Source: YouTube

Source: Google I/O ā€˜26 Keynote

Just finished watching the I/O 2026 keynote. The key theme that I think Google was trying to land - and was successful at - was: Google is an artificial intelligence research and development company. They are firing on all cylinders through the entire stack: from research and bare-metal through the products and its monetization. Whatā€™s more: they have two businesses that are accelerating as a result of AI: cloud and search (and other consumer products). Finally, there was a message that was becoming clear: everyone wants to collaborate with Google because they all see that they are delivering capabilities up and down the stack. Googleā€™s become the ā€œfriendlyā€ company again: They now have deals with Apple, Amazon, OpenAI, Walmart, and every other large corporation. The only ones who are still mad at google are publishing houses dependent on the ever decreasing search traffic.

The research and bare-metal provides Google with cost efficient scale thatā€™s perfect for enterprises

It starts with model research and bare metal (chips).

For model-research, Google Deepmind researches both new models and improvements to said models. On the cutting edge, Google seems to both bet and innovate on moving further away from multimodal models (current generation of models) to ā€œworld models.ā€ Google Omni seems to be the class of models that is taking the frontier to a new place.

This seems to also be the direction in which Yann LeCun also wants to bet on and not just the current LLMs. Demis seems to suggest that the LLMs and world models can both interact and make each other better. We are in the super early innings on this and this space needs to be closely watched.

For bare-metal, Google highlighted TPU v8t and v8i - each designed for training and inference respectively. They highlighted that TPUs help them both achieve scale for large training and achieve incredible speed + scale for inference requirements, which are currently in the range of quadrillion tokens / month.

Then came the announcement of the latest generation of their models - Gemini 3.5 Flash. The key announcement with Gemini 3.5 Flash was: frontier level intelligence (nearly as good as Opus 4.7 and GPT 5.5) but a much faster model. This is an outcome of both model-research and bare-metal coming together.

They werenā€™t meekly hinting things. There was a clear message to enterprises: we have both the capacity and scale to cater to you while delivering more $ savings in token costs because we can.

the agentic era of Google

What followed next was a typical Google tell - it loves the technology - so much - to a fault. It started with how these models are great for agentic development versus start from what could it do. It announced a new version of Antigravity (2.0) with a fully rewritten antigravity-cli (for us TUI lovers), that takes advantage of 3.5Flashā€™s ability for longer time horizon work. They claimed that combined with the agent and harness, it can do multi-day work including complex tasks like building an operating system.[^1]

Itā€™s requesting that you be very bullish on these models and harnesses and highlighted the improvements in Googleā€™s surfaces where these models can improve their services. Starting with:

Gemini: Agents for everyone

The consumer surface - gemini app on the mobile operating systems and https://gemini.google.com/ . Apart from a flex that 3.5 flash is immediately available worldwide to 900M+ users, they are taking advantage of the agentic capabilities to announce an openclaw for every gemini user - Gemini Spark. Itā€™s an always-on agent that starts with working atop your google data to monitor and provide you assistive capabilities. It seems like a nifty implementation of Openclaw that will make sense to most users - especially as it is particularly great interacting with voice.

Again, the message here is: agents for everyone.

Search announced that AI mode is becoming Google Search. Personally, I think it might be okay as Iā€™ve been increasingly liking AI mode as well as AI overviews. When the stakes are low, itā€™s near perfect. When the stakes are high, I do some additional clarifications to ensure that I am not caught with bad information. I think itā€™s yet another salvo against the kind of businesses that depend on google traffic to monetize. =([^2]

Search also highlighted something that I thought was really cool: generative UI and long running agents that do the search for you with up to date information. I think this has some really useful utility. However, I donā€™t understand the arbitrary wall between Gemini Spark and Search agents. The closest I can think of is: some version of ability to use your data versus not / a potential experiment on running ā€œfreeā€ versus ā€œpaidā€ agents. However, I am stretching and the explanation might just be: This is google PAs not talking to each other / waiting to see which has uptake.

Google Shopping was an interesting set of announcements. I really liked the three pillars of innovation: Unified Commerce Protocol, Agent Payments Protocol and Universal Shopping Cart. The number of partners here were really interesting to see. It sends a clear message that they are back to their roots of we shall innovate with better technology and thereā€™s a growing pie. We will see how long that continues.

Workspace

Agents and multi-modal models can bring some nifty innovations to Google Workspace also. Docs Live was a really cool demo of being able to use your voice to both dictate and apparently soon - edit and manage document writing. I am going to be very interested to see the adoption of this.

p.s. that we still needs docs to be written - was an interesting shoehorning that I felt was uniquely Google. :)

Gemini - again.

This was an interesting quirk. Josh came back to highlight both some nifty UI improvements to gemini on mobile along with a more detailed demo of Spark. Itā€™s going to be interesting to see how users will adopt Spark. I can think of this as both the most exciting announcement personally but also the hardest to communicate with people about how to use this.

The technology behind this is so interesting: they have a virtualized container running on the cloud for every user and an agent capable of running the latest models. Purely as a reference implementation for the incredible technologies, this is just šŸ¤Œ.

However, while Josh comes up with some interesting use-cases, i donā€™t think many users will naturally move to planning their block-party the way Josh does. :)

However, this has now shifted from a technology -> a product and a marketing problem. The technology is damn impressive. However, we also get to the part where google usually flubs.

Daily brief

Seems to be CC in gemini. And I am here for it as Iā€™ve tried to setup scheduled actions in gemini to make that happen for me but failed.

Gemini on Mac demo

This was a really useful demo of the raw capabilities of the new model that combined with computer use has some really cool capabilities. I am still not convinced that I am going to be a mac app user. But, this specific demo is making me rethink given this + the fantastic gemini capability to understand my accent.

Google for Creators / Generative media

These didnā€™t hold my attention as I am likely not the audience. I took away notes that summarized as: multimodal + world models = more control over audio, images and video. This also means that we can likely have some richer output from the LLMs over time. Something to keep an eye on.

One note: Google Pics is a bad name. I really donā€™t get why you would even attempt to sully that branding associated with Google Photos and is mostly beloved.

Googleā€™s naming isnā€™t great but this specific one is going to backfire.

Android XR

Audio glasses are interesting, but again I am not the audience.

Closing aka Gemini for Science and some AGI peppering

AI can improve science, not just LLMs, but different other types of models, including my favorite weather prediction model. Personally, I think the AGI and responsibility narrative is required for someone at the scale of Google. Yet, it just felt like talk. I think these initiatives might actually do more good for the world, however, it still feels like an also-ran.

General notes on presentation

Like I said in the intro, this year Google had a clear message and it showed. There was much less meandering (modulo Android XR. Itā€™s easy to be a critic and way way harder to setup an event of this sort. So, kudos to the Google I/O team. Yā€™all delivered one of the best keynotes of the AI era.

I am curious what Ben Thompson thinks about all this and I will wait to see his (likely tending negative) take on Google I/O tomorrow. Itā€™s also clear that you will see our biases here. =)

My only tip for next year would be: continue that message of vertical integration and choose to lead with: here are some real consumer benefits - and look at this incredible tech stack that we developed to make it happen. And developers: yā€™all can take as much of this stack at any layer you want to help you achieve your goals too.

I still remember the panic that was set when OpenAI unveiled ChatGPT and everyone ruled out Google. However, it looks like the company is back in action, leading with technology and nerding out.

At least from the outside, it looks like a company thatā€™s got it together and leveraging its strengths to deliver immense value to users and shsreholders. Fantastic job, onward and upward.

P.S: I specifically havenā€™t talked about antigravity-cli versus gemini-cli here. I want to have the opportunity to evaluate the new cli before providing my thoughts on it. If I have to guess, my worry is that this is a path to lowering the number of credits. Especially as news is starting to come that 3.5 Flash is nearly as expensive as 3.1-pro even as Googleā€™s making that the default model everywhere.

P.P.S: I also have a suspicion that Google is increasingly thinking of itself as a free + ads / pay a subcription amount for its AI offerings. I will expand on that hypothesis in a follow up post.