LLM Data Transfer: Fiber Optics Key by 2028

Listen to this article · 9 min listen

Industry analysts are projecting that a staggering 80% of all internet traffic will originate from or terminate in a data center by 2028. This data explosion, a lot of it coming from the constant demands of large language models (LLMs), is putting incredible pressure on our infrastructure. If we can’t accelerate how we move LLM data, AI’s progress stalls, and the solution is going to come from better fiber optics and optical transport systems.

Key Takeaways

  • By 2028, global data center IP traffic will hit 33.4 zettabytes a year, and to handle the LLM portion of that, we’ll need to grow backbone network capacity by 1.5x.
  • The 400ZR transceiver is already old news; 800ZR and 1.28T coherent optics are taking over fast to keep up with LLM data rates.
  • Photonic integrated circuits (PICs) are shrinking the size and power draw of optical gear by up to 60%, which is essential if we’re going to scale LLM infrastructure.
  • All-optical switching is still new, but it could cut latency for LLM inference by another 15-20% over today’s electronic packet switching.
  • To avoid vendor lock-in and actually speed up LLM deployments, the industry has to standardize on coherent optics and network automation that work together.

The Zettabyte Era: 33.4 Zettabytes of Data Center IP Traffic by 2028

The numbers are hard to ignore. Cisco’s latest annual internet report projects that global data center IP traffic will hit 33.4 zettabytes a year by 2028, a huge jump from 2023. A massive chunk of that growth is coming straight from LLM workloads, training, inference, and fine-tuning. A single training run can easily be terabytes or even petabytes, and getting that data moved quickly between GPUs, storage, and different data centers is absolutely essential for making any progress. I’ve seen too many projects underestimate this data volume until they hit a wall. Without powerful, high-capacity fiber optic networks, the entire LLM development pipeline simply stops.

To handle that much data, you obviously need more network capacity. This requires both higher speeds at the edge and a full-blown overhaul of the core and metro networks connecting these massive compute clusters. A recent LightCounting analysis says backbone capacity has to grow by 1.5x every two years just to keep from falling behind LLM demand. If you don’t invest in this network layer, you’re throwing away the performance you paid for in the model itself. The system is always limited by its slowest part, and a supercomputer that’s starved for data isn’t a supercomputer at all, it’s just a very expensive space heater.

800ZR Transceivers: The New Baseline for Inter-Data Center Connectivity

It feels like just yesterday that 400ZR transceivers were the new thing for data center interconnect (DCI) applications, but by mid-2026, 800ZR is the new standard and we’re already seeing 1.28T on the horizon. That’s how fast LLMs are pushing bandwidth demand. We’re getting these speeds over distance because of coherent optics, which pack data onto both the amplitude and phase of the light wave. A 400ZR module gets you 400 Gigabits per second (Gbps) over about 120 kilometers, but LLM clusters need far more total bandwidth than that. So, we’re turning to 800ZR and now 1.28T solutions to double or triple capacity on the same number of fibers.

The upgrade to 800ZR brings both raw speed and much better spectral efficiency, which is just as important. The pressure on existing fiber is intense because deploying new fiber is a slow, expensive headache. Coherent optics let us use advanced modulation like 16QAM or 32QAM to cram more bits into each symbol, squeezing more throughput from every optical channel. This lets operators get more out of the fiber they already have in the ground, a huge deal for anyone trying to manage costs. I’ve seen projects where a smart upgrade to faster coherent optics bought them several years before they had to think about trenching again, delivering an immediate ROI for their LLM initiatives.

Photonic Integrated Circuits (PICs): Shrinking the Optical Footprint by 60%

People often forget about the physical size and power draw of optical gear, but these are huge factors in whether you can actually scale an LLM environment. That’s where photonic integrated circuits (PICs) are making a real difference, cutting the size and power of transceivers and switches by up to 60% compared to discrete optical components. A PIC puts all the optical jobs, lasers, modulators, detectors, onto a single chip, just like an electronic IC. This shrinking of hardware directly leads to more efficient data centers.

Think about the power costs. Cooling packed server racks is a huge line item on any OpEx budget. Since PICs reduce the power needed for optical transport, they directly lower energy bills and help with sustainability goals by shrinking the carbon footprint. Their smaller size also means you get higher port density in your switches and routers, letting you connect more LLM servers in the same rack. That density is everything for hyperscalers trying to cram as much compute as possible into every square foot. Without PICs, the physical space required for all the optical gear to handle the expected LLM data flows would quickly become unmanageable.

The Latency Conundrum: All-Optical Switching for 15-20% Lower Latency

Everyone talks about bandwidth, but latency is just as important for real-time LLM inference and GPU-to-GPU chatter, and all-optical switching is promising to slash it by another 15-20% compared to standard electronic packet switching. Your typical network switch has to convert light to electricity to process a packet, then convert it back to light to send it on its way. Each one of those OEO (optical-to-electrical-to-optical) conversions adds a bit of delay, and it adds up fast. For a conversational AI or a complex multi-modal model, a few extra milliseconds can kill the user experience and cripple the model’s performance.

All-optical (or photonic) switches try to avoid that OEO conversion completely by keeping the data as light the whole time. It’s still a pretty new technology, especially for big deployments, but progress in silicon photonics is making it a real possibility. Think about it: a network where light signals are routed directly without ever becoming electrons. This gets rid of the OEO delay and the power burned during those conversions. Of course, managing and routing optical signals at that scale is incredibly complex. But the payoff for latency-sensitive LLM applications is huge. I’m convinced that as LLMs get embedded deeper into real-time systems, we’re going to see all-optical switching go from a niche idea to a core requirement.

The Conventional Wisdom Misses the Interoperability Challenge

Most of the talk in the industry is about the raw speed of new optical tech. And sure, speed is great, but everyone seems to be missing a huge problem: interoperability. The assumption is that as long as vendors keep shipping faster transceivers and better components, we’re all set. In my experience, that’s a massive oversimplification. If we don’t get serious about multi-vendor standards for things like coherent optics and automation protocols, we’re headed for a world of fragmented networks and vendor lock-in that will absolutely slow down how we deploy and scale LLM infrastructure.

Open standards like the OIF’s definitions for 400ZR and 800ZR are a good start, but the real work is in the implementation. Vendors can interpret the specs in slightly different ways, and you don’t find the compatibility problems until you’re trying to get everything to work together during integration. It’s a nightmare for operators trying to build a network with gear from more than one company. We need more than just paper standards. What we really need are tough, multi-vendor plugfests and certification programs to prove things actually work together at both the physical and protocol layers. If the industry doesn’t make this a priority, all this fast optical gear for LLMs will get bogged down by the mess of trying to integrate systems that won’t talk to each other which means slower deployments and more operational pain. Verifiable interoperability needs to be just as important as raw performance.

These advances in fiber optics aren’t just small steps forward. They are the foundation for the future growth of LLMs. As the demand for LLM processing and data transfer keeps climbing, the network underneath has to evolve faster than ever before. In the end, the organizations that invest smartly in high-capacity fiber, next-gen coherent optics, and power-saving PICs will be the ones that can actually get the most out of their AI. In this race, the network has become a genuine accelerator, not just the plumbing.

What is optical transport in the context of LLMs?

It’s using light signals over fiber optic cables to move the massive datasets required for LLMs. This involves connecting servers, data centers, and storage with the high speed and low latency needed for training and inference.

Why are fiber optics particularly important for LLM data transfer?

Compared to old-school copper cables, fiber gives you way more bandwidth with less signal loss over long distances. LLMs chew through petabytes of data, so you need a network that can handle terabits per second without slowdowns or delays, and only fiber can do that.

What is a coherent optical transceiver and why is it relevant for LLMs?

It’s a transceiver that uses fancy modulation to pack more data onto a light wave by encoding it on both the amplitude and phase. This is how we get 400 Gbps, 800 Gbps, and even 1.28 Tbps speeds over long-haul fiber, which is perfect for linking the distributed data centers that run LLMs.

How do Photonic Integrated Circuits (PICs) benefit LLM infrastructure?

They put a bunch of optical parts onto one chip. This makes networking gear smaller, cheaper, and more power-efficient, which means you can build more compact and scalable hardware for dense LLM data centers.

What role does network latency play in LLM performance?

Latency is the delay in data getting from point A to point B. For LLMs, especially in real-time chats or distributed training, high latency means slow model responses and inefficient GPU communication. Lowering latency directly improves user experience and makes training cycles faster.

Editorial Team

The Editorial Team at LLM Growth is a group of writers and editors who cover technology, products, and how teams ship. They research each story and publish under a shared desk byline — experts in their fields, working as one team.