In brief
Google DeepMind is Alphabet’s AI division, born in April 2023 from the merger of DeepMind (founded in London in 2010) and Google Brain (founded in 2011). It is led by Demis Hassabis, the original co-founder of DeepMind. The entity concentrates fundamental research and product deployment under one roof, with a structural advantage rare in the industry: a proprietary compute infrastructure — the TPUs — that reduces its Nvidia dependency and lowers its training and inference costs.
Google DeepMind is to date one of three organizations in the world capable of developing frontier models (along with OpenAI and Anthropic). Its singularity lies in the breadth of its portfolio: from computational biology (AlphaFold, 2024 Nobel Prize in Chemistry) to consumer language models (Gemini), through open-source models for developers (Gemma).
In short: Google DeepMind behaves like a fully integrated old-style automaker — it builds the car (Gemini), but also the engine (the TPUs), the assembly line (Alphabet datacenters) and the distribution network (Search, Android, Workspace, billions of touchpoints). Where OpenAI buys its GPUs from Nvidia and its cloud from Microsoft, DeepMind controls the entire chain end to end. This vertical integration is slow to set up, but once acquired, it dramatically reduces marginal costs and accelerates iteration.
Profile
| Founded | April 2023 (merger); DeepMind founded in 2010 |
| Headquarters | London, United Kingdom |
| Parent company | Alphabet Inc. |
| CEO | Demis Hassabis |
| Status | Alphabet division (listed on NASDAQ: GOOGL) |
History
DeepMind was founded in 2010 in London by Demis Hassabis, Shane Legg, and Mustafa Suleyman. Hassabis brought three distinct lives to the project: chess prodigy (master at 13), video game developer (Theme Park, 1994), and cognitive neuroscience researcher (PhD UCL, postdoc MIT and Harvard). The founding project: build systems capable of learning from raw experience, without hard-coded rules.
In 2014, Google acquired DeepMind for approximately $500 million. DeepMind continued to operate as a relatively autonomous unit from the UK, pursuing a long-term research agenda — AlphaGo (2016), AlphaFold (2020) — while Google Brain, founded in 2011 by Jeff Dean and Andrew Ng, handled research applied to Google’s products at scale.
The coexistence became costly redundancy: Brain developed PaLM and LaMDA, DeepMind developed Gopher and Sparrow — two competing LLM families within the same group. The arrival of ChatGPT in late 2022 forced the decision. In April 2023, Alphabet merged the two entities. Demis Hassabis took the helm of Google DeepMind; Jeff Dean was promoted to Chief Scientist of Alphabet.
In short: the 2023 merger solves a problem worth several hundred million dollars per year. Two laboratories were building two competing LLM families in parallel — duplicated training, infrastructure, teams — without coordination. ChatGPT triggered consolidation: Alphabet decided to align energies on a single line (Gemini) rather than continuing to fund two divergent trajectories.
Models and products
AlphaGo and AlphaFold: scientific credibility
AlphaGo defeated world Go champion Lee Sedol 4-1 in 2016 — the first victory of an AI system over a human expert in a game long considered inaccessible to machines. The event established DeepMind’s international scientific credibility.
AlphaFold marks a more lasting turning point. AlphaFold 2 (2020) predicts the three-dimensional structure of proteins with precision below one angstrom — comparable to experimental methods that take years and cost millions. In 2024, Demis Hassabis and John Jumper received the Nobel Prize in Chemistry for this work. AlphaFold 3 was made available to the scientific community (non-commercial use) in November 2024, then publicly in February 2025.
The Gemini family
Gemini is Google DeepMind’s line of multimodal language models, launched in late 2023. It spans multiple capability levels (Pro, Flash, Flash-Lite, Nano) according to use case — from embedded mobile to heavy datacenter inference.
Gemini 2.5 Pro and Flash, presented at Google I/O 2025, marked a recognized resurgence in form. In late 2025, Gemini 3 Deep Think reached gold medal level at the International Mathematical Olympiad (IMO 2025) and excelled in chemistry and physics. Gemini 3.1 Pro has been available since February 19, 2026.
The rest of 2026 shows the two branches of the line-up moving at different speeds, which is worth keeping in mind to avoid picking the wrong model. The Pro branch has stayed at 3.1 — Gemini 3.1 Pro and Gemini 3.1 Deep Think are still the heavy reasoning models as of 6 September 2026, and Google announces a “3.5 Pro coming soon”. The Flash branch, by contrast, has shipped version after version: Gemini 3.5, then 3.6, then 3.7 in August — only three weeks after 3.6 — and 3.8 Flash, presented by Google as its most intelligent workhorse model for coding and agents.
At Google I/O 2026, in May, Google introduces Gemini Omni, a model able to generate from any input — images, audio, video and text combined — starting with video. The line-up is completed by specialised models: Gemini 3.5 Flash-Lite for volume, Gemini 3.5 Transcribe for speech recognition, Gemini 3.8 Flash Cyber, Gemini Image, Gemini Audio, Gemini Embedding 2, and Gemini Nano embedded in the Pixel 11.
In short: at Google, the version number does not tell you the power. In September 2026, the most “advanced” model by number is a Flash 3.8, while heavy reasoning still sits on an older Pro 3.1. Comparing two Gemini models by their number leads you to the wrong choice.
Gemma: the open-source line
Gemma is Google DeepMind’s open model line, designed for developers who want to deploy locally. Gemma 3 (March 2025) runs on a single GPU, offers a 128,000-token context window, supports over 140 languages, and handles text, images, and short video. Specialized variants round out the catalog: TxGemma for therapeutic development, DolphinGemma for animal communication research.
Positioning
Infrastructure as differentiator
Google DeepMind’s most tangible advantage is not a model — it is a chip. Google has been developing its TPUs (Tensor Processing Units) since 2013, deployed in production from 2015. The seventh generation, Ironwood (April 2025), is optimized for inference: 192 GB of HBM3e memory, 4,614 TFLOPs per chip, and Superpod configurations of 9,216 chips reaching 42.5 Exaflops. At comparable scale, TPUs are approximately twice as cost-effective as Nvidia GPUs, with 67% lower energy consumption.
In 2026, Google announces the eighth generation and splits it into two chips: TPU 8t for training (a 9,600-chip superpod, 2 petabytes of shared memory, 121 ExaFLOPS) and TPU 8i for inference (288 GB of HBM, 384 MB of on-chip SRAM). Google claims up to twice Ironwood’s performance-per-watt, and 80% more performance-per-dollar on inference. General availability announced for later in the year. [vendor figures, not validated by an independent benchmark]
All Gemini models have been trained on TPUs. Google controls both the silicon and the software stack — giving it cost control and iteration speed that competitors cannot replicate in the short term.
In short: keep three significant numbers on the TPU advantage. Cost: roughly 2× cheaper per TFLOP than an equivalent Nvidia GPU at datacenter scale. Energy: 67% lower consumption per Google’s figures. Capacity: an Ironwood Superpod aggregates 9,216 chips to reach 42.5 Exaflops — the equivalent of several national supercomputers concentrated on a single training load. The differentiator is not a published benchmark; it is the ability to mobilize this internal capacity without filing an Nvidia ticket or waiting for a competitor to free up GPUs.
Integration into the Google ecosystem
Google DeepMind’s deployment strategy differs structurally from that of OpenAI (which builds its audience through ChatGPT) or Anthropic (enterprise-focused). Google distributes its models through billions of existing entry points: Search, Android, Google Workspace, Google Translate. This eliminates audience acquisition costs.
The limitation is symmetrical: Google DeepMind is dependent on Alphabet’s product-level decisions. Integration decisions depend on agendas that extend beyond AI logic alone — regulation, advertising, cloud strategy.
Strengths and limitations
Strengths: proprietary infrastructure (TPUs), global reach through the Google ecosystem, established scientific credibility (Nobel, AlphaFold), native multimodal capabilities, open-source portfolio (Gemma).
Limitations: Gemini revenues not disclosed separately (financial opacity), dependence on Alphabet’s decisions, late start in the conversational assistant market against ChatGPT, historical tension between fundamental research culture (DeepMind) and rapid deployment imperatives (Brain).
Key takeaways
- Google DeepMind is the most vertically integrated actor in the AI race: from silicon (TPUs) to frontier models (Gemini) through fundamental research (AlphaFold).
- Its strength is not an isolated technological breakthrough, but a combination of assets — infrastructure, distribution, scientific depth — that is difficult to replicate.
- TPUs give Google a structural cost and iteration-speed advantage on training and inference.
- Its main challenge is to translate this structural advantage into perceived leadership in the language model market, against OpenAI’s cultural head start.