Fiber connectivity for hyperscalers

AI Data Centers Eat Fiber Faster Than We Can Build The AI infrastructure race started with GPUs. Now, it is rapidly becoming a fiber and connectivity race. Hyperscalers and content/cloud networks already account for roughly 75% of global used international bandwidth. And AI is adding another major layer of demand. That pressure is spreading across the entire network: → Inside data centers → Between data centers → Across metro and terrestrial backbones → Across subsea networks Optical technology is moving fast: 400G → 800G → 1.6T → 3.2T We can squeeze more capacity from existing fiber with faster optics, more wavelengths and better spectral efficiency. But eventually, we need more physical infrastructure. More fiber. More routes. More subsea capacity. And the investment cycle is already accelerating. Subsea cable investment is expected to exceed $13B between 2025 and 2027, roughly double the previous three-year period. At the same time, hyperscalers are increasingly building global connectivity infrastructure themselves. Meta’s Project Waterworth plans more than 50,000 km of subsea cable across five continents, while Google continues expanding its own terrestrial and subsea footprint. Something fundamental is changing: Hyperscalers are no longer only customers of telecom infrastructure. They are increasingly becoming builders and owners of it. The AI boom may have started inside the data center. But its impact will reach all the way to the ocean floor. #AI #DataCenters #FiberOptics #Subsea #Telecom #DigitalInfrastructure

From T1 to 400G On Friday, we turned up our first 32 × 400G optical wave system for a hyperscaler.  From P.O. to live traffic, the project took less than 30 days.  It certainly helps that Gigabit Fiber has next generation fiber infrastructure strategically prepositioned, but this feat would have been unimaginable during the era of T‑carrier circuits provided by the incumbent monopolies (think AT&T and Verizon), when provisioning a T-1 could take months or even years. Today, hyperscale customers expect terabit‑scale connectivity on demand, and modern optical transport makes it possible. In the 1970s a T‑1 (DS1) could carry 24 voice channels or 1.544 Mb/s of data. Yes, I was alive in the 1970’s, but when I started in telecom in the late 1990’s, T‑1s were still the standard wholesale circuit.  Fast forward to today: a 400G wavelength moves 400 Gb/s, roughly 260 000× the throughput of a T‑1. Aggregate 32 of these massive pipes and you get 12.8 Tb/s — that’s more than 8 million times a T‑1. Why is 400G the new baseline? Economics & efficiency: 400G ports follow the IEEE standard and deliver four times the throughput of 100G while using fewer ports and less power.  One 400G interface costs less and draws less energy than four 100G interfaces. Demand is exploding: Zayo’s 2025 Bandwidth Report shows customers adopting 400G for its cost efficiency, space savings and future‑proofing.  By 2024, 400G capacity purchases accounted for the largest share of total terabits purchased, and software/tech companies sourced 82.15 % of their wavelength capacity in 400G increments. Fiber boom: In the same report, Zayo notes a 600% increase in metro dark‑fiber purchases from 2020–2024, including a 268% year‑over‑year spike in 2024. Hyperscalers account for more than half of Zayo’s metro dark‑fiber purchases, as they want control over massive bandwidth pipes as their critical functions are built upon this foundational infrastructure. As AI, machine learning and cloud workloads surge, 400G has become the new baseline. It offers enough headroom for modern applications and scales easily to terabit levels by combining wavelengths and today 400G waves are the new T‑1 – the standard building block for next‑generation networks. What this means for our industry The era after the breakup of AT&T crowned the T‑1 circuit as the gold standard. Over thirty years later, I am still in the industry and the 400G has taken that mantle. The technology, economics and demand all points to the same conclusion: 400G waves are the new T‑1, and the providers who can deploy them quickly will lead the next phase of the bandwidth boom.

🚨 What’s the AI Data Centre World Missing? Right now, the AI infrastructure race is laser-focused on power: More megawatts, more cooling, more grid deals. But there’s a quieter critical risk few are addressing: Old fibre networks that can’t keep up. Because AI isn’t just about speed. It’s about scale and connectivity across every layer. Here’s where the game has changed: → Thousands of GPUs need to communicate across racks, rows and even buildings ... simultaneously. → Copper? Forget it! Hits hard limits on distance and density. → Fibre is the backbone. But most networks were built for yesterday’s workloads, not today’s AI-scale east-west traffic. → It’s not just about laying more fibre. It’s also about lasers, transceivers, cabling density and optical design. 💥 Your compute might be world-class. But if the network can’t move the data, it’s just silicon waiting around. 💣 Where It Goes Wrong (and I’ve seen it firsthand): → GPU clusters stalled for weeks due to unavailable high-density fibre paths. → Massive power upgrades… while backend teams scrambled to redesign networks. → Cooling chaos caused by underestimated transceiver density. → Millions in sunk costs because the fibre layer was an afterthought. ✅ I just published this article diving deep into why fibre is becoming the real bottleneck of hyperscale AI. I also explain how operators and investors can steer clear of costly errors. If you care about data centre performance, resilience, and ROI, this is a conversation you can’t afford to miss. Are you seeing these challenges where you work? What’s been your biggest network constraint in scaling for AI? 💬 Let’s compare notes. Your shared experiences might help someone else dodge an expensive failure.

🚨 Big milestone in photonic interconnects: Lightmatter has demonstrated a 16-wavelength bidirectional optical link on standard single-mode fibre. Why this matters: 🔹 Hyperscale AI infrastructure is hitting bandwidth and radix limits faster than expected. 🔹 Traditional co-packaged optics (CPO) solutions can’t keep pace with trillion-parameter models and emerging workloads. 🔹 By delivering 800G bidirectional bandwidth per fibre with robust thermal and polarisation performance, Lightmatter just set a new benchmark for fibre bandwidth density. For hyperscalers, this means more capacity on existing fibre, reduced CapEx and OpEx, and a path to higher scalability without waiting for exotic infrastructure. At 650 Group, LLC, we see this as a game-changer for data center efficiency and scalability. It brings advanced CPO closer to commercial reality and directly addresses one of the most pressing bottlenecks in AI development: networking. https://lnkd.in/gDGpvn7G #AI #Datacenters #Networking #Optical #CPO #Hyperscale

The Hyperscale Difference: Why AWS, Azure, and Google Cable Differently It's not just clean wiring. The way these hyperscalers design their Optical Fiber Containment is a direct reflection of their core network strategy: AWS ☁️ (Resilience Scheme): Design mandates focus on physical separation and redundancy across Availability Zones (AZs). Their containment ensures multiple, distinct pathways to guarantee maximum uptime. Show extensive, robust wire mesh baskets and traditional ladder trays carrying dense fiber runs, reflecting their scale and foundational infrastructure. Azure 🟦 (Compliance Scheme): Design prioritizes standardization and zoning for enterprise integration. Their containment facilitates auditable pathways and clear boundaries for hybrid cloud and high-security zones. Focus on structured, rigid fiber runner systems (raceways), possibly with specific mounting brackets, to highlight their modularity and enterprise compliance Google 🟡 (Performance Scheme): Design is built around extreme performance and custom hardware. Their containment uses dense, bespoke raceways to safeguard signal integrity and optimize routing for their high-core-count fiber. Combine various elements: baskets, trays, and ladder racks for diverse cable types, with a prominent Subzero Frame - like contained rack to emphasize high-density, performance-focused infrastructure. The Takeaway: Their fiber containment systems are a physical map of their unique cloud missions. AWS prioritizes uptime, Azure prioritizes integration, and Google prioritizes speed. #DataCenter #CloudComputing #AWS #Azure #GCP #OpticalFiber #Networking

Microsoft has $80 billion in Azure orders it can't fulfill because it can't plug in the GPUs already sitting in its warehouses. Your enterprise team has the power, the rack space, and a signed PO. You're still six months back in line for trunk fiber. Same line. Let that sit for a second. On the Q2 earnings call, Satya Nadella admitted the bottleneck isn't silicon and it isn't capital, it's electricity to install hardware they already own. The most valuable company in the world is being held up by the most boring layer in the building. And power is just the headline. The same physical-layer squeeze is everywhere. Power transformer lead times have stretched to 128 weeks. At least one major fiber manufacturer has already sold its 2026 inventory. Multimode pricing jumped more than 70% between December and January. None of those numbers describe a chip. They describe everything that has to happen before the chip turns on. This is the part the capex headlines skip. A GB300 rack pulls 120kW and wants something north of 1,000 fibers terminated against a polarity scheme that doesn't tolerate field improvisation. The MPO trunks, the LC breakouts, the patch panels, that's the actual gating item. The GPU is just the part that gets the press release. And this doesn't just hit the hyperscalers. You're not building a 100MW greenfield campus in Virginia. You're retrofitting AI into a room specced for 8kW racks in 2017, with a three-person network team and a contractor crew that has never terminated a 144-fiber trunk in their lives. You have the watts. You have the floor. You have CFO approval. What you don't have is a vendor who treats your 60-rack refresh like it matters when there's a $6 billion PO ahead of you on the same production line. Somewhere between "we'll get to it" and "have you tried Q3 of next year?" The smart enterprise teams I'm watching right now stopped fighting for queue position and started picking partners whose production lines aren't consumed by hyperscale mega-orders in the first place. When your manufacturer isn't tooled for 50,000 racks, your 50 racks become the priority by default. Capex doesn't move you up the queue. Choosing a vendor that isn't in the queue does. The physical layer always wins. #AIInfrastructure #DataCenter #EnterpriseIT #SupplyChain #CloudComputing

🔴 Researchers from TSMC present the blueprint for next-generation optical engines at the #IEDM. The paper "EPIC-BOE An Electronic Photonic Chiplet Integration Technology with IC Processes for Broadband Optical Engine Applications" proves that leveraging TSMC 3DFabric for broadband optical engines will define the next decade of #CoPackagedOptics and #GenerativeAI. Future generative AI systems demand massive parallelism, high bandwidth density, and extreme energy efficiency. To meet these demands, this research team developed the first full integration technology for broadband optical engines, seamlessly connecting everything from the fiber directly to the CoWoS system. 1️⃣ High Fiber Count Integration: #VerticalCouplers & #3DFabric Unlike conventional edge couplers that suffer from severe beachfront warpage issues when scaling, this solution utilizes advanced vertical couplers. This completely bypasses the physical limitations of edge coupling, successfully integrating 40 to 80 fibers per row to achieve unprecedented I/O density. 2️⃣ Ultra Broadband Coverage: #SiliconNitride & #OpticalEngine The innovative process flow incorporates silicon nitride waveguides and polarization control devices. This enables a massive broadband coverage spanning from 1260 nm to 1360 nm, providing the immense optical bandwidth necessary for future data-hungry AI architectures. 3️⃣ System Level PPA Enhancement: #CPO & #TSMC By fabricating this compact co-packaged optics module utilizing standard IC processes, the architecture realizes significant power, performance, and area enhancements. It establishes a highly scalable and reliable manufacturing pathway for multiple row counts in extreme density AI clusters. 💡 My Take: As the physical footprint of optical I/O becomes a massive bottleneck for scaling AI accelerators, simply squeezing more edge couplers onto the die edge is no longer a viable strategy. Transitioning to an IC process-driven, vertical coupling architecture fundamentally rewrites the rules of Co-Packaged Optics. By directly integrating the photonic chiplets into advanced 2.5D and 3D packaging platforms like CoWoS, the industry can eliminate the warpage and real estate constraints of traditional fiber attachment. This is the exact hardware foundation required to seamlessly scale terabit-level optical interconnects for next-generation generative AI clusters. 👇 Link in the comments #AdvancedPackaging #SiliconPhotonics #OpticalInterconnects #3DIC #HardwareArchitecture #AIHardware #DataCenter #Optoelectronics NVIDIA AMD Broadcom Marvell Technology Intel ASE Group Amkor Technology, Inc. Applied Materials ASML Lam Research Lumentum Coherent Corp.

Hyperscalers aren’t just buying transport anymore — they’re quietly rewriting the optical playbook. AWS's decision to develop its own long-haul DWDM transponders goes beyond just a cost-saving strategy; it is also about sovereignty. When artificial intelligence drives a 73% increase in bandwidth while reducing power consumption by approximately 35% compared to vendor equipment, the concept of "owning the stack" shifts from being an ideological stance to a practical economic necessity. By controlling the photonics, software, and automation, AWS is transforming from merely being a customer of the optical ecosystem to becoming a network original equipment manufacturer (OEM) for itself. The shift from whitebox solutions in switching to the optical core signifies a significant change that should grab the attention of every telecommunications company and vendor. If hyperscalers take on more of the DWDM value chain, what is the future for traditional transport providers and carrier DCI strategies? Should they focus on “managed complexity” and SLAs, or begin to lose margins to in-house cloud designs? Here’s the uncomfortable question: If AWS can design its own optical destiny, what’s your plan to avoid becoming “just the dark fiber provider” in the AI era? Link to article:https://lnkd.in/gbGE9gDr #BellLabsConsulting

BANDWIDTH Most conversations about connectivity, especially AI connectivity start here. But FIBER ... who controls this asset is the more important question. Dark Fiber and Wavelength services both move massive amounts of data over fiber optic infrastructure. On the surface they can look similar. Under the hood, they solve very different problems. WAVELENGTH (lit service) gives you a dedicated optical channel on a carrier’s fiber network. > You get high capacity quickly, > the provider manages the optical equipment, > and deployment can be relatively fast. **For many enterprises, this is the simplest path to scaling bandwidth.** DARK FIBER, on the other hand, is raw fiber infrastructure. > You lease or obtain rights to the fiber itself and bring your own optical equipment to light it. > That means higher upfront investment BUT you also have full control over capacity, optics, and network design. This is why the choice often comes down to three things: • Control: Who manages the optical layer • Scale: How much bandwidth growth you expect • Economics: Short-term service vs long-term infrastructure strategy Wavelength services today commonly deliver 100G, 400G, and increasingly 800G per circuit. HOWEVER... a single dark fiber pair can ultimately support tens of terabits per second when you layer multiple wavelengths across the spectrum. That’s why #hyperscalers, #AI infrastructure providers, and large data center operators often look closely at dark fiber when planning long-term capacity. The right choice depends on your architecture, growth trajectory, and operational model. >> Sometimes simplicity wins. >> Sometimes control wins. Understanding the difference is the first step in making the right call. #Networking #FiberOptics #DarkFiber #Wavelength #AIInfrastructure #DataCenters #DCI #DCO Follow for more #AI_Infrastructure_Media

At a recent panel at MetroConnect, James Tindall (SVP & GM, Nokia) said something that deserves more attention: Hyperscalers want — and need — visibility and control. APIs are critical. He’s right. But this isn’t a technology preference. It’s a commercial and operational mandate. Hyperscalers operate as software-defined infrastructure companies. They expect infrastructure to behave like code: • Programmatic availability • Deterministic routing • Structured digital inventory • Real-time telemetry • API-driven provisioning • Lifecycle transparency If your fiber network cannot be queried, validated, provisioned, and monitored digitally — it becomes friction in their deployment model. And friction gets removed. But APIs alone are not the answer. APIs sitting on top of inconsistent location records or incomplete path data simply automate ambiguity. Here’s a real-world example: If a hyperscaler asks for four diverse paths into a data center, “diverse” is not a marketing term. It means physical route separation — different conduits, different laterals, different rights-of-way. If you cannot validate route provenance at the segment level, you cannot guarantee diversity. And if a backhoe hits both “diverse” paths because they shared the same conduit for 400 meters, the SLA conversation becomes very expensive. Visibility requires discipline. It requires Location Truth. It requires clean, normalized inventory down to the segment and path level. It requires knowing not just that a building is on-net, but how it is on-net — through which route, under which jurisdiction, with what latency profile, and with what operational exposure. The industry is shifting from asking: “Can you provide fiber?” To asking: “Can your network behave like software?” For AI workloads, sovereign compliance, and global data center expansion, infrastructure must be machine-trustable. This is not just an operational shift. It’s a competitive filter. The backbone of the digital economy isn’t just fiber. It’s trusted, programmable infrastructure. Join the movement to less friction, better experiences and growth.