Google Warns Agentic AI Will Increase Demand for Data Centre CPUs

Google has warned that the rapid development of agentic artificial intelligence (AI) could significantly increase demand for central processing units (CPUs) in data centres, potentially changing how AI infrastructure is designed and operated.

While much of the current AI boom has focused on graphics processing units (GPUs) and specialised AI accelerators, the growing adoption of autonomous AI agents is creating additional requirements for traditional computing hardware.

Mark Lohmeyer, vice president and general manager for AI and computing infrastructure at Google Cloud, told EE Times that agentic AI introduces a different type of workload compared with conventional chatbot services.

Traditional AI chatbots generally respond to individual prompts, relying heavily on GPUs or Google’s Tensor Processing Units (TPUs) to generate responses. Agentic AI systems, however, can break down complex instructions into smaller tasks, assign them to multiple agents and independently carry out activities such as accessing websites, interacting with APIs, writing code and executing software.

These activities place considerably greater demands on CPUs, which are responsible for managing processes, coordinating workloads and handling many of the tasks required to operate autonomous agents.

Agentic AI Creates New Infrastructure Challenges

One of the biggest challenges is the unpredictable nature of agentic AI workloads.

According to Lohmeyer, AI agents can remain idle for extended periods while waiting for information or external services to respond, before suddenly becoming highly active when additional processing is required.

Unlike traditional computing workloads, these systems must maintain their operating state and context over potentially days or weeks, even when they are not actively processing requests.

This creates new challenges for data centre operators, particularly when attempting to balance computing resources, maintain performance and control infrastructure costs.

Google believes that hundreds or thousands of AI agents could be activated rapidly, creating significantly more demand for computing resources than traditional applications designed around human users.

The company expects CPU requirements to become a much larger part of AI infrastructure as the industry moves away from primarily training large foundation models towards running autonomous AI systems at scale.

Google Looks Towards More Efficient Computing

Google believes that supporting agentic AI will require improvements across the entire data centre infrastructure, including processors, memory, storage and networking.

One approach being explored involves effectively putting inactive AI agents into hibernation.

Rather than keeping an agent running continuously and consuming valuable CPU resources, its state and context can be transferred to storage. The agent can then be quickly restored when further processing is required.

This approach could allow data centres to support significantly more agents without requiring dedicated computing resources for every individual agent.

Google Cloud is also using micro virtual machines (microVMs) to provide isolated environments for individual agents.

These lightweight virtual machines offer security boundaries between workloads while requiring fewer resources than conventional virtual machines. Google suggests that a server capable of running around 100 traditional virtual machines could potentially support approximately 1,000 microVMs.

Security will also become increasingly important as autonomous agents gain the ability to access external websites, applications, APIs and business information.

Google is investing in additional security controls, including agent gateways that can restrict and monitor outbound network connections.

Intel, AMD and Arm CPUs All Have a Role

Google does not believe there will be a single processor architecture suitable for every agentic AI workload.

Its Google Cloud infrastructure already uses x86 processors from Intel and AMD, alongside its own Arm-based Axion CPUs.

Different applications will require different processor characteristics. Data-intensive AI agents may benefit from processors offering high individual core performance and substantial local storage, while web-serving workloads may benefit more from processors offering large numbers of CPU cores.

Google believes customers will increasingly need the flexibility to select processors that best match their individual AI workloads.

The growing importance of CPUs in AI infrastructure is also attracting attention elsewhere in the semiconductor industry. Earlier in 2026, Morgan Stanley highlighted how the transition towards autonomous AI could increase demand for CPUs and memory alongside the GPUs that have dominated much of the industry’s recent investment.

A Changing AI Data Centre Landscape

The increasing adoption of agentic AI could represent an important shift in data centre architecture.

Although GPUs and specialised accelerators will remain essential for AI model training and inference, the supporting infrastructure required to operate autonomous agents is likely to place considerably more emphasis on general-purpose computing.

Processors, memory, storage and high-performance networking will all play increasingly important roles as organisations deploy larger numbers of autonomous AI systems.

For Google and other cloud providers, the challenge will be delivering the additional computing capacity required while keeping infrastructure costs under control.

If agentic AI adoption continues to accelerate, CPUs could become one of the most important components of the next generation of AI data centres.