Google Cloud chief reveals the long term: a decade of silicon and the energy battle behind the AI ​​boom

According to Kurian, Google’s preparations began well before the current hype cycle. “We’ve been working on TPUs since 2014… way before AI was hot,” Kurian said, referring to Google’s custom Tensor processing units. The decision to invest early was driven by the fundamental belief that chip architecture could be radically redesigned to accelerate machine learning.

The energetic premonition

Google’s foresight about the physical constraints of computing may have been more crucial than the silicon itself. While much of the industry was focused on speed, Google was calculating the electrical cost of that speed.

“We also knew that the most problematic thing that was going to happen would be energy, because energy and data centers were going to become a bottleneck alongside chips,” Kurian said.

This prediction influenced the design of their infrastructure. Kurian said Google designed its machines “to be extremely efficient and deliver the maximum number of flops per unit of energy.” This efficiency is now a critical competitive advantage as AI adoption increases, putting unprecedented pressure on global power grids.

Kurian said the energy challenge is more complex than simply seeking more power, noting that not all energy sources are compatible with the specific requirements of AI training. “If you’re running a cluster for training…the peak you get with that calculation consumes so much power that you can’t handle that from some forms of power generation,” he said.

To combat this, Google is pursuing a three-pronged strategy: diversifying energy sources, using AI to manage thermodynamic exchanges within data centers, and developing fundamental technologies to create new forms of energy. In a moment of recursive innovation, Kurian said that “the control systems that monitor thermodynamics in our data centers are all governed by our AI platform.”

The “zero sum” fallacy

Despite Google’s investment in its own silicon for a decade, Kurian pushed back against the narrative that the rise of custom chips threatens industry giants like Nvidia. He says the press often portrays the chip market as a “zero-sum game,” a view he considers incorrect.

“For those of us who have worked on AI infrastructure, there are many types of chips and systems optimized for many types of models,” Kurian said.

He called the relationship with Nvidia a partnership rather than a rivalry, noting that Google optimizes its Gemini models for Nvidia GPUs and recently collaborated to enable Gemini to run on Nvidia clusters while protecting Google’s intellectual property. “As the market grows,” he said, “we create opportunities for everyone.”

The full stack advantage

Kurian attributed Google Cloud’s status as the “fastest growing” major cloud provider to its ability to offer a complete “stack” of technologies. According to him, doing AI well requires owning each layer: “the energy, the chips or systems infrastructure, the models, the tools and the applications,” noting that Google is the only player to offer all of the above.

However, he said this vertical integration does not amount to a “closed” system. He argued that businesses demand choice, citing the fact that 95% of large companies use cloud technology from multiple vendors. Therefore, Google’s strategy allows customers to mix and match, using Google TPUs or Nvidia GPUs, as well as Google’s Gemini models alongside those from other vendors.

Despite the advanced infrastructure, Kurian offered a reality check for businesses rushing toward AI. He identified three main reasons why enterprise AI projects fail to launch: poor architectural design, “dirty” data, and a lack of testing for model security and compromise. Additionally, many organizations fail simply because “they haven’t thought through how to measure ROI.”

For this story, Fortune journalists used generative AI as a research tool. An editor verified the accuracy of the information before publishing it.