Meta's Petal Cable: The First 1 Petabit-per-Second Transatlantic Link

Nearly all intercontinental internet traffic runs through cables on the ocean floor, and Meta just announced the biggest single jump in that capacity anyone has built. On September 21, 2026, Meta's engineering team published details on Petal — the first transoceanic subsea cable designed for petabit-class capacity and the first to deploy multi-core fiber at scale. Spanning roughly 7,000 km between France and the United States, Petal will carry 1 petabit per second (1,000 terabits per second) once it enters service in 2029, doubling what today's fastest transatlantic cable systems can move.

This is telecom infrastructure, not an AI product — but it's the kind of infrastructure that AI compute demand increasingly rides on, since moving datasets, model weights, and inference traffic between continents is bounded by exactly this kind of physical capacity.

TL;DR — what people are actually asking

Why capacity per fiber, not just more fibers

Approximately 99% of intercontinental data traffic — nearly every cross-continental message, call, or video stream — travels through glass fiber strands on the ocean floor, not satellites. Since the 1980s, the industry has scaled subsea capacity through a series of step-changes: erbium-doped fiber amplifiers, then coherent optical transmission in the 2010s that drove 10x-plus capacity gains, and more recently spatial division multiplexing (SDM) — simply adding more fiber pairs per cable. Meta's own cable history traces that path directly: Marea (its first transatlantic investment) shipped with 8 fiber pairs, Amitié with 16, and Anjana with 24 — the first 0.5 petabit-per-second transatlantic system.

Petal breaks from that pattern. Instead of scaling to 48 physical fiber pairs to double Anjana's capacity again, Meta opted for 2-core fiber — a single fiber strand engineered to carry two independent light-propagating cores — in a 24-fiber-pair cable that's the capacity equivalent of 48 pairs. The stated payoff isn't just capacity; it's efficiency. Meta's own comparison is that one petabit-class cable built this way uses meaningfully less material and carbon than building two separate 0.5-petabit cables to hit the same combined capacity — fewer repeaters, less power infrastructure, less physical cable laid across 7,000 km of ocean floor.

The engineering problem 2-core fiber actually solves

Two specific challenges had to be solved to make 2-core fiber viable at this scale, and both are about keeping the two cores from interfering with each other over thousands of kilometers:

1. Low attenuation at standard dimensions. The fiber still has to hold to the industry-standard 125-micrometer outer width — about the diameter of a human hair — while adding a second core, achieved using ultra-pure synthetic silica in manufacturing.
2. Minimizing crosstalk between the two cores. Achieved by carefully controlling refractive index contrast between each core and the surrounding medium, plus counter-propagating the optical signals in each core — sending light in opposite directions through the two cores rather than the same direction — which Meta says results in "nearly immeasurable" crosstalk.

The repeater design is its own piece of engineering: rather than trying to amplify both cores together, Petal's single-body 96-amplifier repeater uses a Fan-In/Fan-Out (FIFO) interface to temporarily split the 2-core fiber into two ordinary single-core fibers at each repeater, amplify each with well-understood single-core amplification technology, then recombine them back into 2-core fiber. That design choice — reusing proven single-core amplification rather than inventing new 2-core amplification from scratch — is what lets Petal stay within existing subsea power-feeding limits (up to 18kV) instead of requiring a new hardware qualification process across the entire subsea ecosystem.

The partners actually building it

Petal is a three-way engineering effort, not a solo Meta project:

• NEC is the turnkey system supplier — engineering and qualifying the full petabit-class system (cable, repeaters, FIFO interfaces, powering), and responsible for manufacturing and installing the finished cable. NEC's own Eduardo Mateo called it "a major technological milestone in the history of global telecommunications," reflecting sustained R&D investment in petabit-class SDM repeaters specifically.
• Sumitomo Electric Industries developed and manufactures the 2-core fiber itself — its "2C Z-PLUS ULL Fiber" — bringing roughly four decades of ultra-low-loss submarine fiber manufacturing experience to a genuinely new fiber geometry.
• Orange is supporting the French landing and the terrestrial interconnection into the broader European network, extending Petal's reach beyond the single transatlantic hop.

The three paths Meta considered, and why it picked this one

Meta's own engineering writeup names three concrete options it weighed for reaching petabit-class capacity on this route, which is a useful way to see why 2-core fiber won over the more conventional alternatives:

1. Scale fiber-pair count conventionally — push from Anjana's 24 fiber pairs to 48, the straightforward extension of the SDM approach the industry has used for a decade.
2. Expand the optical transmission band — add the L-band on top of the existing C-band, the same approach Meta used on its PLCN cable, resulting in a 24-fiber-pair system with combined C+L transmission.
3. Adopt 2-core fiber — the path Meta actually chose, packing the equivalent of 48 fiber pairs' capacity into a 24-fiber-pair physical system.

The first two options are proven, lower-risk extensions of existing subsea technology. The third is the one that required Sumitomo Electric to develop genuinely new fiber and NEC to engineer a new repeater architecture (the FIFO system) around it — a materially higher-risk bet that Meta is explicitly framing as worth taking specifically because of the material and carbon savings of building one cable instead of two to hit the same total capacity. That's a meaningful signal about how the economics of subsea infrastructure are shifting: the cost of laying and powering physical cable across an ocean is now enough of a constraint that a harder engineering problem (new fiber physics) can beat an easier one (more of the same fiber) on total lifecycle cost.

Why this matters if you build or run AI infrastructure

Meta's own announcement frames Petal around general internet demand and connectivity resilience, not AI training specifically — and it's worth taking that framing at face value rather than assuming every large infrastructure announcement in 2026 is secretly an AI story. That said, the practical relevance to anyone running AI workloads at scale is real, even if indirect:

• Cross-region model and data sync. Any organization training or serving models across US and European data centers is bound by the actual physical bandwidth available between them — this is the layer beneath "just replicate the dataset to another region," not an abstraction that scales for free.
• Inference latency for a global user base. Serving a model with acceptable latency to users on both sides of the Atlantic depends on backbone capacity like this existing and not being a bottleneck, independent of how fast the model itself runs.
• This is a multi-year infrastructure bet, not a quarter-to-quarter capacity fix. A 2029 in-service date means Petal is a response to demand curves Meta is modeling years out, not a reaction to current congestion — a useful data point for anyone trying to gauge how seriously infrastructure players are taking long-term cross-continental data demand, AI-driven or otherwise.

For readers specifically tracking AI compute infrastructure rather than telecom engineering, the honest takeaway is narrower than the headline capacity number suggests: Petal is Meta continuing to build the general-purpose internet backbone it has invested in for years, at a genuinely new order of technical capability, in a period where AI is one of several forces increasing demand for exactly this kind of bandwidth — not an announcement built around AI training or inference as its primary justification.

What Petal doesn't tell you

A few honest limitations worth naming before treating this as a settled capacity story:

• 2029 is a long lead time. Nothing about Petal changes transatlantic bandwidth availability today or in the next several years — it's a bet on 2029-era demand, not a near-term fix for any current congestion.
• No pricing or wholesale-capacity terms were disclosed. Meta's own cables have historically been used for Meta's own traffic plus some wholesale capacity sold to other network operators; Petal's announcement doesn't specify how much of its capacity, if any, will be available to third parties versus reserved for Meta's own internal traffic.
• The AI-demand framing in this post is inference, not Meta's stated rationale. Meta's own writeup discusses general internet and communications growth, not AI training or inference specifically — the connection to AI infrastructure demand drawn in this post is a reasonable but independent read of what cross-continental bandwidth gets used for in 2026, not something Meta itself is claiming as Petal's primary purpose.
• This is one route. Petal covers France-to-US specifically; it says nothing about capacity constraints on other major routes (transpacific, Asia-Europe) that also carry significant AI-related cross-region traffic.

Related on explainx.ai

• What Is Terafab AI? Elon Musk's Compute Infrastructure Bet
• SpaceX AI-1: Solar Orbital Datacenter, Space Compute
• Zuckerberg: AI Future for Everyone (WSJ Essay)
• Meta Hatch: AI Agent Inference Spend
• Stanford Accelerator Cost Breakdown

Primary source: Meta Engineering, Inside Petal: Building the World's First Petabit-Class Transoceanic Subsea Cable (September 21, 2026)

This post reflects Meta's own September 21, 2026 engineering announcement. Petal is expected to enter service in 2029; specifications, partners, and capacity figures are as stated in Meta's announcement and have not been independently verified by explainx.ai. Follow @explainx_ai for updates.