Cisco says it has more than tripled GPU utilization and expanded AI computing capacity sixfold within its existing data center footprint. The results suggest businesses may be able to scale artificial intelligence more efficiently by improving infrastructure management rather than relying solely on new construction.
WHAT’S HAPPENING
As companies race to expand artificial intelligence, Cisco is demonstrating that one of the biggest opportunities may be improving the computing infrastructure they already have.
In a November 2026 report, Cisco detailed how it transformed separate AI computing clusters into a shared, enterprise-wide system designed to support large-scale AI operations.
The overhaul was driven partly by the rollout of MyAgent, Cisco’s internal AI assistant serving approximately 90,000 employees, alongside growing demand for its customer-facing Cisco IQ platform.
Previously, individual teams operated dedicated AI computing resources. That arrangement worked for early experiments but created inefficiencies as AI became part of daily operations.
Some expensive graphics processing units (GPUs) sat underused while other projects needed additional capacity.
Cisco addressed the problem by building what it calls a multi-tenant AI factory, combining shared computing resources, high-speed networking, security, and real-time performance monitoring.
According to Cisco, the results include:
-
GPU utilization increased from 15% to approximately 50%.
-
Computing capacity expanded from about 250 to more than 1,568 GPUs.
-
The expansion occurred within existing data center facilities.
-
Enterprise service availability targets reached 99.99% or higher.
These figures are company-reported, not independently audited.
Importantly, Cisco added substantial computing hardware. Its achievement was fitting more capacity into existing facilities while using that capacity more effectively.
WHY IT MATTERS
The AI industry’s infrastructure expansion is creating enormous demand for electricity, data centers, cooling systems, and advanced computer chips.
But purchasing additional hardware does not automatically mean using it efficiently.
Companies that assign dedicated computing resources to individual departments can end up paying for capacity that sits idle much of the time.
Cisco’s approach demonstrates an alternative: pooling resources so workloads can share available computing power.
The company also redesigned networking to reduce bottlenecks, installed more efficient rack cooling systems, and incorporated security protections throughout the infrastructure.
Another important element is cost management.
AI services consume computing resources each time they process requests and generate responses. Monitoring those workloads helps businesses understand which activities justify their operating expenses.
Cisco determined that running approximately 80% of its steady-state AI inference workloads on its own infrastructure offered the most favorable economics for its requirements.
That calculation will differ for other organizations depending on workload volume, equipment expenses, cloud pricing, and operational capabilities.
The central lesson is that AI efficiency depends not only on how powerful the hardware is, but on how intelligently businesses operate it.
WHO BENEFITS
Large enterprises deploying AI may reduce wasted computing capacity and improve cost predictability by consolidating fragmented infrastructure.
IT departments can gain better visibility into hardware utilization, workload performance, security, and operational expenses.
Employees and customers may benefit from more reliable AI services as organizations improve system availability and manage demand.
Infrastructure and networking providers could see increased demand for technology supporting shared computing, workload management, and high-density cooling.
WHO LOSES
Businesses operating inefficient AI infrastructure may face higher costs than competitors that use their computing capacity more effectively.
Organizations relying on isolated computing clusters risk paying for underused equipment while purchasing additional capacity unnecessarily.
Companies without effective AI cost monitoring may struggle to determine whether their infrastructure spending is producing measurable business value.
Smaller businesses may also find it difficult to justify the specialized hardware and operational expertise required for large on-premises AI deployments.
WHAT HAPPENS NEXT
Cisco is preparing its infrastructure for the growing demands of autonomous AI agents.
Unlike traditional AI applications that respond to individual prompts, agentic systems can carry out sequences of tasks, generating more complex and unpredictable computing workloads.
Cisco is also pursuing improved automated workload routing, GPU resource management, and monitoring of AI operating expenses.
The next important test will be whether similar approaches can deliver comparable improvements across other large organizations.
For businesses considering major AI investments, Cisco’s experience offers a practical starting point: measure existing infrastructure utilization, identify stranded computing capacity, compare cloud and on-premises costs, and evaluate whether resources can be shared more efficiently.
Not every organization needs to build its own AI factory. But nearly every organization deploying AI at scale needs to understand how effectively its computing resources are being used.
The larger signal: The next phase of AI growth may depend as much on infrastructure efficiency as on constructing new data centers and purchasing additional chips. Cisco’s reported results show how shared computing, better monitoring, and integrated security can support substantial growth within existing facilities. As AI moves from experimentation into everyday business operations, the companies that manage computing resources most effectively may gain a significant competitive advantage.