Public Cloud Capacity Is Not Guaranteed: Why Organizations Need Hybrid AI Infrastructure

By Rena Erez, Chief Marketing Officer, PacketLight Networks

TL;DR

Public cloud remains central to AI, but capacity is not always available in the required region, at the required scale or within the required timeframe.
Data sovereignty, security and infrastructure control are also keeping sensitive workloads and proprietary data in private facilities.
The result is hybrid AI infrastructure spanning private data centers, colocation facilities, edge locations and multiple clouds.
A secure optical transport layer connects the locations where compute and data are available while preserving visibility, security and control over data in motion.

Public cloud capacity is no longer an unlimited assumption

For years, infrastructure plans were built around a simple assumption: when more compute was needed, additional public-cloud capacity would be available. AI is testing that assumption.

In a July 2026 Sourcery interview, MongoDB CEO CJ Desai described a Fortune 100 enterprise in Texas that had begun decommissioning a data center. When its hyperscaler could not provide the additional capacity it needed, including capacity for AI workloads, the company retained the facility and decided to run the workloads on its own infrastructure.

Desai also described a US telecommunications company that added a second hyperscaler after its existing provider could not supply capacity in the required region.

Neither case represents a rejection of public cloud. Both demonstrate why organizations need more than one viable location for their AI workloads.

Two reasons organizations are retaining private AI infrastructure

1. Capacity may not be available where it is needed

AI capacity depends on suitable data-center space, high-density power and cooling, accelerators, networking and regional availability. Capacity in another region may not solve the problem. Data residency, sovereignty, regulatory or contractual requirements may restrict where sensitive data can be stored and processed. Latency-sensitive workloads may also require compute to remain close to users, data sources or operational systems.

A cloud service may therefore be available globally, but not in the required location, at the required scale or within the required timeframe.

The International Energy Agency projects that global data-center electricity consumption will more than double by 2030. It also estimates that grid constraints could put approximately 20% of planned data-center projects at risk of delays. Uptime Institute identifies forecasting capacity and securing power as growing management concerns.

2. Data sovereignty and control remain essential

Capacity is the newer pressure, but sovereignty remains fundamental. Organizations may need to keep data within an approved jurisdiction, restrict access, protect intellectual property or maintain direct operational control. Desai also cited a customer in France that kept its environment off the public cloud because data sovereignty was its highest priority.

Sovereignty applies to data in motion as well as data at rest. When sensitive information moves between a private data center, a colocation site and a cloud-interconnection facility, the transport layer also requires appropriate security, visibility and control.

What is hybrid AI infrastructure?

Hybrid AI infrastructure is an architecture in which AI workloads, applications and data operate across private data centers, colocation facilities, edge locations and one or more public clouds.

This model preserves choice. Sensitive data can remain in a controlled environment, training or inference can run where accelerators and power are available, and cloud services can still provide elasticity, specialized models and regional reach.

Uptime Institute reported in 2025 that 45% of IT workloads still resided in corporate facilities, confirming that private infrastructure remains a foundation of hybrid IT.

These locations are useful only if they can work together. That makes data center interconnect part of the AI architecture, not an afterthought.

What the optical network must deliver

Optical networking does not create compute capacity. It connects the locations where capacity and data are available, allowing organizations to place workloads without surrendering control of the transport layer.

Scalable bandwidth: DWDM carries multiple high-capacity wavelengths over a fiber pair to support sustained traffic between facilities.
Direct interconnect: Dedicated optical capacity can link private data centers, colocation sites and facilities where cloud interconnection is available.
Visibility and resilience: Performance monitoring, alarms, protection and restoration help operations teams maintain services across the transport path.
Flexible routing and restoration: ROADMs allow wavelengths to be remotely added, dropped or rerouted as connectivity requirements change. In networks with alternate paths, automatic channel restoration can redirect affected services following a fiber or route failure.
Data-in-motion security: Layer-1 optical encryption secures traffic between sites at full throughput, without consuming server resources or requiring application-level changes.
Room to grow: Additional wavelengths or compact systems can add capacity without requiring a wholesale network redesign.

How PacketLight Networks supports hybrid AI infrastructure

PacketLight Networks provides compact, carrier-grade DWDM and OTN systems for high-capacity connectivity between private data centers, colocation facilities and cloud-interconnection locations.

PacketLight’s compact 1U systems support service rates from 10G to 800G and scale to as much as 8T per device. Integrated transport and management capabilities provide visibility into services, performance, optical conditions and faults. Organizations can start with the capacity they need, then add wavelengths or deploy additional compact systems as traffic grows.

For sensitive and regulated traffic, relevant PacketLight platforms provide built-in Layer-1 AES-256 encryption. This secures data in motion at full throughput without consuming server compute resources or requiring changes to the applications being transported.

PacketLight solutions also support future quantum-security strategies through QKD security and partner-enabled post-quantum cryptography. QKD, PQC or a hybrid approach can be incorporated according to the network architecture, risk profile and security requirements.

Support for alien wavelengths and open interfaces can enable integration with compatible third-party DWDM infrastructure. This helps organizations use suitable existing assets and reduces unnecessary dependence on a single optical-system supplier.

Five questions to ask before fixing AI workloads in one location

Can the required compute be obtained in each target region and within the required timeframe?
Which workloads or data sets must remain in a private facility or approved jurisdiction?
Which workloads require compute close to users, data sources or operational systems?
Can private, colocation and cloud environments be connected without redesigning the network around one provider?
Do the available fiber and DWDM infrastructure provide the required capacity, encryption, resilience, monitoring and room for future growth?

From cloud-first to capacity-ready

The lesson is not that public cloud has failed. It remains indispensable to modern IT and AI. The lesson is that infrastructure plans need more than one viable path when the preferred cloud, region or facility does not have sufficient capacity available.

A capacity-ready strategy preserves private and colocation options, uses more than one cloud where justified and treats connectivity as a strategic layer.

A secure, scalable optical network enables these locations to function as a distributed AI infrastructure while helping organizations maintain control over their data, transport capacity and infrastructure choices.

Frequently asked questions

Why are organizations retaining private data centers for AI?

Private facilities provide greater control over data, security and operations. They also provide another location for AI workloads when public-cloud capacity is unavailable in the required region, at the required scale or within the required timeframe.

Does hybrid AI infrastructure replace public cloud?

No. It combines public cloud with private data centers, colocation facilities and, in some architectures, edge locations. Workloads can run where capacity, data residency, security, latency and operational requirements can best be met.

What is the role of DWDM in AI infrastructure?

DWDM transmits multiple high-capacity wavelengths over a fiber pair. In a distributed AI environment, it can provide scalable connectivity between private data centers, colocation sites and facilities offering cloud interconnection.

How does optical networking support data sovereignty?

A private optical network gives organizations greater visibility and control over data moving between approved locations. Layer-1 encryption protects data in motion, while performance monitoring supports operational oversight. Combined with data-residency policies, access controls and compliance governance, optical networking supports a broader data-sovereignty strategy.

How does PacketLight Networks secure data in motion?

Relevant PacketLight platforms provide built-in Layer-1 AES-256 encryption at full throughput without consuming server compute resources. PacketLight solutions can also support quantum-security strategies through QKD readiness and partner-enabled PQC. The final design depends on the selected platforms, topology, risk profile and compliance requirements.

Can existing DWDM infrastructure support new AI capacity?

Often it can, subject to an optical design review. The channel plan, passband, channel width, fiber condition, amplification and end-to-end optical margin must support the selected wavelength and data rate.

Sources

Sourcery, “Breaking: Inside AI’s Biggest Downstream Winner: The Surge in AI Database Demand,” interview with MongoDB CEO CJ Desai, July 20, 2026.
International Energy Agency, Energy and AI, April 10, 2025.
Uptime Institute, Uptime Institute Global Data Center Survey 2025, UII Keynote Report 180, July 2025.