Shane BuckleyA few months ago, I wrote about why I believe enterprise AI is evolving toward Hybrid AI, with organizations using different models and services for different workloads rather than relying on a single provider. The economics, performance, security, governance, and data requirements of each application will ultimately determine the right combination of models.
That shift is already beginning to take shape, and perhaps faster than many expected, including me.
Recent reporting from The New York Times highlighted how AT&T has rapidly increased its use of open models as it looks to reduce the cost of AI. Other organizations are making similar moves as open-weight models become increasingly capable, customizable, and economical. This is Hybrid AI beginning to emerge in practice.
Organizations certainly won’t abandon frontier models, nor will open-weight models become the answer for every workload. Instead, they will increasingly match models to workloads, combining frontier, open-weight, specialized, and privately deployed models based on what matters most for each application and business requirement.
But as that transition accelerates, another question becomes increasingly important: Is the infrastructure beneath AI ready for what comes next?
How Hybrid AI Changes Infrastructure Requirements
The conversation about AI infrastructure today is understandably dominated by compute. GPUs, power, cooling, and data center capacity have become strategic considerations as organizations race to support increasingly demanding AI workloads. But the AI infrastructure is only part of the equation.
As organizations deploy more AI models, AI-enabled applications, and autonomous agents, the number of interactions among them will multiply. Models will communicate with applications. Agents will access data, invoke services, and interact with other agents. Workloads will span public cloud, private infrastructure, SaaS, and the edge.
Every one of those interactions creates data in motion.
Hybrid AI won’t simply increase demand for compute. It will drive more network traffic, create new application dependencies, and generate growing volumes of telemetry across increasingly distributed hybrid cloud infrastructure. Organizations therefore need to prepare for two related challenges: the network carrying all that data must scale, and the visibility infrastructure used to turn it into intelligence must scale with it.
Why Telemetry Architecture Must Scale Efficiently
We’re already seeing the industry respond. For years, the prevailing approach was to collect enormous volumes of telemetry data and move it into centralized platforms for analysis. As data volumes grow, the economics of continually moving, duplicating, storing, and ingesting all of that data become increasingly difficult.
We’re seeing this shift across the Gigamon partner ecosystem. Innovations from Splunk, Elastic, Cribl, and others reflect a broader move toward distributed data lakes, federated access, and more flexible telemetry architectures. The approaches differ, but the direction is clear: store data where it makes sense, then access and analyze it when needed without continually moving or duplicating it.
As telemetry volumes grow, the answer can’t simply be to move and store proportionally more data. Hybrid AI will accelerate the need for more efficient architectures.
Scaling Network Visibility and Telemetry Efficiency Together
The same principle applies to the network. As AI-driven applications, agents, models, and services generate more data in motion, networks are evolving toward higher-capacity infrastructure. The visibility infrastructure supporting them must keep pace.
At Gigamon, we’re scaling the Gigamon Deep Observability Pipeline and the underlying visibility infrastructure for 800Gbps-class networks, helping organizations maintain visibility into data in motion as network speeds and traffic volumes increase.
But scale alone isn’t enough. Organizations also need to efficiently extract the context that matters and make it available to the tools and platforms that can put it to work.
Gigamon Application Metadata Intelligence (AMI), for example, transforms high-volume network traffic into application-aware metadata. Approximately 100 Gbps of packet traffic can yield 1 Gbps of application metadata, making valuable context available without requiring equivalent packet volumes downstream. We’re also improving telemetry efficiency by reducing metadata volume while preserving valuable application context.
The objective isn’t to create more data simply because more data exists. It’s to efficiently transform growing volumes of data in motion into higher-quality intelligence.
Preparing Hybrid AI Infrastructure for What Comes Next
The shift toward Hybrid AI is already underway. As organizations match different models to different workloads and applications and agents become increasingly autonomous, interactions among models, applications, services, and data will multiply.
Preparing for that future requires thinking beyond the next model or GPU investment. Organizations will need networks capable of handling rapidly expanding volumes of data in motion, visibility infrastructure that keeps pace with higher network speeds, and telemetry architectures that can economically store, access, and extract value from growing data volumes. Most importantly, they will need to efficiently deliver the context that matters to the people, tools, and AI systems making increasingly consequential decisions.
Hybrid AI is changing the infrastructure equation. Being ready means building for both scale and efficiency, from the network carrying the data to the visibility and telemetry architecture that ultimately turn it into intelligence.
CONTINUE THE DISCUSSION
People are talking about this in the Gigamon Community’s AI Exchange group.
Share your thoughts today
