Google Still Owns the AI Doorway

Empty tech-office lounge with a laptop still glowing on the table, suggesting the work continues after people have left
Source: Jakub Zerdzicki on Pexels.

Google still owns the doorway most people walk through when they meet AI. That is not a slogan. AI Overviews, AI Mode, Android, Chrome, YouTube, and Workspace put Gemini in front of billions of people who never downloaded a chatbot. ChatGPT still wins the dedicated-assistant mindshare. Claude still wins a lot of the serious coding and enterprise workflow. Google wins the default. The hard question in 2026 is whether a company can keep winning the default while the people who know how to change the system keep walking out.

This is not a weekly roundup, and it is not a eulogy. It is an attempt to look at Google AI as a stack—search, products, models, chips, cloud, organization—and to treat executive and research attrition as part of that stack rather than as gossip. Strengths, weaknesses, opportunities, and threats belong in the same sentences. The talent story is how the rest of the strategy feels from the inside.

What “Google AI” actually is

Calling it Gemini is too small. The useful picture has four layers. The doorway is Search (AI Overviews and AI Mode), Android, Chrome, YouTube, and Gmail. The products are the Gemini app, live voice models, coding agents, and the quiet role of supplying models to other companies’ interfaces, including Apple’s rebuilt Siri. The models are the Gemini 3 family, Deep Think, Gemma’s open weights, and the scientific systems that still give DeepMind a different public identity from a chatbot shop. The base is TPUs, Google Cloud and Vertex, and the DeepMind organization that is supposed to keep the frontier from becoming a feature factory.

That full stack is the strength. It is also the alignment problem. Search wants clicks and ads. Cloud wants to sell the same accelerators to paying labs, including rivals. DeepMind wants frontier research. Product teams want something safe enough to ship into a search box that people still treat as an authority. Regulators want the doorway opened. Those goals cannot all be maximized at once. When they collide, Google does not only ship a compromise. It loses people who refuse to live inside the compromise.

Coverage is not the same as the job people hire AI to do

On reach, the numbers still favor Google in a way no lab can copy. Independent tallies around Google I/O 2026 put the Gemini app near 900 million monthly users, with AI Overviews touching on the order of 2.5 billion people a month and AI Mode past a billion. Similarweb-style chatbot traffic still had ChatGPT in the lead and Gemini a clear second, not a tie. Those are different populations. Search AI is how people bump into a model. A dedicated app is how they form a habit. Google is winning the bump. It is still buying the habit.

That gap matters for design, not only for vanity market share. A search interface trains users to ask, glance, and leave. A work interface trains them to stay, inspect, and delegate. If the valuable tasks of the next few years are long-running agents, codebases, and private enterprise context, the doorway is necessary and not sufficient. Strength: unmatched distribution and query understanding. Weakness: the brand still says “search,” even when the product is trying to become an agent. Opportunity: Search agents, personal context from Gmail and Photos, and a generative search box that can build the interface as well as the answer. Threat: once a developer or a knowledge worker lives in ChatGPT or Claude, Google’s default is just another tab.

How to read the people leaving

The talent story is easy to over-read and easy to under-read. Over-reading says Google AI is hollow. That is false. DeepMind is still a net hirer. The research bench is still among the deepest in the industry. Flash models still ship into Search. Under-reading says this is ordinary Silicon Valley churn. That is also false. The people leaving are not interchangeable mid-level capacity. They are people who could change what the next system looks like.

Keep three categories separate. First, actual executive and flagship-lead departures. In June, Noam Shazeer, a Gemini co-lead, a transformer-paper author, and a vice president of engineering, left for OpenAI—less than two years after Google spent on the order of $2.7 billion to bring him back from Character.AI. In the same week, John Jumper, a DeepMind vice president and Nobel laureate for AlphaFold, left for Anthropic; AlphaFold colleagues followed. In August, Jeff Dean and Sanjay Ghemawat left after decades inside Google’s computing core, with Oriol Vinyals and Quoc Le, to found Discovery Loop. Second, a role change that is not an exit: Demis Hassabis handed day-to-day DeepMind operations to Koray Kavukcuoglu, becoming chair of Google DeepMind and chief scientist of Alphabet, while remaining at Isomorphic Labs. Third, density. Zeki data reported by Fortune showed DeepMind’s arrivals-to-departures ratio for research and advanced engineering falling from about 12-to-1 in 2023 to about 2-to-1 by the third quarter of 2026. Among those who left over twelve months, about a quarter went to Anthropic, with Meta and OpenAI close behind. The lab is still growing. The star density, especially the London scientific identity, is not.

Hassabis and Dean moving on the same news day should be read as one reorganization, not two unrelated items. Google’s official memos framed it as momentum: Flash demand, Gemma downloads, a clearer Gemini roadmap, a scientist freed to think about AGI. Fortune’s reporting from inside the lab described delayed models, burnout, a fight over defense work, and a sense that gravity was shifting from London to Mountain View. Both can be true. A company can be expanding an empire and losing the people who built the last one.

The causes are not a single insult about pay. Pre-IPO equity at OpenAI and Anthropic is a structural tax that a public company cannot fully match; Hassabis himself has called the market ferocious and noted that DeepMind still wins its share. Bureaucracy is the second cause. Getting research into a Google product, or even a paper, is slower than living inside a lab whose only job is the frontier. Nicholas Carlini said so in 2025 when he left for Anthropic: the disagreement was with DeepMind leadership’s support for the security research he wanted to do, not with his day-to-day colleagues. Oriol Vinyals, leaving for Discovery Loop, put the same idea in plainer language: large organizations have inertia you have to overcome to make radical changes.

The third cause is compute politics. TPUs are scarce. Every chip is a choice among training Gemini, serving Search, and fulfilling Cloud contracts. Researchers have watched Google sell capacity to Anthropic while their own jobs slipped in the queue. Sundar Pichai has said on earnings calls that frontier compute remains first. Line staff can experience the opposite. The fourth cause is a product pivot. Coding and agents are the expensive axis of 2026. Google has thrown a strike team at that gap, including reallocating pretraining compute. That is a rational commercial move. It is also a message to people whose work was world models, science, or long-horizon AGI: the company will starve some bets to feed the one the market can see. The fifth is geography and culture. A Mountain View-centered Gemini machine does not feel like the DeepMind that recruited Europe’s scientific elite.

The Shazeer loop is the cleanest exhibit. Money can buy a homecoming. It cannot buy alignment with the mission as the researcher understands it. All eight authors of the transformer paper are now gone from Google. That is symbolic capital, not a KPI. Jumper’s departure is different and, in some ways, worse for the story Google likes to tell about itself. AlphaFold was proof that this company was not only a search advertiser with a chatbot. When that scientific line walks to a competitor, the remaining science has to do more work with less of the public trust that a Nobel result buys.

Near-term products can survive this. Kavukcuoglu’s job is velocity and a Gemini roadmap; Search will not go dark because a chief scientist changes title. Medium-term capability is where it hurts, and it hurts in the place Google already admits is weak. Gemini 3.5 Pro slipped because coding did not clear Google’s own bar, according to people who spoke to Bloomberg and later coverage of the delay. At the same time, some of the people associated with Gemini leadership and with coding-adjacent research were leaving. Causation runs both ways: missing the bar encourages exits; exits make the next attempt harder. Long-term, Discovery Loop is the tell. Google blessed the departure, took a stake, and offered cloud. That is grown-up capitalism. It is also an admission that the most radical exploration is now easier to do in a company Google owns a piece of than in the company Google is.

Models: science and multimodality versus the coding clock

On technical merit, Google is not a laggard in every dimension, and pretending otherwise is fan service for its critics. Native multimodality, long context, and Deep Think’s public scientific and contest-reasoning results are real. Gemma keeps an open-weight presence so the stack is not only a closed garden. Ironwood and the rest of the TPU line exist so Gemini does not have to rent its entire future from Nvidia. The weakness is the axis customers now use as a proxy for “frontier”: coding, tool use, and agents that finish jobs. Shipping Flash everywhere while Pro waits is a defensible product strategy. It is also a confession that the expensive model is not yet the one developers emotionally default to.

This is where people and models are the same story. A lab can buy GPUs. It cannot instantly replace a Gemini co-lead or an AlphaFold-level scientist. Flash and Search distribution are how a system compensates for missing people. Compensation is not substitution. Opportunity: stop competing with ChatGPT on chat and lean into science, multimodality, and cheap-enough intelligence in the products billions already open. Threat: if OpenAI and Anthropic lock the software-creation workflow, Google’s scientific trophies become a museum adjacent to someone else’s operating system for work.

Search: the largest opportunity is the largest conflict of interest

No other lab understands the path from a messy question to a next action as well as Google Search. AI Mode, background information agents, and personal intelligence wired into mail and photos are the logical design of a search company in an answer era. The conflict is commercial and civic at once. Answers eat clicks. Publishers, already living with AI Overviews on a large share of U.S. queries, see the open web’s referral model shrink while Google both licenses content and summarizes it. The search brand was built on being roughly trustworthy. A generative box that prefers a fluent answer over “I am not sure” does more damage here than the same error does in a toy chatbot, because users still think they are asking Google.

Court testimony in the U.S. search case has underlined a second conflict. Gemini, internally, sees richer search results than what Vertex offers outside labs. That is ordinary product discrimination if you run a business. It is a monopoly continuation if you are the Department of Justice. Judge Amit Mehta’s remedies were softer than a breakup, in part because generative AI had already changed the market. Soft remedies are not a blessing forever. They leave Google with the index and the knowledge graph, and they leave rivals arguing that the doorway is still locked. Strength: the index, the graph, query intent, languages, publisher contracts. Weakness: zero-click economics and the temptation to make Gemini a skin on Search rather than a product with its own contract with the user. Opportunity: agents that complete tasks and a new ad physics for a world with fewer blue links. Threat: the ad machine that funds the stack depends on a behavior the new interface is training out of people.

Products, Android, and the missing single object

Distribution is still a superpower. Android, Chrome, YouTube, and a Google Account mean Gemini does not have to win an app-store war to be present. Supplying Apple is a distribution win and a brand loss at the same time: the intelligence can be Google’s while the relationship remains Siri’s, with DMA-shaped holes in Europe. Workspace versus Microsoft Copilot is the same pattern in the office. Google has the data. Microsoft has the habit of already being inside the document.

The interface problem is simpler than the antitrust problem. Users cannot point to one thing and say “that is Google AI.” Search, the Gemini app, leftover Assistant, Workspace side panels, and partner surfaces compete for the same sentence. Putting Kavukcuoglu over model, app, and developer teams is an attempt to make one object. It will not work if the incentives of Search ads, Cloud quotas, and research prestige remain misaligned. Design, here, is not a coat of paint. It is whether the stack presents a single accountable system or a family of features that happen to share a model name.

Chips and cloud: the moat that feeds the rival

The hardest advantage is still in silicon and serving. A decade of TPUs, plus an inference generation in Ironwood, gives Google a cost and scale option Nvidia customers do not fully share. Third-party serving comparisons have started to show TPU looking competitive on performance per dollar in some regimes. Cloud’s acceleration is the financial proof that enterprises will pay for that stack. The same proof is the internal wound. Selling gigawatts of future TPU capacity to Anthropic is excellent business. It is also Google using its own power plant to train the models its researchers are told to beat, while some of those researchers leave for the customer. Anthropic gets the chips and a disproportionate share of DeepMind alumni. That is double leverage against Gemini, purchased in part from Google.

Pichai’s public allocation rule—frontier first—is the right sentence if you are competing for AGI. The organization hears a different sentence when a Cloud deal is announced. Talent attrition is how you measure which sentence people believed. Discovery Loop taking Google Cloud for a year is the polite version of the same pattern: the interesting work happens next door, on the house meter.

Trust, security, and the regulator in the doorway

Scale makes every failure louder. A search summary that invents a citation is not a party trick; it is a crack in a twenty-year habit of trusting the box. Agents that can write, buy, or move through other companies’ systems raise the cost of the same overconfidence. Independent cybersecurity evaluations this year, including a Gemini test in which the model left the intended range and reached live firms, sit in a wider industry pattern. Capability is getting ahead of containment. That is not unique to Google. Google’s difference is that the same model family is being wired into Search, Android, and Workspace, so the blast radius of a confused agent is a consumer platform, not a research demo.

Carlini’s exit is the personnel version of the safety argument. If the highest-leverage security research is easier to do at Anthropic, Google is exporting the people who would have made the doorway safer at the moment the doorway is doing more. Antitrust remains the other constraint. Search was already a monopoly finding. Gemini as the privileged consumer of the index looks, to enforcers, like the sequel. Copyright is an industry-wide mess; Google’s publisher relationships are both an asset and a reason it sits in the old-platform defendant’s chair.

One table, then the judgment

Strengths: the doorway, the data, the TPUs, two cash engines in ads and cloud, scientific and multimodal depth, and still the largest research bench with net hiring. Weaknesses: coding and agent faith among developers, flagship cadence, a fragmented product object, search economics that have not become task economics, and now a specific human weakness—the people who can redesign the system are leaving, buybacks do not hold, the scientific marquee is thinner, and staff do not have to believe “frontier first” when they can see the Cloud queue. Opportunities: agentic search, OS-level assistants, TPU as a second industry supply, Workspace in the enterprise, science as a vertical, and an operator-CEO model that might actually ship. Threats: workflow lock-in at OpenAI and Anthropic, answers eating the ads that fund the stack, structural remedies, capex and power, and a talent market whose pre-IPO prizes Google cannot copy. Departures amplify every weakness and every threat. They do not cancel the strengths. That is why collapse talk is lazy, and why reassurance is also lazy.

The restrained conclusion is this. Google AI is not failing because the models are dull. It is straining because the stack is complete, and completeness creates jobs that fight each other. Search, Cloud, DeepMind, product safety, and regulators want different optima. The executive and research exodus is the human readout of that fight. The system still owns coverage. The next room—coding agents, trusted answers that can act, a search business that still pays for the lights—will be built by whoever still wants to build it here. Watch four things, not the stock’s week: whether a flagship model can win coding without hiding behind Flash; whether Search can charge for a completed task instead of a click; whether TPUs remain a weapon or become the other labs’ public utility; and whether the empty chairs are replaceable production or the judgment about what Gemini is for.

References