On October 20, the Big Technology podcast released a long interview with Sundar Pichai, CEO of Google and Alphabet. Pichai said plainly that Google Cloud is deliberately giving up some revenue: compute goes to the teams building frontier models first, and to customers after that. He also shared where he thinks AI is heading over the next couple of years.
What I find more interesting in this interview is not what Pichai said, but where he dodged the question. How much of Google's code is written by AI - "the metric has become meaningless." What's next with Apple - "nothing specific." Where Gemini 4 Argon falls short at coding - "there are pockets." Below are the key points from the new interview.
1. Chat turns into an agent that works without you
The main product news: Google is preparing a personal agent for regular users. According to Pichai, it will run in the cloud around the clock, remember your context and be available in the Gemini app, in Search and in Gmail. The agent will run on Gemini 4 Argon.
Competitors have already shipped their agents. On September 29, OpenAI introduced Dots - always-on agents in ChatGPT, and on October 21 Anthropic answered with Claude Duet, which watches your email and calendar and messages you when it needs your decision. The interviewer asked directly where Google's answer to Dots is. Pichai said the goal for Gemini is to become a personal, proactive and powerful assistant, and Google showed the first steps toward it at I/O this year.
You won't have to decide whether something is "a chat question" or "a task for the agent": Gemini will figure it out. Pichai says that a year ago almost all requests to the app were regular chats, and now the share of agent tasks keeps growing.
The interface isn't going away, though. Pichai's example is voice input: he just talks into his phone, and Gemini fixes even unusual names of his friends. A screen will still be needed, he says, when you want to watch a video or look at what the AI made for you.
What it means: OpenAI, Anthropic and Google have shown the same thing within a month. AI stops waiting for you to write first: it lives in your email, calendar and work tools, runs in the background and comes to you when it needs your decision. "Ask and get an answer" turns into "assign and check the result." Dots and Duet already work this way, but Pichai gave no dates for Google's own agent.
2. Gemini inside Siri and billions in Anthropic
Pichai confirmed that Apple builds Gemini models into Siri, customizing them for its needs, and that the partnership is "working well." Asked whether it will expand, he said there was "nothing specific to comment."
On Anthropic, in which Google, as the interviewer points out, has invested and committed tens of billions of dollars, Pichai kept it simple: from the start the team looked like one that would be at the frontier. Another candid moment: when the interviewer noted that Google had also gone into its own "code red" over OpenAI, Pichai didn't push back. He said everyone has to respond with some level of urgency right now, and that OpenAI has pushed Google too.
What it means: Gemini reaches people not only through its own app, but also through other companies' products like Siri. And Google both competes with Anthropic and profits from its success.
3. AI writes Google's code not just more, but better
Pichai declined to say what share of Google's code is written by AI. He used to cite 25%, and now calls that number meaningless: AI is involved in almost everything one way or another.
Instead, he gave two other numbers. In the four weeks since Argon launched, the number of code changes (inside Google they are called CLs, changelists) grew by 30%. And rollbacks, meaning changes that had to be reverted because of errors, dropped by about a third.
What it means: if the numbers hold, the second one is the interesting part. Everyone is writing more code, but fewer rollbacks means the code got more reliable. There is no way to check this from the outside: the data is internal. And Pichai himself admits that Argon has "pockets where we need to do better" in coding.
4. Even Google doesn't have enough compute
This is the most candid part of the interview. According to Pichai, the rule at Google is this: Google DeepMind gets compute first, for research and frontier models, and that is "non-negotiable." Other businesses, including Cloud, plan around what's left.
A few more numbers from Pichai:
- in the first 72 hours, Argon processed almost 25 trillion tokens, about three times more than Gemini 3.0 in the same period;
- Google is serving Argon while "capacity constrained," meaning demand exceeds the servers available;
- today's compute is the result of decisions he made in 2024, and planning has to look 2-3 years ahead;
- for a customer asking for compute for 2027 right now, helping is "incredibly difficult."
What it means: the compute shortage doesn't only affect people who use AI. Recently I wanted to add another memory kit to my computer - and the prices bite: in less than a year it got about one and a half times more expensive. Graphics cards are a whole other story. Down the chain, this pushes up the price of every device that uses these components: laptops, smartphones, game consoles.
It's not about greedy stores. AI data centers are buying up memory, and manufacturers are shifting factories to server chips because that pays better. Gartner expects memory and SSD prices combined to rise about 130% by the end of 2026, and PC prices to go up about 17% on average. Pichai himself named memory among the things that could run short.
The same goes for AI itself: top models will come with limits, and responses will be slower at peak hours, and not only at Google. So if you were planning to upgrade your computer, waiting "until it gets cheaper" over the next year looks pointless.
5. A bubble? "This is not like fiber"
The interviewer asked directly: railroads and fiber were also overbuilt at some point. Pichai answered that the fiber comparison doesn't fit. Back then cables sat unused for years, while today the compute is in use every day.
He doesn't deny that overbuilding is possible, though. Building a data center takes more than 2,000 components: memory, electricity, land, people. A shortage of any of them or an economic shock could break the plans. But he is "absolutely" confident that Google will find a use for everything it builds.
What it means: the head of a company building AI data centers is unlikely to say "yes, it's a bubble." The more useful part here is that he listed himself where things could break.
6. Models help build models, and that's worrying
Google co-founder Sergey Brin has said that Argon was built to a large extent with the help of the model itself. Pichai clarified: true recursive self-improvement is still "pretty far," and for now these are agents helping researchers.
On safety he was more cautious than you might expect. Models are not perfectly aligned, it is "an active area of research, and there are clear gaps," and it is "reasonable to be worried." That's why Google monitors models in real time while they work and keeps them sandboxed.
As the next step he names cross-industry standards and independent evaluators, like auditors in accounting. And he adds that regulation will play a role too.
What it means: the head of one of the largest AI companies is asking for outside auditors himself. That's a good sign, but for now it's just words.
7. Data centers in orbit
The most sci-fi part: Google has already sent its TPU chips to space and is waiting for their first "hello world." The idea is to get solar power around the clock, with no clouds and no night, and to avoid building data centers in people's backyards.
Pichai compares it to Waymo's self-driving cars and quantum computing: Waymo took almost 20 years, quantum about 10, and space is only 2-3 years into the journey. He expects real results in 10-20 years.
What it means: this won't affect your life in the coming years. But it shows how seriously Google takes the energy shortage for AI.
What to take away from this interview
The main thing: nobody is planning to slow down. Google plans compute 2-3 years ahead, puts its own cloud in line behind researchers and is already sending chips to space. Pichai describes his approach as "confident, aggressive, but disciplined" and has no doubt that everything built will find a use. Investment in AI is only growing.
Pichai ended the interview on an optimistic note. He believes we are at the very beginning: most of the progress, in healthcare for example, is still ahead. AI could make cheaper things that have only been getting more expensive for decades, and help discover new materials. The scale of this, in his words, is hard to grasp today. But he immediately added that optimism doesn't cancel the need to take seriously all the ways this technology could go wrong.
The most practical idea of the interview came almost in passing. Pichai agreed that you can build a great product without the most powerful model: frontier models are "slower and more expensive." That's why Google has both Argon and the fast Flash models.
The same approach works for you. Hard tasks - large codebases, long research, a document hundreds of pages long - go to Gemini 4 Argon. A stream of simple tasks - emails, translations, short answers - goes through Gemini 3.8 Flash: on GPTunneL it costs four times less than Argon per token ($0.003 vs $0.012 per 1,000 input tokens).
Both models run on one balance, with no subscription, so you can switch right in the middle of a task. We covered what Argon can do and where it shines in our article on the Gemini 4 Argon launch.
Sources
- Sundar Pichai: Google's Frontier Imperative, Compute Scarcity Dilemmas, Future AI Products and Businesses, Big Technology, Alex Kantrowitz, 10/20/2026 - full interview transcript: agents, Apple and Anthropic, compute, code, safety, data centers in space.
- Memory prices will stay "crazy" through to at least 2027: Gartner, CRN, 2026 - Gartner forecast: memory and SSD +130% by the end of 2026, PCs +17%.