Google and Sundar Pichai
Ashish Bhatia
Google is losing people at the top of its research organization at the same moment that the open-weight layer of the AI market has been taken over by Chinese labs and the U.S. government has quietly made open models the easiest thing in America to ship. Those three facts point at an opportunity that Google has not taken: releasing the weights of a frontier-class Gemini model, not just the small ones.
To be clear about the terms, an open-weight model is one whose trained parameters are published for anyone to download, run on their own hardware, and modify. It is different from open-source software in the strict sense, because the training data and training code usually stay private. What you get is the finished model, free to inspect and deploy. This is what people mean when they talk about Moonshot AI’s Kimi K3, DeepSeek v4, or Alibaba’s Qwen 3.8-Max being “open.”
What Actually Happened In Six Weeks
On June 19, Noam Shazeer, a co-lead on Gemini and a co-author of the seminal 2017 “Attention Is All You Need” paper that introduced the Transformer architecture underlying every major model today, left for OpenAI. Forty-eight hours later, John Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold, left for Anthropic. Alphabet shares dropped more than 5% the following Monday. Two AlphaFold veterans, Jonas Adler and Alexander Pritzel, followed Jumper to Anthropic.
Then came August 5. Hassabis moved to Chairman of Google DeepMind and Chief Scientist of Alphabet, with CTO Koray Kavukcuoglu taking over operations as senior vice president reporting directly to Sundar Pichai. Hassabis said he wanted “the time and space to focus on the big picture.” The same day, Dean announced Discovery Loop, a public benefit corporation aimed at automating scientific experimentation, co-founded with Sanjay Ghemawat, Oriol Vinyals and Quoc Le, and funded by Radical Ventures and Khosla Ventures with Alphabet itself participating.
Fortune’s reporting describes Gemini 3.5 Pro missing three consecutive release dates, engineers working 60-hour weeks, and persistent burnout from what one described as playing catch-up ever since ChatGPT launched. Ghemawat told reporters the group wanted to “build something differently than how things are built at Google right now,” and pointed at infrastructure optimized for consumer products rather than research.
Putting that together, that describes an organization under sustained pressure to ship flawless proprietary releases on a schedule its competitors set.
The Two Pressures Squeezing The American AI Stack
The first pressure is concentration. Menlo Ventures’ most recent enterprise survey put Anthropic at 40% of enterprise LLM API spend and OpenAI at 27%, with Google at 21%. Together the two vendors hold roughly two-thirds of the market for the layer that increasingly sits underneath enterprise workloads like customer service, code generation, document processing, and internal search at large companies. Many enterprise leaders are raising the same three concerns about that: pricing control, concentration of power, and the inability to run anything in an air-gapped or regulated environment.
The second pressure is that the open alternative is now mostly Chinese. Data compiled from OpenRouter, a routing service that sits in front of many models and therefore sees real production traffic rather than downloads, shows Chinese-origin models at roughly 61% of tokens consumed by May 2026. DeepSeek alone routes about 17.6%. Alibaba’s Qwen passed a billion cumulative downloads on Hugging Face and now accounts for something like 40% of all new model derivatives created there. Meta’s open-weight share of routed volume fell below 1%. Google’s fell from about 37% to 13% year over year.
What Google Already Does In The Open, And Where It Stops
Google is not absent from open weights. Gemma 4, released April 2, 2026 under an Apache 2.0 license, comes in four configurations topping out at a 31-billion-parameter dense model, and Google’s own framing emphasizes “digital sovereignty” and control over your own data and infrastructure. On the Q2 earnings call, Pichai said the Gemma family had passed 900 million cumulative downloads, with more than 300 million of those coming from Gemma 4 since April.
Those are real numbers from a real open-weight program with a genuinely permissive license, which is more than Meta offered for most of the Llama era. Still, Gemma tops out well below Gemini’s frontier tier, so a company that wants Google’s best reasoning has to rent it through an API. When Meta released Muse Glimmer on August 10, a 30-billion-parameter model distilled, meaning trained to mimic a larger model at a fraction of the size, from its proprietary Muse Spark and also Apache 2.0 licensed, it landed in the same weight class and made the same trade. Meta’s Alexandr Wang has promised open weights for Muse Spark 1.2 “soon,” which would be a real move up the ladder if it happens on the frontier model rather than a distilled version of it.
The one American entrant that has actually broken past that ceiling is not a model lab at all. Nvidia shipped Nemotron 3 Ultra on June 4, a 550-billion-parameter mixture-of-experts model with about 55 billion active parameters per token, released under the Linux Foundation’s OpenMDW-1.1 license, which permits commercial use. On Artificial Analysis’s Intelligence Index, a benchmark aggregator that scores models across reasoning, coding and knowledge tasks, Ultra scored 48, the highest of any US-origin open-weight model and well ahead of Gemma 4’s 31B model at 39. It still trails China’s Kimi K3, which scored 57.1.
So there is an American frontier-adjacent open-weight model, and Nvidia built it rather than a research lab, which says something about incentives. Google, OpenAI and Anthropic sell inference and subscriptions, so an open frontier model competes directly with the product they charge for. Nemotron proves the American open-weight ceiling can move.
The Case For Opening A Frontier Gemini
Google holds four assets here that its U.S. competitors do not.
1. Opening a frontier model is the fastest way to cut into OpenAI’s and Anthropic’s lead. Google is not competing from a position of strength right now. One assessment of the 2026 field gave Google a momentum score of 3 out of 10, against OpenAI’s 10 and Anthropic’s 8, noting that coding has become the dominant use case driving AI adoption and that “Gemini is largely absent from that conversation” while Codex and Claude Code define it. Shipping a better proprietary model on the same closed terms OpenAI and Anthropic already win on will not close a gap like that.
Opening a frontier Gemini would compete on a different axis entirely: price, not the leaderboard. Meta’s Llama releases already demonstrated the mechanism, putting a hard ceiling on what OpenAI and Anthropic could charge for comparable capability. The same dynamic just repeated itself. When Moonshot AI’s Kimi K3, DeepSeek v4 and Z.ai’s GLM-5.2 arrived within weeks of each other in July, OpenAI cut GPT-5.6 Luna’s price by 80%, from $1 to $0.20 per million input tokens. A Google-made frontier model released the same way would apply that pressure directly to the two companies actually beating Google today, and it would hand Google a bargaining chip it does not currently have.
The same release would do double duty on the China front. Chinese-origin models already account for roughly 61% of tokens routed through OpenRouter, the same measure cited earlier, while Google’s own share of that traffic has fallen to about 13%. Staying closed does nothing to move that split, since a proprietary Gemini is not even competing in the category where the share is being lost. A frontier Gemini released openly would be the first mainstream, lab-built American alternative to Kimi, DeepSeek and Qwen, the first real chance to pull that 61% number down instead of watching it climb.
2. The research lineage is Google’s, and so is the developer love. The Transformer came out of Google Brain. So did much of the talent now running research at OpenAI and Anthropic. Beyond AI specifically, Google has been one of the most consistent forces behind open source software for two decades, and developers love it. About 10% of Alphabet’s full-time workforce contributes to open source projects in a typical year, and more than 20,000 outside developer accounts committed code to Alphabet-led projects in 2025 alone, according to Google’s own open source retrospective. That goodwill already shows up in AI specifically: the Gemmaverse, the community of developers fine-tuning and remixing Gemma, has produced more than 60,000 model variants since the family launched in 2024.
3. Google has run this play before, at scale, profitably. Android, Chromium, Kubernetes and TensorFlow were all released openly while Google monetized the layer around them: distribution, cloud, ads, and services. Kubernetes is the closest analogy. Google gave away the orchestration standard and then sold the managed infrastructure to run it on. Alphabet’s Q2 numbers suggest that surface is stronger than ever, with Cloud revenue up 82%, a $514 billion backlog, and nearly 90% of the Fortune 100 using Gemini Enterprise.
4. Washington has removed the regulatory friction. The White House framework finalized in early August evaluates the cyber capabilities of frontier models before release, and excludes open-weight models from that review entirely. National Cyber Director Sean Cairncross has said the administration wants to build U.S. open source and make it “the preferential adoption by planet Earth.” Whatever one thinks of that policy on the merits, and there are serious people who think the exclusion is a mistake, it means an American lab releasing frontier weights today faces a lighter compliance path than one shipping the same capability behind an API.
Alphabet shares fell 5.4% on August 5, the day Google announced that Demis Hassabis was handing over day-to-day control of Google DeepMind and that Jeff Dean was leaving after 27 years. It was the second time in six weeks that a top researcher’s departure moved the stock.
A Speculative Second Benefit: Lower Stakes Per Release
Every proprietary frontier release is a referendum on the company. Gemini 3.5 Pro missing three deadlines became a story about whether Google can compete at all. That dynamic raises the internal bar for shipping, which raises review burden, which raises the number of people who can say no, which is a reasonable description of what the departing researchers were complaining about.
Open weights change the shape of that risk. A published model gets evaluated, fine-tuned, and corrected by thousands of outside developers, and the release becomes the start of a process rather than a verdict on it. Gemma has already given Google a working example of this mode, at a scale where the stakes are lower.
I would not claim this fixes burnout or bureaucracy. Those come from structure, headcount, and unclear ownership, and no licensing decision touches them. But a team that ships into a community that improves its work is in a different psychological position than a team shipping into a leaderboard.
The Three Strongest Objections
“Opening frontier weights helps China more than it hurts them.” Chinese labs have advanced in part by distilling and building on published Western work, so publishing a frontier Gemini hands them a free jump forward and accelerates Chinese competitors. Where I think it falls short is the counterfactual. Chinese open models already lead production usage globally without Google’s participation, and Kimi, Qwen and DeepSeek are now the base models that others build on. Google staying closed does nothing to slow that down, and it leaves the enterprise choosing between an American API and a Chinese open model, which is the choice that produces the dependency the objection is worried about in the first place.
“There is no way to monetize an open core at this scale.” The strongest version notes that Gemini API revenue is real, that Google has no Red Hat-style support business to fall back on, and that giving away the model gives away the thing customers pay for. The response is that Google’s monetization surface is unusually broad. Most enterprises will not self-host a frontier model even when they legally can, because serving it well is hard, so the managed API business persists for the majority of workloads. Meanwhile the weights being free makes Google infrastructure the natural place to run them, and Search, YouTube, Workspace and other Google properties monetize the underlying capability directly regardless of who else has a copy.
“The safety risk at frontier scale is different in kind.” The strongest opposition is from Anthropic, laid out in its position on open-weight models: safeguards can be removed from published weights, a released model cannot be recalled or updated, and the developer loses any ability to detect and disrupt misuse or control who has access. Anthropic stops short of calling for a ban and argues instead for mandatory pre-release safety testing of any sufficiently capable model, open or closed. Both of those points hold. Where I land is that the marginal risk depends on what already exists. Frontier-capable open weights are already circulating from labs operating under different safety regimes, so the question is whether the open tier includes an entrant that publishes evaluations, documents known failure modes, and funds external red-teaming. That case is stronger when a safety-invested lab is in the pool than when it sits out.
What To Watch
Watch whether Meta ships open weights for Muse Spark 1.2 itself rather than another distilled model in the 30-billion class, whether Gemma 5 closes the gap to Nemotron 3 Ultra’s Intelligence Index score, and whether Google’s share of routed open-weight tokens recovers from 13% or keeps sliding. The market asked Google a question on August 5 about whether it can still lead. The open-weight lane is the one place where Google’s own history counts as an advantage, and right now a chipmaker is occupying that ground instead of a model lab. Will Google take the baton?