32
AI inference is reshaping network demand, creating an opportunity for service providers to turn high-capacity, intelligent connectivity into new revenue as enterprises and hyperscalers scale AI workloads.
AI has taken the world by storm and gone from a niche technology to a boardroom imperative. According to a McKinsey survey, nearly nine out of ten organizations are now using AI, and many have begun to use AI agents, making it one of the fastest-adopted enterprise technologies in history.
But while much of the AI conversation has focused on GPUs, chips, and data centers, businesses are quickly realizing that without the underlying communication network, those investments risk becoming little more than desert islands. Without high-speed, scalable connectivity, billions of dollars in AI infrastructure spending will not deliver their full value.
AI inferencing and network demand grow together
In AI’s first chapter, the industry was mostly focused on developing new models. To enable the training of larger, more intelligent models, hyperscalers and AI labs invested in massive GPU clusters. These efforts remain relevant to this day as more capable and advanced frontier models are pursued by different companies aiming for Artificial General Intelligence (AGI), which has been referred to as the type of AI that can exceed human cognitive capabilities. However, an AI model only creates value—and revenue—when it’s put to use. It’s during inference that models provide answers and valuable insights that can be monetized. Training is about building capabilities; inference is about usage and turning these capabilities into services.
Every time a model is used, when it answers a question, flags a fraud pattern, or helps an engineer debug code in real time, that’s inference. As AI adoption grows with more users and agents and as AI is infused into more digital services and experiences, inference demand grows exponentially. This year is on pace to be the first year that global spending on inference will overtake spending on training, a shift that reflects how quickly AI has moved from niche use cases into something pervasive in everyday life. It will help generate revenue as businesses embed AI into products and services, turning AI from experimentation into an ongoing source of commercial value.
Simultaneously, another part of the AI ecosystem is seeing substantial monetization growth – the underlying network infrastructure that carries inferencing traffic. Ciena’s recent Wave Services report projects the market for high-capacity “wave” network services – the optical connections that move AI traffic between data centers, clouds, and continents – will grow at an 18 percent annual rate and pass $50 billion by 2030. For context, the world’s largest telecom and network operators have grown overall revenue by less than 1 percent a year, on average, over the past twelve years according to McKinsey.
The growth of AI inference is translating directly into new revenue for telecom and hyperscale network providers, rather than simply more traffic to carry at flat or falling margins.
What (or how) are we inferring?
Training an AI model is manageable, predictable. It is much easier to predict how much capacity is needed and for what time period. AI Inference brings a different set of network requirements.
Inference is distributed in nature, quickly shifting, and hard to predict. It is triggered by billions of human requests, and by AI agents operating continuously in the background, monitoring conditions and actioning tasks autonomously. As agents continue to multiply, AI-enhanced IoT will create new flows, and other AI applications will optimize network usage running scheduled tasks in off-peak times.
Effectively, the AI era – and particularly inferencing – can potentially render the network peak hours obsolete, with traffic becoming a constant-yet unpredictable load that can shift across cloud providers and regions by the minute, if not the second, and can be so data heavy, with a single prompt now potentially carrying text, images and video.
Today’s networks weren’t built for this level of demand. They were designed for steadier traffic patterns, more centralized applications, and a world consisting of more manual processes to reroute traffic or add capacity. Add multi-cloud AI deployments, where a single application might span private infrastructure, several public clouds, and edge locations, and the limits of that architecture become obvious. This is another beast entirely, and simply adding bandwidth won’t solve the challenge.
The AI era requires network connectivity that behaves more like cloud infrastructure. Future networks will need on-demand optical bandwidth allocation so that service providers can give customers the ability to scale optical capacity up or down to meet changing requirements. This will allow enterprises to benefit from the flexibility to scale their network connectivity alongside AI workloads and pay for the capacity they need when they need it.
For service providers, this creates a new revenue opportunity. Instead of selling connectivity primarily as rigid capacity, they can introduce differentiated, consumption-based services aligned with how customers increasingly consume cloud and AI infrastructure.
Inference requires intelligence
This is where the opportunity rests for service providers to capture a part of the overwhelming investment in AI – on the infrastructure side. Recent Ciena research reveals that 90 percent of service providers expect high-capacity AI-driven network services – connectivity to AI infrastructure required by enterprises, hyperscalers, and neoscalers – to be a primary revenue driver over the next three to five years, with 56% citing them as their primary source of net-new revenue.
But for AI to be successful, a network that no longer solely acts as a conduit shipping data from A to B or C is key. Findings from that same survey reveals the majority (88%) of respondents had a strong sense of urgency on the need for optical network upgrades to support the demands of premium enterprise AI service level agreements.
To capitalize on AI, service providers need high-capacity, robust optical networks powered by advanced software that can steer traffic intelligently and dynamically.
In effect, the age of AI inference requires an intelligent and adaptive architecture – ideally one that enables granular control over how traffic is routed so that different AI workloads can be prioritized for latency, bandwidth and resilience based on requirements at the time.
By giving service providers real-time visibility into their networks potential disruptions can be predicted and mitigated before they impact users. To achieve greater visibility and control, service providers are adopting coherent routing platforms that integrate ultra-scalable coherent optical transport directly into the IP layer. They’re also utilizing intelligent software that automates and provides telemetry to manage traffic, particularly the kind that skews from traditional patterns. Other advances like hyper-rail photonics and full-spectrum transponders, will also play a key role to support the growing network demands driven by AI.
As AI adoptions accelerates and inference becomes the dominant AI workload, service providers are making investments to handle the growing AI-driven demands. Providers that can deliver the capacity, flexibility and intelligence needed to move AI traffic reliably and efficiently, will be best placed to capture new revenue opportunities.
Brodie Gage is Ciena’s chief product and technology officer, leading Product Line Management, Systems Engineering and Introduction, and Quality, while working with R&D to drive the company’s technology direction. An electrical engineering graduate of McGill University, Gage joined Ciena in systems engineering before moving through strategic product marketing and product line management, ultimately helping shape the company’s optical portfolio and joining its executive leadership team in 2023.
