AI Fabrics × Open Line Systems — The Political Fight Coming in 2027

Two trends in optical infrastructure are about to run into each other. Hyperscaler AI fabrics push aggressive disaggregation — programmable transport, vendor-neutral interfaces, customer-controlled SDN. Carrier DWDM, where most of the long-haul gear lives, is moving in the same direction but with completely different operational assumptions: closed support contracts, vendor-certified configurations, deterministic ops tooling.

These two worlds increasingly need to interconnect. AI training runs cross-datacentre for the largest workloads. The fabric the hyperscaler controls extends into transport the carrier controls. The boundary between them is where the political collision happens. Schedule: 2027.

The two architectures

Hyperscaler AI fabric is a flat high-radix mesh between GPU clusters. Pluggable or co-packaged optics, short-reach wavelengths, hyperscaler-owned SDN. Operational model: "we own end-to-end, we configure ourselves, vendor support is a fallback".

Carrier DWDM is a hierarchy of line systems, amplifier chains, ROADMs, transponder shelves. Vendor-supplied control plane with operator extensions. Operational model: "the NMS tells us something's wrong, we open a ticket, we don't change config without vendor change-control".

Both work in their respective scopes. The collision is at the boundary.

What's actually colliding

Five interop pain points, observable today, getting worse as 2027 approaches.

Wavelength provisioning. Hyperscaler wants programmatic API. Carrier wants ticketed provisioning with change-control. Multi-week change windows don't survive AI fabric extensions. Programmable APIs don't survive the carrier's existing NMS layer.

Telemetry detail. Hyperscaler wants high-frequency OSNR, BER, polarisation, dispersion per wavelength. Carrier's NMS exposes a coarser dataset. The granularity gap is built into the management plane.

SLA structure. Carriers run availability-and-resolution-time SLAs. Hyperscalers run tail-latency and throughput-percentile SLOs. The two measurement structures don't translate cleanly. A link that meets the carrier SLA can fail the hyperscaler SLO.

Failure-domain ownership. A wavelength fails between two hyperscaler sites traversing carrier transport. Carrier sees a transport problem. Hyperscaler sees a fabric problem. The boundary is exactly where ownership is unclear, and resolution times suffer for it.

Optical software stack. Hyperscaler wants their software on the transport equipment. Carrier doesn't allow it. Hyperscaler wants the carrier's optical control plane in their SDN. Carrier doesn't expose the APIs.

Why 2027 specifically

Three factors converge.

Multi-site AI training becomes default. Single-DC frontier models give way to multi-DC training in 2026. By 2027, multi-site is the production pattern, not the corner case. Fabric extension into carrier transport is the standard traffic flow.

Carrier OLS deployments mature. By 2027 the major carriers have OLS in production for new builds and migration of legacy. The "programmable transport" capability hyperscalers need is finally there. Question shifts from "can we do this" to "how do we operate this together".

Standards work lands. OIF and ITU disaggregated optical control plane documents start producing usable specs around 2027. The political negotiation about whose model wins happens at the same time.

Three plausible resolutions

Hyperscaler model wins for cross-site AI fabric. Carriers offer a "programmable wavelength service" tier matching hyperscaler requirements. Carrier-side SLA adapts to performance SLOs. Hyperscalers want this, have the position to push for it.

Carrier model holds for traditional transport. Hyperscaler builds their own private transport for AI-fabric-class connectivity, carrier transport stays the fallback. Carriers prefer this — preserves operational model on bulk revenue, AI fabric becomes separate niche.

Hybrid wins. Carriers offer two tiers: traditional managed transport for general workloads, programmable transport for AI fabric extension. Distinct pricing and ops models. Most plausible. Also messiest.

The hybrid is my best guess. Carriers don't want to lose the AI fabric revenue. Hyperscalers don't want to build transport everywhere. Both have reasons to compromise.

The uncomfortable part

Carriers want operational ownership. Their value proposition depends on operating the transport plane. If hyperscalers run the SDN, carriers become wholesale lambda providers, not managed transport vendors. Margin difference is significant.

Hyperscalers want operational control. Fabric performance depends on tight integration between optical telemetry and SDN decisions. If they can't see the optics in detail, fabric performance degrades. Performance difference is significant.

Both sides are right within their frames. The frame compatibility is the problem.

Three signals to watch in 2027

A major hyperscaler announcing a private long-haul transport build. Signals negotiation broke down. Hyperscaler is deploying their own.

A major carrier announcing a "programmable wavelength service" tier with hyperscaler-style telemetry. Signals carrier-side accommodation. First mover gets either followed quickly or punished by enterprise customers.

OIF or ITU publishing a multi-stakeholder document on disaggregated optical control plane. Signals standards-body convergence is real. Doesn't have to be perfect, just exist as reference point for contracts.

By Q4 2027, the direction should be visible. 2028 procurement plans for both hyperscalers and carriers will reflect whichever way 2027 pointed.

Enterprises buying carrier transport stay neutral on the resolution. The collision doesn't directly affect enterprise procurement until late in the decade. But the carrier-side ops model that emerges trickles down to enterprise contracts eventually.

The negotiations have started already, mostly behind closed doors. The shape of optical transport for the next decade is being decided right now.