Sundar Pichai on AI Agents, Open Models and Google’s Compute Constraints

At Google I/O 2026, Matthew Berman speaks with Google and Alphabet CEO Sundar Pichai about agents, security, open models, competition and computing capacity. Their discussion describes product plans and industry conditions in May 2026; predictions and company claims should be read in that context.

Berman asks whether agents will become the main entry point to the internet. Pichai says developers have moved from accepting code completions toward orchestrating agents, and he expects agents to become a core part of web use. He emphasizes user control, agency and transparency as conditions for their adoption.

Pichai does not expect agents to replace every interaction. People may welcome automation of chores such as filling out renewal forms while still enjoying shopping for a gift, finding entertainment or pursuing meaningful interests. In his view, agents can free people to use the web for those purposes.

Berman asks how users can trust an agent to curate their information. Pichai compares the issue to spam filtering and Waymo, saying trust must be earned through delivered value, operational experience and evidence. His description of Waymo’s safety is his assessment, not an independently established comparison in this interview.

Pichai says Google is deliberately introducing Gemini Spark through first-party services such as Gmail and Calendar before expanding to third-party connections, MCP, browser use and computer use. He frames the staged approach as a way to gather feedback and give users control. Google’s rollout plans were specific to the time of the event.

Berman worries that agents may put another buffer between people and the raw internet. Pichai responds that users still value direct relationships with creators and trusted sources, and often enjoy discovery and shopping themselves. He says agents should connect people to the wider ecosystem rather than eliminate that connection.

Pichai acknowledges that agents add an intermediary layer, which can help users reach what they want but also abstracts content. He expects a new balance as creation tools expand. Berman says curation may help amid low-quality content while leaving room for exploration; Pichai agrees that exploration is a human need.

Berman turns to AI-assisted cyberattacks and asks whether Google is seeing an increase. Pichai describes Google’s security history and says its teams use agentic workflows to find and patch vulnerabilities. He reports that improving models have helped them detect more over the previous two years, without giving a quantified attack trend.

Pichai cites Anthropic’s Mythos as an important moment in specialized cyber capabilities. He then describes Google’s CodeMender as a tool used internally and being prepared for external users to identify vulnerabilities, generate patches, test them and deploy fixes. These are his descriptions of the tool’s workflow, not evidence that every fix works without oversight.

He also points to the completed acquisition of Wiz and its real-time monitoring capabilities, envisioning a combination with CodeMender. Pichai welcomes cross-industry collaboration on cybersecurity and on marking AI-generated content, including SynthID, as a necessary response to the technology’s risks.

Berman contrasts Anthropic’s restricted release of Mythos with OpenAI’s cyber model and asks whether Google would hold back a dangerous model. Pichai says the answer depends on its capabilities and how much it changes the existing frontier. Small advances may be suitable for release, while a dramatic jump warrants a more cautious process.

Pichai compares that decision to established security disclosure: Project Zero finds vulnerabilities and normally gives vendors time to patch before disclosure. He advocates working closely with government when capabilities change sharply, while ensuring enough defenders have access to patch their systems. His mention of one-to-two-percent versus twenty-percent changes is an illustration of judgment, not a published release threshold.

Berman asks why Google does not release a large frontier model as open source. Pichai points to Google’s history with Chromium, Android and Kubernetes and its Gemma model family. He says frontier models require substantial capital and research investment and may yield new techniques, so Google takes what he calls a balanced approach between investment and an open ecosystem.

Berman asks whether the economics of open models are viable for US startups, given the cost of training and competitors serving the resulting model. Pichai says Google also invests in Flash and Flash-Lite to offer a range of options. He expects demand for an open ecosystem even if its proximity to the frontier changes as technology advances.

Berman notes strong models from China and asks why US companies would not use low-cost options such as DeepSeek. Pichai says enterprises choose solutions based on reliability, predictability, safety and security as well as price. He expects both open and closed models and providers that build services around open models.

Pichai says that with suitable licenses he worries less about an open model’s country of origin than whether the US is keeping up at the frontier. He argues that community scrutiny can help establish trust, while acknowledging that AI models are not identical to conventional open-source software. Openness by itself does not guarantee a model’s safety or suitability.

Berman raises a different dependence: Chinese developers might optimize models for their own chips, drawing users into another country’s ecosystem. Pichai says applications should be built so that the underlying model can change as the field evolves. He considers it too early to decide whether the proposed dependency will become a real problem.

Asked why Google emphasizes fast workhorse models, Pichai points to the need to serve billions of users across Search and Gemini and provide affordable tools to developers. He says efficiency and accessibility matter alongside frontier capability, particularly when an agentic workflow repeatedly calls a model.

Pichai reports hearing from CIOs concerned about AI spending and argues that Gemini 3.5 Flash can help manage cost and latency. He says Google uses a blend of Pro and Flash internally and encourages companies to consider a similar mix. These are Google’s own product assessments, not an independent cost comparison for every workload.

Berman asks whether the ultimate race is to reach self-improving AI first. Pichai says the industry should avoid a dangerous race dynamic and deploy technology responsibly. He observes that perceived leadership shifts quickly as labs release models on different schedules and compete with different strengths.

Pichai considers recursive self-improvement a possible scenario, not an established result. If a capability jump comparable to a major cyber breakthrough occurs, he says, it requires more responsible handling and a broader societal conversation rather than a decision by one company alone.

Berman asks about the compute behind Google’s models, products and cloud customers. Pichai says Google made major long-term investments, though in hindsight it could have invested more. Rising costs, including memory, mean a given budget may buy less computing capacity than previously planned.

Google plans separately for cloud customers and internal needs, Pichai says, while making tradeoffs across both. He gives the example of customers asking for the infrastructure behind a fast Gemini 3.5 Flash demonstration. Planning and allocation remain difficult even with long-term investment.

Asked directly whether demand exceeds Google’s available compute, Pichai says it does. That makes efficiency more valuable: a larger or more capable model may reach fewer people if it is too expensive to serve. He describes model size, capability and broad access as ongoing tradeoffs rather than a single optimization.

For the final question, Berman asks about the largest bottleneck. Pichai says constraints shift as each one is addressed: permitting and constructing data centers, supplying power, and obtaining core components all matter. Memory may be the limiting factor at one moment and something else at another, so he sees bottlenecks across the stack.