Agentic AI inference workloads reshape network requirements, with KPMG pointing out that fiber optics and edge computing become key.

Event Background

KPMG US Technology Lead Phil Wong pointed out in a recent interview with RCR Wireless News that as enterprises shift from the AI training phase to inference, especially with the rise of Agentic AI, network infrastructure will face a new round of demand restructuring. Wong emphasized that inference workloads will no longer be confined to traditional centralized data centers but will flow between enterprise cloud environments and AI-dedicated computing infrastructure, and even extend further to the network edge. This change will have a profound impact on fiber optic networks, data center site selection, and enterprise IT architecture.

Technical Analysis: Unique Requirements of Agentic AI for Network Architecture

Agentic AI refers to AI systems capable of autonomously executing complex tasks, interacting with the environment, and making decisions. Unlike traditional passive AI, Agentic AI needs to operate in real time by combining data, context, and memory, which imposes distinctly different requirements on the network:

• High Bandwidth: Agentic AI frequently needs to call upon massive amounts of enterprise business data (such as CRM and ERP system records) from the cloud during operation and perform inference computation on AI-dedicated GPU clusters. This continuous cross-environment data exchange demands extremely high bandwidth, especially when inference requests involve multimodal inputs (text, images, video).
• Low Latency: Many Agentic AI scenarios require millisecond-level responses, such as real-time customer service, automated trading, or industrial control. If network latency is too high, the timeliness of AI's autonomous decisions will be lost.
• Edge Distribution: When physical AI (such as robots and autonomous driving) matures, inference workloads will have to migrate to edge nodes close to end users or devices, further expanding network coverage.

Wong explained, "Agentic AI works best when combined with data, context, and memory. Traffic between the traditional cloud (where enterprise data and record systems reside) and AI-dedicated computation will increase significantly. Additionally, with the development of physical AI, inference traffic will further spread to the network edge, getting closer to the end user."

Enterprise Impact Analysis

Cost Impact (CAPEX / OPEX)

• For enterprises, the network requirements of Agentic AI will directly translate into new capital expenditures.- CAPEX: Enterprises need to invest in higher-bandwidth internet connections (such as dedicated fiber or SD-WAN upgrades) and may deploy local inference servers or GPU nodes at edge sites. For enterprises adopting a hybrid cloud strategy, cross-cloud data transfer costs will become a continuously increasing operational expense.
• OPEX: Electricity costs are currently the biggest operational challenge. Wong points out that power availability is the primary factor limiting AI infrastructure expansion, even more severe than supply chain delays or labor shortages. Enterprises need to evaluate the power reliability of data center locations and consider renewable energy procurement to stabilize OPEX.

Deployment and Operations Impact

• Data Center Location: Traditional data center markets (e.g., Northern Virginia, Silicon Valley) are facing land and power shortages. Wong states that developers are increasingly placing large AI campuses outside hubs, requiring new high-bandwidth fiber routes. For enterprises, this means that data centers hosting AI workloads may be far from their headquarters or primary business areas, increasing network latency and operational complexity.
• Operational Complexity: As inference traffic spreads to the edge, enterprises need to deploy distributed network monitoring and AIOps tools to manage the network health across clouds, core data centers, and edge nodes.
• Security and Compliance: Data flow across environments increases the risk of data breaches. Agentic AI may access sensitive enterprise data, so enterprises must ensure transmission encryption, access control, and data sovereignty compliance, especially in cross-border scenarios.### Network and Fiber Operators

• Equinix, Digital Realty, as data center REITs, need to provide more high-density, high-power-capacity colocation facilities, along with dark fiber or cloud connectivity. Their success will depend on their ability to acquire land outside power-constrained areas.
• Lumen Technologies, Zayo, Crown Castle and other fiber operators face a dilemma: building new routes to remote AI campuses requires massive investment, but ROI is uncertain because these routes bypass traditional population centers, making it difficult to capture diverse customers. Wong bluntly stated: "The challenge for fiber operators is whether they can achieve good returns from routes that do not pass through traditional commercial hubs."

Network Equipment Vendors

• Cisco, Arista, Juniper's high-end switches and routing products will benefit from the growth of east-west traffic within data centers and increased DCI demand between data centers. Equipment supporting 800G/1.6T optical modules and intelligent congestion control will be favored.
• NVIDIA is also competing through its InfiniBand and Spectrum-X networking platforms, which are optimized for AI clusters and are being adopted by hyperscalers, but enterprise customers may require more general-purpose Ethernet solutions.

Edge Computing Platforms

• Cloudflare, Fastly, AWS Wavelength and other edge node providers will benefit from the downward migration of Agentic AI inference, but need to offer GPU acceleration capabilities.

Industry Trend Observations

Shifting Network Weight from Training to Inference

Current AI infrastructure capital expenditure is primarily focused on compute (GPU) capacity. Wong points out that every gigawatt of new compute corresponds to corresponding connectivity demand, and as workloads shift from training to inference and Agentic AI, connectivity demand will further increase. This means that the investment growth rate in network infrastructure will catch up with computing investment in the coming years.

Power Bottlenecks Create New Logic for Data Center Site Selection

Globally, the power consumption of AI data centers is drawing attention. Data from Uptime Institute and other organizations shows that the power demand of a single hyperscale AI cluster can reach hundreds of megawatts, while power grid expansion cycles often take 5–10 years. This forces developers to seek non-traditional regions, such as the US Midwest or Northern Europe, but that also brings new fiber deployment requirements.

Token Consumption and AI Efficiency RebalancingWong predicts that while the adoption of AI and Agentic AI will drive sustained significant growth in computing and storage demands, as token usage and costs rise, enterprises and providers will begin to actively manage consumption. Model optimization, inference scheduling, and token compression technologies will become important areas of investment. This indicates that network demand will not expand indefinitely, but will develop in parallel with efficiency improvements.

Sovereign Cloud and Data Localization

Agentic AI involves sensitive enterprise data, and many countries require data to remain within their borders. This drives the development of sovereign clouds and regional data centers, further increasing the complexity of network fragmentation.

CloudTechDaily Insight

Agentic AI is not merely an upgrade of AI algorithms; it fundamentally transforms the interaction model among cloud, network, and computing. This KPMG analysis reveals several key signals: First, the network is no longer a "supporting role" after computing, but stands alongside electricity as one of the three major bottlenecks in AI infrastructure. Enterprise IT strategies must prioritize network planning in advance, synchronizing it with GPU procurement and cloud selection. Second, the decentralization of data center site selection will reshape the fiber network map, leaving operators facing uncertainties in investment returns, while enterprises must endure longer deployment cycles and higher connection costs. Third, the real-time requirements of Agentic AI force enterprises to reassess the edge capabilities of their IT architecture—not all inference needs to go to the cloud, nor all data needs to be centralized. In the future, enterprise architecture will evolve into a three-layer model of "core training + edge inference + multi-cloud data access," posing unprecedented challenges to the strategic vision of CTOs and CIOs. CloudTechDaily believes that 2026 will be a watershed moment for AI network infrastructure investment: those enterprises and providers that deploy high-bandwidth, low-latency connections early will gain structural advantages in the upcoming Agentic AI race.