Why Google Needs Its Founders Back in Charge of AI

As rivals surge and internal projects stall, direct leadership from Brin may be Google’s best hope to regain momentum in advanced model development.

This article originally appeared on Weijin Research on Huxiu on August 9, 2026. Original Chinese title: 「创始人模式才能挽救谷歌AI」. It has been translated and adapted for an English-speaking audience.

A person observing a model training curve displayed on a screen in a server room. The chart shows training loss (blue line) and validation loss (orange line) decreasing across epochs, illustrating typical machine learning model convergence.

Quick Take

There is a joke about how compute is allocated inside Google: if you work at Google, you get 100 cards; if you leave to start a company, you get 1,000.

Google Cloud gives its cards to Anthropic first, to make money, and then rents compute at a high price from Musk's SpaceX.

Jeff Dean decided to resign and start his own company, and Google immediately sent over funding and compute. But reportedly, Dean and his team left because they wanted to use Nvidia or AMD GPUs instead of TPUs. According to a Wall Street Journal report, they believe Google's infrastructure is well suited to large consumer applications, large advertising systems, and search, but the infrastructure their research wants is different; they hope for something more flexible than a TPU.

Actually, if Demis Hassabis eventually resigned and, based in London, led all of Europe in the AI for Science cause he is truly devoted to, Google would likewise follow him and send compute.

It's just that staying at Google doesn't work!

Google was dealt a good hand, but never formed a trump card in models. The launch of ChatGPT at the end of 2022 threatened search, its lifeline; in 2026, Anthropic's Claude Code again raised red flags inside Google, and after that, Chinese open-source models barged into the frontier, leaving Gemini behind at least in pace.

Since the beginning of this year, Google's AI leaders have begun to leave one after another, with the three most emblematic: David Silver, Hassabis's longtime startup partner and the standard-bearer of reinforcement learning, left to found a startup; Noam Shazeer, one of the principal inventors of the Transformer, jumped to OpenAI two years after returning to Google from his own startup; and most recently, John Jumper, who shared the 2024 Nobel Prize with Hassabis, defected to Anthropic.

Clearly, they think that going to a frontier Lab, or starting their own Neo Lab, is better than staying at Google.

They've all left. What about Gemini?

There is a claim circulating that Google is abandoning frontier models.

This episode has been interpreted as Google Cloud's Thomas Kurian winning the internal contest against DeepMind. There is another view that this may be a major adjustment in Google's AI strategy: giving up the competition in frontier models and instead devoting itself to becoming infrastructure and getting AI's killer applications right. Google putting its money into TPU + GPU + Google Cloud may yield better returns than pouring it into the frontier-model competition. Apple does not participate in the frontier-model competition at all, and it is doing just fine.

There are also suggestions that Google should simply open-source Gemini and, with the strong support of TPUs and Google Cloud, compete with Chinese open-source models.

Or turn Google into the world's largest venture capital group: release all the geniuses nurtured inside Google to start companies and change the world, with their AI projects all running on Google Cloud. Start with Dean and Hassabis.

For reorganizing its AI team, Google can look at two comparables: one is Meta, spending heavily to poach talent from all directions; the other is Musk, who spent $60 billion to acquire Cursor and made SpaceX an AI infrastructure company.

A more credible account is that Google co-founder Sergey Brin will personally take charge of the Gemini large model. More precisely, he will personally direct the training of Gemini 4.

On the second-quarter 2026 earnings call, when an analyst directly asked how compute is allocated among search, model training, and the cloud business, Google CEO Sundar Pichai confirmed that Google first determines how much compute frontier AGI development needs, takes it as the "baseline" that must be guaranteed, and only then allocates the remaining resources among businesses such as search, YouTube, and Cloud.

At the same time, Google has begun the pretraining of Gemini 4, calling it the company's "most ambitious pretraining project to date." Pichai acknowledged that the next phase of frontier competition requires "larger foundation models," and said that substantial compute and R&D resources are being invested in Gemini 4.

More important is Brin's thinking. He does not view models, infrastructure, and applications as three mutually exclusive paths. In his logic, the three form a loop:

Flowchart showing how Google's early Gemini model was applied across different business areas, including cloud and enterprise AI, search and Android applications, and supporting infrastructure like TPUs and data centers.

If Google does not have its own frontier models, this loop breaks: the next-generation interface for search could be constrained by competitors; Google Cloud would degenerate into an infrastructure provider that supplies compute for others; TPUs would struggle to achieve deep software-hardware co-design around Google's own frontier models; Android, Workspace, YouTube, and Gemini applications could only use an intelligence layer not controlled by Google; and the advantages Google has in users, data, products, and compute could not converge into a unified AI system.

Therefore, frontier models are not competitors to the cloud business; they are the precondition for the cloud business, the search business, and the application ecosystem to achieve differentiation. Pichai also made it clear on the earnings call that having its own models enables Google to optimize a complete integrated solution; although Google Cloud will also offer third-party models, Gemini remains the core of its full-stack advantage.

Brin's belief in frontier models has not changed. In his past public statements, he also held that the long-term trend in machine learning is not toward ever more unrelated specialized models, but toward capabilities gradually converging into more general models: specialized models can serve as tools for experimentation and breakthroughs, but successful experience will ultimately be absorbed into general models. Google DeepMind will maintain its formation:

Google Gemini model lineup and strategic functions. Two-column table showing model tiers (Gemini 4 base, Pro, Flash, Flash-Lite, Gemma, and specialized models) paired with their respective use cases and capabilities.

If Google AI enters "founder mode," Brin will turn Gemini 4 into a battle that must be won.

Gemini 3.5 Pro's delay, especially because its coding capability failed to meet internal expectations, exposed that Google's problems are not only insufficient compute but also training schemes, data, post-training, product coordination, and decision-making efficiency.

Brin will most likely give Gemini 4 extraordinary resource guarantees: a larger foundation model, more complete multimodal pretraining, stronger reasoning and tool-use capabilities, and a compute quota independent of the general business budget. The goal is not to briefly lead on existing leaderboards, but to anticipate the competitive frontier at the time Gemini 4 is released and aim at that position in advance.

Brin will make coding intelligence the breakthrough point and force the company to eat its own dog food.

Lately, Brin has been working at the headquarters campus in Mountain View, Silicon Valley, near the Gemini team. He personally examines loss curves, writes code, and constantly demands that the team increase speed and reduce bureaucracy. A founder who implements the company's AI strategy by writing code in this way will inevitably bring enormous authority and compress organizational layers.