Netflix Open Connect Explained: Architecture and Delivery in 2026

During Q1 2026, Netflix accounted for roughly 15% of global downstream internet traffic — a figure that climbs past 30% in some North American ISP networks during evening peak hours. Nearly all of that volume bypasses the public internet entirely. It moves through Netflix Open Connect, a purpose-built content delivery network that embeds hardware directly inside ISP facilities across more than 6,000 locations in over 100 countries. This article breaks down the Open Connect architecture as it operates in 2026, examines what changed from earlier iterations, maps the failure modes that matter, and offers a decision framework for engineers evaluating whether a first-party CDN build makes sense versus leveraging existing commercial infrastructure.

How Netflix Open Connect Works in 2026

The core premise has not changed since Open Connect launched in 2012: push bytes as close to the eyeball as physics allows. What has changed is the scale, the hardware, and the control plane sophistication.

Appliance Tiers

As of 2026, Netflix deploys two primary Open Connect Appliance (OCA) form factors. The storage-dense appliance packs up to 560 TB of NVMe flash in a 2U chassis, designed for ISP sites that serve large subscriber populations with deep catalog depth. The throughput-optimized appliance targets peering facilities and exchange points, pushing 400 Gbps from a single server using custom FreeBSD-based software and kernel-bypass networking via DPDK. Netflix reported in its most recent Open Connect documentation that a single appliance now serves upward of 100 Gbps of sustained throughput during peak, a figure that has roughly doubled since 2022.

Content Placement and Predictive Fill

Netflix's control plane determines what content lives on each OCA. This is not simple popularity-based caching. The fill algorithm incorporates regional viewing patterns, upcoming release schedules, personalized recommendation model outputs, and time-of-day demand curves. By the time a subscriber presses play, the title is already resident on the local appliance with high probability. For new releases, Netflix pre-positions encoded assets hours before the global launch window, coordinating across thousands of appliances simultaneously.

Steering and Failover

Client devices receive a ranked list of OCA endpoints from Netflix's steering service (Open Connect's equivalent of DNS-based request routing, but operating at the application layer). If the primary OCA is degraded, the client falls back to secondary or tertiary appliances, potentially at a peering location or even a transit-connected site. This steering is informed by real-time telemetry: per-session throughput, appliance CPU and disk utilization, and network-level health signals. The fallback chain is deterministic and tested continuously.

Netflix ISP Peering and the Settlement-Free Model

Open Connect operates on a settlement-free interconnection basis. Netflix does not pay ISPs for OCA hosting, and ISPs do not pay Netflix. The value exchange is straightforward: the ISP eliminates enormous volumes of transit traffic (which it would otherwise pay for), and Netflix gets lower latency and higher throughput to that ISP's subscribers.

This model works because Netflix's traffic profile is overwhelmingly asymmetric and predictable — large objects, downstream only, with well-understood demand curves. For ISPs, accepting an OCA is almost always a net cost reduction. As of 2026, Netflix maintains peering at over 200 Internet Exchange Points globally and has direct private network interconnection with thousands of ISPs. The program remains open to any ISP meeting minimum traffic thresholds, typically a few hundred Mbps of Netflix traffic.

What Changed in 2025–2026

Several architectural shifts are worth noting for engineers tracking this space:

• AV1 at scale: Netflix completed its migration to AV1 as the default codec for most device classes in late 2025. This reduced per-stream bitrate by 20–30% compared to VP9 and H.264 at equivalent perceptual quality, directly reducing the storage and bandwidth demands on each OCA.
• QUIC adoption: Netflix's client-side stack now uses QUIC for the majority of streaming sessions, improving connection migration on mobile networks and reducing tail latency during congestion. The OCA software stack was updated to support this natively.
• Increased appliance density: The 2026-generation appliance ships with Gen5 NVMe drives and PCIe 5.0, enabling the 560 TB capacity in a 2U footprint. This is relevant for ISPs with limited rack space.
• Live events infrastructure: With Netflix expanding into live sports and events through 2025–2026, the Open Connect team added low-latency live ingest paths. These bypass the standard pre-fill pipeline and use a real-time push model from origin to edge OCAs, targeting sub-four-second glass-to-glass latency.

Failure Modes and Production Resilience

No infrastructure discussion is complete without examining how the system fails. Open Connect is designed around the assumption that individual appliances are expendable. Here are the failure scenarios that matter:

OCA Disk Failure

Each appliance runs a custom storage layer that stripes content across drives without traditional RAID. When a drive fails, the appliance continues serving content from remaining drives. Missing content is re-fetched from upstream fill sources during the next fill window. Netflix accepts temporary catalog gaps on a single appliance because the steering layer routes around them.

Full Appliance Outage

If an OCA goes offline entirely — power loss, motherboard failure, network partition — the steering service detects the absence within seconds via health-check failures and removes it from the candidate list. Subscribers are redirected to the next-nearest appliance. In dense ISP deployments, this might be another OCA in the same facility. In sparse deployments, it could be a peering-site appliance or even a cross-region fallback, with corresponding latency increase.

ISP-Wide Connectivity Loss

When an ISP loses upstream connectivity entirely, locally cached content on embedded OCAs remains available to subscribers within that network. This is one of the underappreciated resilience benefits of the embedded model: the OCA functions as a local cache even during upstream outages, serving whatever is already on disk. New content and catalog updates will not arrive until connectivity restores, but in-progress sessions and cached titles continue to work.

Control Plane Partition

If OCAs lose contact with Netflix's central control plane, they continue serving content using their last-known configuration. Fill operations pause, but read-path serving is unaffected. This is a deliberate design choice: the data plane is decoupled from the control plane, and the appliances are built to operate autonomously for extended periods.

Build vs. Buy: A Decision Framework for CDN Strategy

Open Connect is a compelling case study, but it is not a template most organizations should copy directly. Building a first-party CDN at Netflix's scale requires dedicated hardware engineering, custom OS development, a global logistics operation for appliance deployment and maintenance, and the negotiating leverage that comes with representing 15%+ of an ISP's traffic. The economics only work because Netflix's content catalog is finite, its traffic is write-once-read-many, and its scale justifies the capital expenditure.

For organizations delivering video, software, or large assets at scale but without Netflix-level traffic concentration, the calculus favors commercial CDN infrastructure. The key variables in that decision:

For media companies, game publishers, and SaaS platforms operating in the 25 TB to 2 PB monthly range, commercial CDNs offer the edge coverage and fault tolerance without the capital and operational burden of a bespoke build. BlazingCDN's media delivery infrastructure delivers stability and fault tolerance comparable to Amazon CloudFront while pricing starts at $4 per TB and scales down to $2 per TB at 2 PB+ commitments — a cost structure that puts enterprise-grade delivery within reach of teams that do not have Netflix's engineering budget. Sony is among its clients, and the platform supports 100% uptime SLAs with fast scaling under demand spikes.

FAQ

How does Netflix Open Connect work at the network level?

Netflix embeds appliances (OCAs) inside ISP networks via settlement-free interconnection agreements. The OCAs are filled with content during off-peak hours based on predictive demand models. When a subscriber initiates playback, Netflix's steering service directs the session to the optimal local OCA based on real-time health and proximity data.

What are Netflix Open Connect Appliances, and who maintains them?

OCAs are purpose-built 2U servers running a custom FreeBSD stack, optimized for high-throughput storage and network I/O. As of 2026, they hold up to 560 TB of NVMe flash and can sustain 100+ Gbps of output. Netflix owns and manages the software remotely; the ISP provides rack space, power, and network connectivity.

How does Netflix ISP peering differ from traditional CDN peering?

Traditional CDNs peer at exchange points and deliver traffic to ISPs over those links. Netflix goes further by embedding servers inside the ISP's own network, eliminating even the last peering hop. This is closer to a private cache than a traditional CDN edge node. The settlement-free model means no money changes hands for the interconnection itself.

Can other companies replicate the Open Connect model?

Technically, yes. Practically, it requires the traffic scale to make ISP partnerships worthwhile, the engineering capacity to build and maintain custom appliance software, and the operational logistics to deploy and support hardware in thousands of locations. For most organizations, a multi-CDN strategy using commercial providers achieves comparable end-user performance without the capital investment.

How does Netflix handle live streaming on Open Connect in 2026?

Live events use a real-time push model from Netflix's origin infrastructure to edge OCAs, bypassing the standard batch-fill pipeline. The target is sub-four-second glass-to-glass latency. This is a newer capability, added to support Netflix's expansion into live sports and events during 2025–2026.

Next Step: Benchmark Your Own Edge Delivery

If you operate a delivery pipeline serving video, large binaries, or high-volume static assets, here is a useful exercise for this week: instrument your cache-hit ratio by edge location and overlay it with your per-TB egress cost. The ratio between those two numbers — effective cost per cache-hit TB — is the single best indicator of whether your current CDN topology is right-sized. If you find locations where cache-hit ratios drop below 85% while costs stay flat, you have a steering or fill problem worth investigating before your next traffic spike.