For the previous few years, the AI infrastructure race has been pushed by the belief that bigger artificial intelligence fashions require bigger information facilities. That concept has fueled a unprecedented wave of spending.
In rural Richland Parish, Louisiana, for instance, Meta is constructing one of many world’s largest AI infrastructure initiatives. Its Hyperion campus is predicted to value more than $50 billion and run on about 5 gigawatts of energy, roughly the output of 5 nuclear reactors.
However as these services develop, so does the resistance to them. Throughout the nation, communities are pushing back over information facilities’ calls for on energy and water, and the impacts they’re having on rural and suburban areas. The larger the challenge, the extra doubtless it’s to change into a political goal.
However Tom Leighton, cofounder and CEO of Akamai, the $18 billion content material supply community firm that powers a big share of the world’s net site visitors, would argue that’s not even the worst half. The Meta challenge, like so many others, rests on the belief that firms capable of focus probably the most computing energy shall be finest positioned to construct the subsequent technology of AI methods. And Leighton believes that assumption applies extra clearly to AI coaching than to AI inference (the “considering” that AI does when making use of its coaching to real-world information).
“The following problem for AI is what it would take to run these fashions and their derivatives all over the place,” he tells Quick Firm, sharing his perception that the business could also be making an attempt to unravel too many issues with the identical monumental constructing.
Leighton’s 28-year-old firm has spent the previous a number of years increasing right into a cloud platform for AI. Somewhat than making an attempt to match the hyperscalers information middle for information middle, Akamai is betting on a distinct strategy provide a much less disruptive path for increasing AI infrastructure, one which depends extra closely on a community of current services as an alternative of concentrating monumental calls for for land and energy in a single neighborhood.
“Agentic AI wants low latency, excessive efficiency, and inexpensive economics that big, centralized information facilities can’t present,” Leighton says.
His proposed answer can also be impressed by what made Akamai a participant in web infrastructure to start with in the course of the dot-com period. Within the late Nineties, Leighton helped handle the online’s rising pains, when each request going again to a handful of centralized servers was choking the web’s progress. Now, Akamai is making an attempt to deliver a model of the structure that saved the online to AI.
And Leighton’s received the AI kingmaker Nvidia backing him.
Greater AI information facilities gained’t remedy each AI downside
Leighton’s proposed answer is partly an financial one. Working an AI mannequin is a distinct downside from coaching one, and lots of inference duties don’t require the most important mannequin, probably the most highly effective chips, or a request touring 1000’s of miles to an enormous centralized campus.
For enterprises pursuing agentic AI, the controversy extends past securing graphics processing unit (GPU) capability. Corporations more and more must resolve the place inference ought to run, the place enterprise information is processed, how rapidly AI brokers should reply, what it prices to maneuver information throughout areas, and the way safety insurance policies are enforced as soon as these methods transfer into manufacturing. In different phrases, deploying AI brokers more and more turns into an infrastructure design downside—not only a mannequin choice downside.
Leighton argues these operational selections, not uncooked compute alone, will decide whether or not AI purposes ship acceptable efficiency, reliability and economics at scale.
Inference efficiency relies upon partly on how rapidly an AI system can reply when a choice is required. Community distance, information location, device calls, and utility design can all have an effect on response time.
“If an AI request has to journey 1000’s of miles, contact information in one other location, name instruments, after which return a solution to a consumer in actual time, the velocity of the information middle is just one a part of the equation,” Leighton says. “Proximity issues. In actual fact, it might matter much more as AI evolves.”
Leighton’s view attracts on Akamai’s expertise with net infrastructure in the course of the late Nineties.
As the online expanded, centralized origin servers struggled to deal with world demand. “Web sites had been constructed round centralized origin servers. As demand went world, each consumer request needed to journey again to a small variety of central locations,” Leighton recollects. “The end result was gradual efficiency, web sites would go down or freeze up throughout site visitors spikes, and there was a whole lot of frustration for customers.”
The business known as the issue the “World Vast Wait.” Leighton, an MIT utilized mathematician, helped develop another based mostly on distributing content material and computation nearer to customers. Algorithms decided the place requests needs to be served.
Almost three a long time later, Leighton believes AI infrastructure now faces a associated distribution downside, and argues {that a} main engineering problem shall be coordinating inference throughout 1000’s of knowledge middle places whereas sustaining constant efficiency, safety, and reliability as a unified system.
“Anybody can construct an information middle, however making many places act intelligently collectively, below altering demand, with constant efficiency and belief, is one other problem altogether,” he says.
Can distributed AI outperform centralized clouds?
The central part of Akamai’s technique is AI Grid, an orchestration layer developed with Nvidia. AI Grid determines whether or not an inference workload ought to run in a centralized AI manufacturing unit, a regional cloud, or one in all Akamai’s edge places. The choice relies on latency necessities, working prices, and efficiency wants.
Leighton argues that routing selections can materially have an effect on inference efficiency. GPU value and availability stay important constraints, however the business is asking step-one questions in a step-two market.
“The larger questions are round the place inference ought to run, what information it wants, how rapidly the response should come again, what it would value to maneuver the information, and what safety coverage must be enforced alongside the best way,” he says. “The solutions to these questions can have a a lot larger affect on whether or not AI reaches its potential.”
For Leighton, the definition of scale itself is altering. In the course of the coaching period, scale meant concentrating as a lot compute as doable inside a single AI manufacturing unit. Within the agent period, he argues, scale more and more relies on how successfully infrastructure can distribute inference throughout many places whereas retaining latency, information motion, and prices below management.
“GPU availability is a significant difficulty at the moment, however GPUs will not be the answer to each AI downside,” he argues. “Many inference duties don’t want the most important mannequin or the most important cluster or the costliest compute. What they want is the power to marry the proper mannequin, in the proper place, information and second, at the best value.”
Akamai says buyer demand is starting to replicate this strategy. Earlier this yr, the corporate disclosed a four-year, $200 million settlement with an unnamed main U.S. know-how firm to deploy one of many world’s largest clusters of Nvidia RTX PRO 6000 Blackwell GPUs on its platform.
Three months later, Akamai introduced a seven-year, $1.8 billion cloud infrastructure commitment from a number one U.S. frontier AI mannequin developer. Insider reviews recognized the corporate as Anthropic. The settlement is the most important contract in Akamai’s 28-year historical past.
Collectively, the 2 agreements characterize roughly $2 billion in dedicated cloud enterprise from clients Akamai didn’t have two years in the past. Akamai’s cloud infrastructure income has additionally grown 40% yr over yr.
Leighton declined to establish the businesses or talk about the workloads behind the agreements. He says buyer evaluations now embody a broader set of operational questions. “When clients consider AI infrastructure, in fact they have a look at scale and efficiency,” he says. “However they’re additionally asking how workloads carry out in manufacturing, how prices evolve over time, how dependable the infrastructure is, and whether or not it offers them the flexibleness to adapt as AI utilization modifications.”
With out naming further clients, Leighton says Akamai is supporting manufacturing AI deployments for an AI-powered video intelligence platform in India and a U.S.-based client AI firm. Each require low-latency inference and don’t rely on giant centralized coaching clusters.
The actual value of AI goes far past GPUs
Many firms consider AI infrastructure by token costs, GPU utilization, and mannequin endpoint prices. Manufacturing methods additionally create prices associated to context retrieval, API calls, storage reads, and community site visitors throughout zones and areas.
Akamai claims its structure can cut back inference latency by as a lot as 2.5 occasions in contrast with conventional hyperscaler infrastructure, and decrease inference prices by as much as 86%. Its revealed benchmarks, carried out utilizing Nvidia’s methodology, present RTX PRO 6000 Blackwell GPUs on Akamai’s cloud delivering as much as 1.63 occasions the inference throughput of Nvidia H100 GPUs. The system sustained roughly 24,000 tokens per second per server below 100 concurrent requests.
Leighton says the business viability of AI infrastructure will rely on whether or not these methods can ship acceptable efficiency, reliability, and price.
“As a CEO, I’m at all times skeptical of obscure economics,” he says. “It’s straightforward to say the longer term is larger information facilities and extra GPUs. It’s more durable to indicate how that structure delivers the proper efficiency, reliability, and price when AI is working all over the place, on a regular basis. Inference is the place AI should present worthwhile ROI.”
Leighton’s background consists of work as a theoretical pc scientist. He holds greater than 50 patents and has served as Akamai’s CEO for greater than twenty years.
“As a mathematician, I’m skeptical of straight-line considering. The truth that one structure labored for the primary part of AI doesn’t imply it would work for each part that follows,” he says. “As a theoretical pc scientist, I’m conscious that what ignites a technological revolution isn’t the factor that lets it survive in the actual world. The early promise of AI isn’t any exception. The huge centralized clouds that began this growth aren’t constructed for the extremely distributed actuality of what comes subsequent.”
Wall Avenue continues to be pricing AI just like the cloud period
Wall Avenue has spent a lot of 2026 evaluating how Akamai’s cloud enterprise matches with its legacy operations. The corporate’s content material supply community enterprise has been shrinking for years. Its AI cloud enlargement has additionally required substantial spending on {hardware} and infrastructure.
Rising reminiscence costs, pushed by robust demand for AI {hardware}, elevated part prices as Akamai bought 1000’s of Blackwell GPUs. Margins compressed, and earnings per share declined at the same time as income grew. Traders continued to worth the corporate largely as a mature infrastructure supplier.
The $1.8 billion dedication led to Akamai’s largest single-day inventory rally in additional than twenty years. The settlement supplies proof of demand for Akamai’s distributed AI infrastructure. It additionally underscores the quantity of capital required to construct and function that infrastructure at scale.
“It highlights each the chance and the self-discipline required to pursue it,” Leighton says. “We’re making important investments as a result of we imagine AI shall be a significant driver of cloud demand, however we’re not making an attempt to easily copy the hyperscaler mannequin. Our benefit is that we already function one of many world’s most distributed platforms, and energy and defend giant elements of the web. The investments we’re making are about extending that platform for cloud and AI.”
The AI business has invested closely in GPUs, giant clusters, and extra highly effective fashions. Leighton argues that inference would require further infrastructure designed round distribution, latency, and price. “Within the early days of the online, many individuals assumed the reply was simply extra central infrastructure. It was not. AI is reaching the same level,” he says.
Whether or not Akamai’s distributed mannequin can compete successfully with centralized cloud suppliers will rely on buyer demand, technical efficiency, and the economics of working the community at scale.
