Agentic AI shows few signs of slowing, and some analysts expect the wave to generate as much as a trillion dollars of capital spending over the next several years. That forecast comes from Alan Weckel, co-founder and analyst at 650 Group, whose bullish outlook anchored a recent industry roundtable on the rise of the “neocloud” — a new class of vendors selling GPUs as a service, exemplified by fast-growing independents such as CoreWeave, Nebius, and Lambda.
Weckel’s forecast implies commensurate growth in networking for AI data centers, which neocloud providers and the major generative-AI vendors are building at a rapid pace worldwide — stoking talk of a bubble in the process. In his telling, the build-out means hundreds of boxes entering data centers every minute, racks deployed continuously, and terabits of connectivity required between the buildings being linked.
Why AI data centers must be networked
New AI facilities increasingly need to be connected to one another because the largest AI workloads have outgrown single buildings. Training frontier foundation models — now measured in trillions of parameters — requires more power than any one site can supply, so operators spread GPUs across multiple facilities and stitch them together. Vendors are also building dedicated, separate plants for training and for inference, and inference has its own data-gravity problem: the datasets a model needs at prediction time are not always local, which pulls still more traffic onto inter-datacenter links.
The rise of the “neoscaler”
The roundtable introduced the term “neoscaler” for the emerging group of cloud vendors that rival traditional hyperscalers in scale but operate in a more decentralized fashion. Their strategy is to network together sometimes far-flung GPU clusters and offer them on demand — to enterprises, to generative-AI vendors, and even to hyperscalers running AI workloads.
Mark Bieberich, vice-president of portfolio marketing at optical-networking firm Ciena, described the model as metered GPU compute: pay-as-you-grow businesses that are well capitalized and expanding both compute and the network infrastructure beneath it. Ciena, which specializes in spectrally efficient optical systems, says it works with roughly two dozen neoscalers that have already decided to build their own optical network infrastructure, and sees what Bieberich called fairly sustainable demand over the next five to ten years.
Sovereign AI adds another wave
Beyond the commercial build-out, Weckel pointed to sovereign AI as a further driver of data-center construction spending. Nations increasingly treat AI infrastructure as a matter of data sovereignty: for regulatory and national-security reasons, many governments will find it difficult to host AI workloads outside the country of origin, which multiplies the number of facilities that must be built and interconnected.
Powering it all
The power demands are reshaping site selection. Operators are courting new municipalities and new energy sources as existing grids strain. Weckel cited Meta as emblematic: the company operates a roughly one-gigawatt data-center campus in Ohio and is planning an approximately five-gigawatt facility in Louisiana. Microsoft, meanwhile, is building a multi-gigawatt center in Wisconsin that it plans to link with a companion site in Atlanta. Operators are looking increasingly at nuclear and hydroelectric generation, with solar and wind serving mainly as backup in hybrid configurations. Geography drives the details: in the water-scarce but capital-rich Middle East, operators are spending heavily on desalination to produce cooling water, while in water-rich regions such as Canada and Northern Europe, hydropower is becoming a primary energy source.
Limitations and what to watch
Much of this outlook rests on analyst projections rather than realized demand, and trillion-dollar capex forecasts assume that agentic AI adoption — and the revenue to justify the infrastructure — materializes on schedule. The bubble question is live: several neocloud business models depend on sustained GPU scarcity and long-term contracts with a small number of AI labs, a concentration risk worth monitoring. Power and water constraints could also slow builds regardless of demand. Useful signals to watch include whether announced multi-gigawatt campuses reach completion on their stated timelines, how quickly optical-networking orders from neoscalers grow, and whether sovereign-AI mandates translate into funded projects. Background on the vendors named is available from 650 Group and Ciena.
Related reading on this site: emerging trends in data engineering for 2026 and x402 and payments for AI agents.
