Some expect a wave of agentization sweeping the AI industry, which is showing few signs of abating, to generate as much as a trillion dollars of capital spending over the next few years.
The forecast for agentic AI is “absolutely an insane number,” said 650 Group co-founder and analyst Alan Weckel, whose bullish outlook on the financial prospects of the neocloud industry recently punctuated a roundtable discussion on the forces behind the emergence of a new class of vendors selling GPUs as a service, exemplified by rapidly growing independent ones. Corewave, Nebius and Lambda,
Waeckel’s vision anticipates equally great growth in networking for AI data centers neocloud provider And major generic AI vendors are building at a rapid pace around the world – and creating fear in the process hey bubble,
“That means hundreds of boxes are going into the data center every minute. That means racks are being deployed every minute, and we’re talking about terabits and terabits of connectivity between the buildings that need to be connected,” Weckel continued.
linking data centers
New AI data centers need to be connected or networked, because with the increasing size of AI workloads, companies can no longer train huge foundation models, now measured in trillions of parameters, in a single facility.
“You need to find additional power sources, and so you’re spreading those GPUs across multiple facilities,” Weckel said.
And again, AI vendors are finding that they need to build dedicated, separate plants for it. training and inferenceHe added, “AI doesn’t live in a bubble, As you start predicting things, it becomes very important to bring in those data sets,” “And the data you need for estimation won’t always be local,”
after several AI data centers are connected to the network Supporters of the AI boom see another wave of investment, Weckel said, to meet the growing demand for more compute power to process large-scale generative and agentic AI workloads. it is physical aiA world of AI-powered robots and autonomous transportation.
“The key message is that the multi-trillion dollar spending we will see at the end of the decade will carry us into the next decade,” Wackel added.
seller’s perspective
The online event was sponsored by Ciena, an optical networking vendorCiena AI uses the term neoscaler to identify the new group of cloud vendors, which are big like the traditional hyperscaler cloud giants, but in a more decentralized fashion, Their neoscalers’ strategy is to network together sometimes far-flung GPU clusters, which they make available on demand to enterprise clients, generative AI vendors, and even hyperscalers running AI workloads,
“With the advent of more and more AI workloads and greater demand for AI traffic, these companies need to scale to drive their business,” Mark Biberich, Ciena’s vice president of portfolio marketing, said during the Dec. 4 event. “They need to expand rapidly, not only in terms of their compute infrastructure, but also in terms of the network infrastructure that supports it.”
“Their business really runs on a ‘pay-as-you-grow’ or on-demand mode… as I call it, metered GPU compute,” he said. “These are companies that have very good capital and are building network infrastructure.”
focus on sienna Optical networking with spectral efficiency systemsAnd Biberich said the company is involved with a group of neoscalers, which numbers about two dozen and “they have already decided that they need to build out their own optical network infrastructure to support their businesses.”
Biebrich said Ciena sees “fairly sustainable” demand for its services over the next five to 10 years. “The market really has tremendous growth potential,” he said.
Meanwhile, AI is another area driving huge growth in spending on data center construction sovereign aiAccording to Wakel.
“The enterprise part of the country is very important because of data sovereignty. It will be very difficult for regulatory reasons as well as national security reasons to put AI anywhere outside the country of origin,” he said.
running centers
To power AI data centers, their manufacturers are exploring different ways to generate the huge amounts of power required. This places huge demands on existing power grids. “What’s happening is operators are looking to other municipalities, they’re looking to other power sources,” Weckel said.
Weckel cited META as an example of the growing demand for power by larger AI vendors and users. The social media giant maintains a one gigawatt data center in Ohio. Now Metra is planning its next data center, a 5-gigawatt plant in Louisiana. At the same time, Microsoft is building a multi-gigawatt center in Wisconsin, which the tech giant plans to connect to another center in Atlanta.
“So, we’re looking at much larger energy sources around nuclear and hydroelectric,” Wackel said. “You start putting more and more pins on the map. It requires a lot of new connections that never existed before.”
Data center operators also are using solar and wind power, but only as backup, in hybrid setups, Weckel said. In places like the Middle East, where water is scarce but capital is abundant, data centers are searching for water and spending heavily on desalination plants to produce fresh water to cool the multitude of GPUs within the plants. In water-rich regions such as Canada and Northern Europe, hydropower is becoming a primary energy source.
