billion. Backlog represented future business rather than revenue already recognized.[12]ⁿ²

Alphabet expected $180–190 billion in 2026 capital expenditures, overwhelmingly for technical infrastructure, and planned to fund investment through operating cash flow, debt and equity. Its June 3 presentation described a proposed $80 billion equity raise, including a $10 billion Berkshire Hathaway investment and a $30 billion underwritten offering. After oversubscription, approximately $35 billion of the offering had been priced and allocated, bringing the expected total to approximately $85 billion. These were announced and expected financing amounts.[12]ⁿ²