Thomas Kurian on Google Cloud’s TPUs, AI Agents and Cybersecurity

Matthew Berman interviews Google Cloud CEO Thomas Kurian about compute capacity, the TPU business, Anthropic, agents and security. Their discussion was published in April 2026; company-specific performance and customer examples are Kurian’s account.

Berman asks how Google can train and serve Gemini while also providing compute to competing labs. Kurian says Google earns revenue from its own models and from other companies using its chips. He attributes capacity to years of planning across land, power, data-center construction and silicon.

Kurian says Google diversified energy sources, secured real estate and moved more data-center work toward prefabricated manufacturing. Deploying preassembled and tested units, he argues, reduces time to bring machines online. He also notes Google’s long partnership with NVIDIA alongside its own TPU development.

He says demand for TPUs is coming from AI labs and other fields, including financial markets and high-performance computing. He cites Citadel and the US Department of Energy as examples of broader use, describing this as an expansion beyond the original AI workload.

Asked to compare monetization of different uses, Kurian declines to give a detailed breakdown. He says owning chip intellectual property allows Google to earn margins through multiple channels. He also describes placing TPUs in selected customers’ data centers near trading venues.

Kurian argues that serving varied customers exposes Google to more requirements and aggregates demand when negotiating with suppliers. Berman asks why a company pursuing advanced AI would not reserve all available compute for its own models.

Kurian answers that the business needs continuing cash flow to fund infrastructure and training. Google balances capacity for customers against internal needs and capital requirements, he says; his argument is that external revenue is one way to support further investment.

Asked whether Gemini also faces constraints, Kurian predicts compute demand will exceed supply for the next decade. He argues that owning a chip helps Google’s economics under scarcity. That is his forecast, not a verified guarantee about future supply or costs.

Kurian does not disclose a detailed training-versus-inference allocation. He says cloud accounts for roughly half of Alphabet’s capital investment and notes that Gemini and Google’s models contribute significantly to cloud growth; he offers that only as a broad approximation.

Berman asks about local opposition to data centers. Kurian identifies energy bills and jobs as key community concerns and describes behind-the-meter power, alternative energy sources and a geographically distributed buildout as parts of Google’s approach.

Kurian claims Google pays close attention to power usage effectiveness: the extra facility energy required beyond the computing load. He also claims strong efficiency and community investment. The conversation does not provide independent comparisons or establish that a particular project has no grid impact.

On public confidence in AI, Kurian says benefits must be demonstrated in practical use. He describes a German insurance customer using Gemini-based agents and claims some inquiries went from minutes to seconds without job losses at that organization. This is a company example, not proof about employment or service quality across the economy.

Kurian also describes work with the American Society of Clinical Oncology to help clinicians navigate complicated care guidelines. He says accuracy and avoiding hallucinations are essential, and cites a planned Citigroup wealth-advice application as another possible use. These are descriptions of applications and goals, not evidence that medical or financial outputs are always correct.

Berman asks directly whether Google Cloud is hiring as AI raises productivity. Kurian says the organization is adding product, sales and forward-deployed engineering staff, while demand for its services remains strong.

Kurian says models can both detect code vulnerabilities and help repair them. He describes planned security agents for continuous testing, prioritizing findings and assisting fixes, connecting this work to Google’s Wiz acquisition. Berman contrasts Google’s hiring with Block’s announced layoffs; Kurian says each company makes its own choices according to demand.

Berman asks whether NVIDIA or Google has the lower total cost per token. Kurian says many customers consider Google’s platform best on total cost of ownership and cites strong demand. He does not present a controlled comparison, so the interview does not settle the relative cost of the two platforms.

Kurian says chip performance is only part of the system. He describes TPU 8t scaling to 9,600 chips and 8i to 1,152 in a pod, with high-bandwidth networking and predictable latency. Google later publicly described those configurations and up to two petabytes of shared memory for 8t.

Above the hardware, Kurian points to JAX, XLA, Pathways and PyTorch work as part of the programming and compiler stack. He uses “goodput” to mean effective useful throughput and says Google also optimized tokens or dollars per watt in anticipation of energy constraints.

Asked how much long-term plans change, Kurian says experience compounds across successive generations of hardware and software, while customer feedback still changes priorities. He says the decision to build a specialized inference chip reflected an expectation that model providers ultimately need inference revenue to fund training.

Berman asks about the eighth-generation split. Kurian says Ironwood was a mixed training and inference chip; 8t is mainly for training, although some customers may use it for inference, while 8i is primarily for inference but can also train smaller models.

Kurian describes three broad phases in enterprise AI use: question answering and research; generation of images, audio and video; and agents that use tools and interact with business systems. These are his characterization of product evolution, not mutually exclusive categories.

An enterprise agent may need to work with customer-management or supply-chain software and interpret information returned by those systems. Kurian defines an agent as a module to which a task can be delegated, with skills and tools that may include using a computer.

He cites Xfinity, Walmart, Bosch and Merck as examples of organizations using Google technology in customer care, scheduling, manufacturing and research. These are Kurian’s descriptions of customer use, with scope and outcomes varying by implementation.

Berman asks how those workloads inform the chip split. Kurian says input and output patterns change between question answering, media generation and agents. Longer tasks also affect how much context and intermediate information a system must retain.

Kurian gives key-value cache and memory management as examples. Agents may run for hours, and shuffling state can be costly. Computer use also requires conventional compute, which is why Google works with Intel and develops Arm-based processors as well as AI accelerators.

Inference must often be distributed closer to users to manage latency, while training can be concentrated in large facilities. Kurian says 8i can operate without water cooling, allowing use in a wider range of data centers. He presents these as examples of decisions made across the system.

On co-design, Kurian discusses storage for both large-scale training and low-latency inference. He cites managed Lustre throughput of 10 terabytes per second and a rapid storage system at 15 terabits per second, as company-specified figures for different systems and units.

Kurian also discusses Virgo networking and says Google coordinates networking, storage and compute for agent workloads. His next bottleneck is the cost of virtual machines and local storage for consumer agents that may need to start and stop tools repeatedly.

In his travel-planning example, an agent could query several services and calculate a budget. Kurian argues that keeping a virtual machine running indefinitely would be costly for consumers, making efficient activation, storage and resource sharing important for access at scale.

Berman returns to Anthropic: it uses Google infrastructure while Claude competes with Gemini. Kurian calls this normal for a platform company, saying that different parts of Google can supply and compete with the same organization. He offers the Apple relationship as another example of cooperation alongside ecosystem competition.

When capacity is limited, Kurian says Sundar Pichai and the executive team make allocation decisions across Gemini, Anthropic and other labs. He does not give a formula or allocation percentages, but says owning chips while demand is strong is a favorable position.

Berman asks about a rumored ten-trillion-parameter Mythos model and whether Google is operating at that size. Kurian does not confirm the rumor or disclose an equivalent Gemini parameter count. He says Google expects further Gemini releases and is confident in its ability to serve large models.

Kurian points to disaggregated serving as a way to scale large models and says Google designs models with serving in mind. His assertion that its TPU serving stack is the most efficient is a company claim, not a published independent comparison in this interview.

Asked whether pretraining scaling is slowing, Kurian says he does not see such a slowdown from the chip-design, system-design or capacity perspective he oversees. That answer does not establish that every model continues to improve at the same rate as compute increases.

On data, Kurian moves from unstructured text, audio and video to enterprise records. A citation to a document can be straightforward, he says, but explaining a computed inventory answer from changing tables is more complicated: an agent must select the right fields and make its reasoning traceable.

Kurian says enterprise applications with many fields and controls offer challenging examples for model training and trajectory optimization. Google’s enterprise work, in his view, supplies feedback for teaching Gemini to use complex systems.

Berman asks about criticism that Google is behind on agentic coding. Kurian says many engineers use an internal coding harness called JetSki and send feedback to Google DeepMind. His account describes internal adoption but does not independently measure Google’s position against competitors.

Berman says he reviews few lines in his own small projects and asks how Google can move faster while maintaining quality. Kurian says lines of code are a poor productivity metric: a senior engineer may solve the same problem more compactly. He emphasizes functionality delivered instead.

Kurian says Google retains peer review and uses Gemini to inspect code and scan for security vulnerabilities before senior engineers review it. He also describes exposing cloud tools to a model for incident troubleshooting. These uses aim to reduce routine work, while review remains part of the process he describes.

Berman raises the risk that people may lose understanding as AI generates, reviews and debugs code. Kurian agrees it is an industry risk: a prompt cannot describe every behavior or exception in a complex system. He says humans are still needed to use tools and address security problems.

Kurian asks whether a model can reliably identify mistakes in code it generated itself. He says Google is exploring approaches, including supervisor models, while continuing peer review. Thousands in his organization use the internal tools daily, he says, across coding, testing and review.

Berman asks about Anthropic’s decision to restrict Mythos Preview because of cyber capabilities, and whether Google has a clear public-release threshold for Gemini. Kurian says Google is still working through what that line should be. Anthropic had announced restricted preview access; the interview does not establish a confirmed Mythos parameter count.

Kurian asks what vulnerabilities can also be found using open models, which adversaries may obtain. His argument is that models used to discover flaws require corresponding defensive tools to repair them, prioritize findings and continuously test systems.

He describes the operational challenge of deploying fixes: identifying all locations running vulnerable code and replacing them, not merely writing a patch. Berman asks whether AI vulnerability discovery is an argument against open-source software; Kurian says Google both uses and contributes to open source and intends to help secure it.

Asked what keeps him up at night, Kurian lists long-term capacity planning, data centers, networks and TPUs, and solving important customer problems such as cybersecurity. He says Google’s Gemini Enterprise token volume rose from 10 billion to 16 billion per minute and enterprise users grew 40% sequentially over the period he describes; those are his company figures.

Kurian closes by stressing the need to anticipate customer needs as technology changes quickly. Berman thanks him for the conversation.