Google, DeepMind and Gemini: how the AI business fits together

Google’s AI business connects a research organisation, consumer products, advertising and cloud computing. Alphabet is the parent group, with Google reported through its Services and Cloud segments. Google DeepMind develops intelligence used across those businesses.[1][2]

The names describe different parts of the organisation

Google DeepMind brings together the former DeepMind and Google Brain teams. Their backgrounds were complementary: DeepMind pursued general learning systems, including through games, while Brain developed machine-learning methods and software with applications across Google. The combined organisation is led by Demis Hassabis.[2]

Its responsibilities also reach into products. In October 2024, Google moved the Gemini app team into DeepMind. The stated reason was to shorten the feedback loop between model development, post-training and release in the consumer app. This helps explain why a research organisation can also be responsible for an experience people use every day.[3]

How customers reach Google’s AI

An individual can use the Gemini app and buy a Google AI subscription for additional access and bundled services. The current plans combine AI with benefits such as storage and tools inside Google applications; eligibility varies by feature, country and account. A subscription is therefore a package of product access, rather than a purchase of the model itself.[4]

A software developer can use the Gemini API, an application programming interface that lets their software send a request and receive a model’s response. Google AI Studio provides a place to test prompts, create prototypes and manage API keys. The developer can then build a distinct product around that access: a customer need never open the Gemini app to use the underlying technology.[5]

For organisations building and operating AI systems, Google Cloud offers Gemini Enterprise Agent Platform. Introduced in April 2026 as the evolution of Vertex AI, it brings model development and agent-building services together. The naming matters when reading older partnership announcements: references to Vertex AI describe the predecessor platform, while current documentation uses the newer name.[7]

The platform’s Model Garden includes Google, third-party and open models. Its surrounding services handle work such as connecting enterprise data, managing agent identities and inspecting executions. Those are products with value even when a customer selects another company’s model. Google can compete to supply the intelligence and also earn business from operating the environment around a rival’s intelligence.[6]

Advertising still shapes the economics

Google Services earns most of its revenue through advertising across Search, YouTube and partner properties. AI helps advertisers find audiences and create campaigns. The paying customer can therefore be an advertiser while the person using an AI-enhanced service pays nothing directly.[1]

Advertising represented approximately 68.1% of Alphabet’s total revenue that quarter, calculated from those figures. Cloud has become a substantial business, but its reported total also includes services beyond AI. The accounts therefore cannot be read as a simple comparison between “Gemini sales” and another model developer’s revenue.[8]

Alphabet reports some shared AI research and development outside its operating segments. Cloud’s segment profit therefore excludes certain group-level AI research costs, preventing a simple standalone profit calculation for Google’s entire AI effort.[1]

The same strategy reaches down to chips and out to open models

Google designs tensor processing units, or TPUs, specialised accelerators used for training and running AI. Its Cloud offering makes this computing available to customers as well as supporting Google products. The current TPU catalogue usefully separates shipping status: Ironwood is generally available, while TPU 8t and 8i are still labelled “coming soon”. Designing a processor, announcing it and providing usable capacity are distinct stages.[9]

Google also releases Gemma, a separate family of downloadable models developed using research and technology related to Gemini. The weights—the numerical parameters defining a trained model—come with developer tools. This gives builders a route to work directly with a model, alongside Google’s hosted applications and APIs.[10]

These connections give Google several reasons to improve AI. Better products can retain users and attract advertisers; useful tools can earn subscriptions; developers can buy model access and computing. The company’s position is strongest when those activities reinforce one another. Understanding that structure is more revealing than treating each Gemini feature as an isolated business.

Report an error or suggest a correction