--- title: "The next bottleneck for AI infrastructure: Data is starting to become immobile (excerpt from x)" type: "Topics" locale: "en" url: "https://longbridge.com/en/topics/41066920.md" description: "$Nokia Oyj(NOK.US) The money in AI infrastructure is gradually flowing from single-point computing power to system-level interconnection. The market has been focusing on GPUs for the past two years, which is not wrong. But one thing that is becoming increasingly clear now is that training large models is not about a few cards fighting alone, but about thousands of GPUs collaborating simultaneously. Machines are getting faster, but the overall efficiency of the system is increasingly constrained by data movement capabilities. For the next few years in the AI industry, the essence is solving the same problem: data movement. After PCIe 6.0..." datetime: "2026-05-23T06:41:56.000Z" locales: - [en](https://longbridge.com/en/topics/41066920.md) - [zh-CN](https://longbridge.com/zh-CN/topics/41066920.md) - [zh-HK](https://longbridge.com/zh-HK/topics/41066920.md) author: "[Ryan-](https://longbridge.com/en/profiles/19144604.md)" --- # The next bottleneck for AI infrastructure: Data is starting to become immobile (excerpt from x) $Nokia Oyj(NOK.US) The money for AI infrastructure is gradually shifting from single-point computing power to system-level interconnection. For the past two years, the market has been focused on GPUs, and that's not wrong. But it's becoming increasingly clear that training large models isn't about a few cards working alone; it's about thousands of GPUs collaborating simultaneously. Machines are getting faster, but the overall system efficiency is starting to be increasingly constrained by data movement capabilities. In essence, the AI industry in the coming years is solving the same problem: data movement. After PCIe 6.0, the pressure on copper interconnects regarding distance, power consumption, and signal integrity is becoming more apparent. AI infrastructure is gradually moving from copper-based to a hybrid of optical and copper, especially in long-distance, high-bandwidth scenarios, where optical interconnects will become increasingly important. The core contradiction of the past two years was insufficient computing power. Next, system bottlenecks will slowly spread to memory, networking, and data movement. This is also why NVIDIA, TSMC, telecom operators, and cloud providers are increasingly focusing on optical and networking. In the longer term, AI-RAN and edge AI will become important expansion directions for the next phase. Huang has actually made the future very clear. Huang has actually made the future growth path very clear: cloud AI factories are the present, enterprise and industrial AI are accelerating their deployment, and Physical AI—autonomous driving, robotics, edge AI, AI-RAN—is the next stage. But the problem is, when AI truly enters the real world, every robot, every autonomous vehicle, and every smart grid node will continuously generate massive amounts of real-time data. Many scenarios simply don't allow data to be sent thousands of kilometers to the cloud and wait for results to return because latency itself determines whether the system can function normally. This means the focus of AI infrastructure will gradually expand from "single-point computing power" to backbone optical networks, edge low-latency communication, AI-RAN, data center interconnection, and carrier-grade infrastructure. What will truly matter in the future is not just the GPU itself, but who is responsible for connecting these AI systems. Why Nokia? If you sort out the competitive landscape, the answer isn't actually complicated. There are only a handful of players globally that can truly build large-scale, carrier-grade communication infrastructure. Huawei's technical strength is certainly strong, but geopolitical constraints are very evident within the European and American AI infrastructure systems. Among the remaining Western players, many companies are either more focused on single-point optical modules or more on enterprise networks. There aren't many that truly cover backbone optical networks, IP core routing, carrier infrastructure, and localized supply chain capabilities simultaneously. And Nokia happens to be at the intersection of these segments. NVIDIA's deepening collaboration with Nokia in the direction of AI-RAN and carrier AI networks recently is no coincidence. The truly critical part is the Infinera acquisition. It filled Nokia's most important past weakness: long-distance optical interconnect capabilities and InP (Indium Phosphide) optical chip manufacturing lines. In future AI optical communications, many high-speed optical chips will require InP materials, and there aren't many companies in Europe and America that truly possess scalable InP manufacturing capabilities. After the acquisition, Nokia began moving from a traditional equipment vendor towards an integrated chip + system + backbone network direction. The Infinera acquisition filled in the most critical piece of the puzzle. Nokia was more of an equipment integrator in the past, relying on external suppliers for core optical chips. Price hikes from others hurt, and shortages from others caused supply disruptions. Infinera has an InP optical chip manufacturing plant in San Jose, California. Many high-speed optical chips in future AI optical communications will require InP materials, and there aren't many companies in the European and American markets that truly possess scalable InP production line capabilities. After the acquisition, Nokia began moving towards a full-chain "chip plus system plus backbone network" approach. Profits started to stay in its own hands. The financial figures confirm this change. Overall revenue for the latest quarter grew 4% year-over-year, but the optical networking business grew 20%, net profit surged 245%, and gross margin improved significantly. New factories will gradually start production this year, and new chip solutions can help customers reduce usage costs by 40%. The explosive period might not have truly begun yet, but the trigger conditions are almost ripe. Two changes will gradually shift the market's attention from inside the data center to the entire network. The first change is that computing power is beginning to cross geographical boundaries. A single data center has limits on land, power grid, and cooling. When computing power in one location is insufficient, cloud providers will inevitably connect data centers across multiple cities into a whole. Once data starts moving out of the data center, short-distance interconnection is no longer the only bottleneck; the importance of long-distance backbone optical networks, DCI, and core routing systems will significantly increase. This is NOK's true home turf. The second change is that AI is moving from training to inference, and from the cloud to the edge. Scenarios like autonomous driving, industrial robots, and smart grids have extremely high latency requirements. In many cases, data simply cannot be transmitted thousands of kilometers to the cloud and back for processing. Operators must deploy low-latency edge networks closest to the devices. This will bring a large-scale carrier expansion cycle not seen in the past decade, and Nokia itself is one of the important players in the global carrier backbone network. On the timeline, I think the second half of 2026 to 2027 will see the initial manifestation of demand for industrial private networks, smart grid communications, AI-RAN, etc. The truly large-scale backbone network upgrade cycle might have to wait until late 2027 to 2028, only fully exploding after Vera Rubin and large-scale Agent traffic truly land. Wall Street typically starts pricing things in 12 to 18 months in advance. The biggest difference between NOK and CIEN Both are in the optical communications business, but their directions are different. CIEN is technology leadership-driven: betting on the continued leadership of the WaveLogic series, being first with 1.6T transmission solutions, and reducing power per bit by 50%. Technologically very strong, but squeezed between cloud providers and upstream suppliers, with full order books but not necessarily retaining all the profits. NOK is industry chain integration-driven: betting on the entire AI backbone network upgrade cycle. Not just doing optical networking, but also IP routing, fixed networks, carrier infrastructure, providing one-stop solutions for customer needs. After integrating chips and equipment across the entire chain, the moat will deepen. There's another often-overlooked source of income. NOK doesn't just make money from equipment; it also has an extremely strong patent cash flow. A large number of 4G/5G equipment, local networks, and vehicle networking solutions globally are essentially paying Nokia toll fees. Regardless of industry fluctuations, this revenue stream won't disappear, has extremely high profit margins, and provides a stable cash flow base. Risks also need to be made clear. 1. Traditional telecom operator business capital expenditures have been contracting in recent years, and this drag is real. The high growth of the AI business needs to continuously offset the pressure from the traditional business, and the process is not linear. 2. Having many orders doesn't mean they can smoothly turn into revenue; supply chain pressure exists across the entire industry. 3. The market currently still labels Nokia as a traditional communications company. Re-pricing this label requires time and sustained financial data to drive. Summary When AI starts collaborating across data centers, cities, and edge nodes, the real beneficiaries are long-distance backbone optical networks, AI-RAN, carrier-grade infrastructure, and edge low-latency networks. $Nokia Oyj(NOK.US) is the underlying pipeline for all of this. Backbone networks are its main business. AI-RAN is the direction it's jointly advancing with NVIDIA. Carrier infrastructure is its decades-accumulated customer relationships. InP chip production lines are its unique manufacturing barrier in the European and American markets. Patent income is the toll fee it continuously receives regardless of industry fluctuations. When AI truly steps out of the data center, Nokia is the one collecting the toll on the road. ### Related Stocks - [NOK.US](https://longbridge.com/en/quote/NOK.US.md) - [NOA3.DE](https://longbridge.com/en/quote/NOA3.DE.md) - [INFN.US](https://longbridge.com/en/quote/INFN.US.md)