AI Infra Is Nothing Like the "Classic Cloud Infra"

Complexity is overrated

SubscribeSign in

AI Infra Is Nothing Like the "Classic Cloud Infra"<br>Most builders will never need to mess with GPUs

Raman Sharma<br>May 27, 2026

Share

AI infrastructure, despite the billions flowing into it, is still in its infancy. But the signs are already clear. In its mature state, the AI infra market will look very different from how classic cloud turned out. Here are a few places where the trajectories diverge.

1. The Workload Foundation Is Different<br>Classic cloud was built for web workloads. You built software and deployed it. No concept of “training” existed. AI infrastructure is built for ML workloads, where you need massive compute just to prepare the model, long before anything is deployed.<br>The closest ML equivalent to “deploying a web app” is running inference. But the asymmetry matters. In the classic cloud world, the web was the primary way of shipping digital experiences, so a vast majority of developers (even those not tasked with infra operations) had to build fluency in cloud infrastructure.<br>That’s not true in the ML world. Most app builders (or “agent builders” as they are called today) don’t need to understand AI infra or GPUs at all. Unless you’re training, fine-tuning, or self-hosting models for inference (bastion of the elite), you can “incorporate AI in your apps” entirely through APIs (GPT, Claude, Gemini, 3rd-party-hosted open-weight models, etc). Which is what most people are doing.

2. Concentration Is Inverted<br>Classic cloud showed early concentration on the provider side. Top tier: AWS, Azure, GCP. A struggling mid-tier. Some smaller players like DigitalOcean, Linode, and Vultr. Overall, the number of companies that wanted to own/operate data centres at scale (to provide cloud services to other people) just wasn’t very high.<br>The consumer side was highly distributed: basically, every company that needed to run apps, websites, data pipelines, or anything that needed compute.<br>AI infra flips this. On the provider side, a wave of GPU-specialized neoclouds has emerged alongside the hyperscalers. The consumer side is far more concentrated.<br>Model training is the exclusive domain of a handful of well-capitalized labs.

Inference has a slightly broader customer base, but not by much. The dominant consumers are still the model builders, plus a new category of inference providers running open-weight models at scale.

In classic cloud, everyone was a customer, and three companies were the landlords. In AI infra, everyone wants to be a landlord, and there are about eight customers.

3. In Enterprise, PaaS Wins This Time<br>IaaS was the dominant model for classic. Primarily because large enterprises wanted control (or thought they did). This is why, despite the early popularity of some PaaS providers like Heroku, the industry largely gravitated towards IaaS. It also helped that a lot of cloud was built around infrastructure concepts that were well understood (like virtualization) even before cloud computing became popular. Several organizations had experience running infrastructure in their own data centers. They could translate that knowledge to cloud resources.<br>AI infra has no such legacy. Operating GPUs at scale was never something most enterprises did. For inference in particular, PaaS will be the dominant consumption model. This is especially true for enterprises, but also for the so-called “AI natives.”<br>The most popular closed models, GPT, Claude, and Gemini, are available only through APIs. That makes them structurally closer to Twilio or Stripe than to EC2 or EKS. Even the more sophisticated buyers, those using open-weight models to get 90% of frontier performance at 10% of the cost, are overwhelmingly relying on third-party inference providers rather than self-hosting. The competence (both hardware and software stack) required to run high-performance inference at scale is rare. Enterprises have little incentive to build it.

4. Customer Requirements Haven’t Matured Yet<br>As classic cloud matured, buyer requirements deepened. SLAs. Compliance certifications. Geo-redundancy. Vendor lock-in mitigation. The criteria got more sophisticated over time.<br>AI infra hasn’t reached that stage. The GPU supply crunch is so acute that even the most sophisticated buyers are still primarily focused on securing access. This is why relatively young companies are landing Fortune 500 logos. CoreWeave, at IPO, drew more than half its revenue from Microsoft.<br>The hottest sales motion in AI infra right now is: ‘We have GPUs.’ That’s it. That’s the pitch.<br>When customers are rationing a scarce resource, they don’t negotiate on terms. The requirements sophistication will come. It always does. But for now, availability is the product.

The Pattern<br>Classic cloud matured into a commoditized, IaaS-dominated market with a broad consumer base and consolidated providers. AI infra is on a different trajectory. Fragmented provider landscape....