For the past several years, the AI industry has been focused on a relatively straightforward question. Could the industry build enough compute fast enough to keep up with demand?

The rise of large language models (LLMs) triggered an unprecedented race for infrastructure in which we are still very much in the middle. Nvidia sits at the center of the AI ecosystem, while hyperscalers are expanding datacenters at a pace never seen before. In addition, AI-native cloud providers have emerged to serve organizations unable to secure capacity through traditional channels. Cranes are still going up and power purchase agreements are still being signed.

The industry’s attention has naturally followed these developments and the discussions surrounding them centered on accelerators, networking fabrics, power consumption, cooling systems and datacenter construction. Investors track GPU shipments; enterprises compete for available capacity and cloud providers jockey for access to increasingly scarce infrastructure. The focus is understandable because AI demand has been growing faster than the industry’s ability to satisfy it. That demand is expected to continue accelerating. The Tirias Research white paper Forecasting the Rise of Agentic AI & Interaction Models projects annual LLM inference growing from 990 trillion tokens in 2024 to more than one quintillion (1,008,410 trillion) tokens by 2030, alongside dramatic increases in image and video generation.

Yet as AI deployments mature, the industry is beginning to encounter a different kind of bottleneck. As AI utilization continues to evolve from chat to agentic workloads, the challenge is shifting from obtaining AI infrastructure to putting that infrastructure to work effectively.

At the crux of this challenge is that while running a model is largely a compute problem, running an agent is a systems problem. Agents change both the way infrastructure is used and the amount that is consumed. The Tirias Research forecast describes this evolution as a transition from Wave 1, where AI is used primarily as a conversational assistant that responds to individual prompts, to Wave 2, where autonomous agents reason, invoke tools, maintain context and execute complex, multi-step tasks with minimal human intervention. These agents often operate continuously rather than ending after a single response. As a result, the forecast estimates that the average Wave 2 user consumes 40 times as many tokens as a traditional Wave 1 chat user, while technical users operating asynchronous agents consume orders of magnitude more. As workloads become longer-running, more autonomous and increasingly interconnected, the challenge shifts from simply providing compute to orchestrating complex systems around it.

Recently, Nvidia CEO Jensen Huang described the industry as reaching another “inflection point” driven by agentic AI. An inflection point is not simply growth. It is the point where the trajectory changes and in this case, accelerates. The underlying technology remains important, but the center of gravity shifts. Agents must plan, access tools, interact with external systems, maintain memory and coordinate actions across multiple services. The infrastructure remains essential, but it is no longer sufficient. Success increasingly depends on the systems that connect models, tools, data and actions into reliable workflows.

As Tirias Research Sr. Analyst Kevin Hein distills it, “The question is not whether AI infrastructure remains important, but what becomes important next?”

Infrastructure Is Becoming the Starting Point

The first phase of modern AI was about proving what the technology could do. The second has been about building the capacity to do it at scale. That means training clusters, inference capacity, storage systems and high-performance networking. The emergence of AI-native cloud providers was a direct response to that demand.

Viewed this way, an AI cloud is evolving beyond infrastructure optimized for AI workloads. It is becoming a platform that combines compute, models, data, tools, orchestration and operational services into an environment for building and running AI systems.

But infrastructure is increasingly looking like the beginning of the journey rather than the destination. A production AI deployment today typically involves much more than a model and a GPU. Organizations combine retrieval systems, vector databases, observability platforms, evaluation frameworks, agent architectures, security controls, orchestration layers and third-party integrations. The resulting hardware and software stack can span multiple vendors and multiple clouds.

The challenge is even more pronounced in areas like Physical AI. Building autonomous systems also requires simulation, synthetic data generation, model training, validation, deployment, telemetry collection and continuous retraining. The systems surrounding the model can easily dwarf the model itself.

Obtaining infrastructure remains necessary. But knowing how to assemble everything around it is becoming the harder problem.

This is not be the first time the technology industry has encountered such a transition. The history of computing can be viewed as a history of abstraction. Operating systems reduced the need to understand hardware. Virtualization removed the burden of managing physical servers. Cloud computing eliminated the need to build datacenters. Serverless computing took infrastructure deployment off the table entirely. Each step moved users further from the underlying technology and closer to the outcomes they sought.

AI appears to be following a very similar path. Most organizations still approach AI from the bottom up. They start by selecting models, provisioning infrastructure, configuring frameworks, connecting APIs and assembling workflows piece by piece. It is a lot of work before anything useful gets built. However, what customers increasingly seem to want is something different. They want to start with the problem they are trying to solve and have the platform handle the rest.

The Emergence of Workflow-Centric AI Cloud

This is where the current generation of AI cloud providers becomes interesting. Most still describe themselves as infrastructure companies. Nebius, for example, explicitly positions itself as an AI cloud provider. But looking across its customer stories, product announcements and conference themes reveals something more nuanced.

The recurring theme is not infrastructure capacity but improving the systems surrounding the model. Customer stories focus on orchestration, retrieval, observability, evaluation and deployment rather than simply model performance. The emphasis consistently falls on reducing the complexity of building reliable AI systems rather than exposing additional infrastructure components.

Nebius’ product development remains firmly rooted in its technology stack, but the company increasingly packages those capabilities around customer workflows. This is visible in its support for open models, third-party tools, ecosystem partnerships and higher-level platform services. The pattern suggests a company focused not only on providing infrastructure, but on making that infrastructure easier to apply to the real-world AI development and deployment challenges.

Historically, technology vendors have built products and expected customers to adapt their workflows accordingly. The emerging AI cloud model is reversing that relationship, starting with customer workflows and evolving the platform to support them. Essentially, customer workflows are becoming the product requirements.

When Workflows Become Products

The emergence of workflow-oriented platforms provides an interesting illustration of this trend. Historically, cloud platforms exposed infrastructure components. Customers selected virtual machines, databases, storage systems, networking services, queues and observability tools, then assembled those pieces into applications. The platform provided building blocks. The customer supplied the architecture.

Increasingly, AI platforms are moving one level higher. Rather than exposing individual services, they are beginning to expose workflows.

A useful example of this shift is provided again by Nebius with its recently introduced Agents Blueprint. Infrastructure-as-Code tools such as Terraform made it possible to define and deploy infrastructure consistently, reducing operational complexity and improving repeatability. Nebius Agents Blueprint extends that concept beyond infrastructure by packaging entire AI workflows including models, retrieval systems, orchestration frameworks, observability tools and supporting services, into reusable patterns. The goal is no longer simply to deploy resources. It is to accelerate the creation of working AI systems by reusing established system templates.

A customer support agent, a research assistant, a search application or an autonomous system can increasingly be viewed as a reusable implementation pattern rather than a unique integration project. Blueprints capture the architecture, tooling and operational practices proven in earlier deployments, allowing organizations to start from a working system rather than assembling one from individual services. Each deployment builds on the experience of the last, accelerating delivery while reducing the risk of repeating previous mistakes.

During the cloud era, infrastructure became a service. In the emerging AI era, workflows themselves are becoming the product. This perspective also helps explain why some AI cloud providers are expanding beyond infrastructure into higher-level platform capabilities. Rather than focusing solely on provisioning infrastructure resources, these platforms increasingly help customers assemble, deploy and operate AI applications while abstracting much of the underlying infrastructure complexity.

“This shift reflects a broader industry trend. If customer workflows become the primary unit of consumption, then understanding how customers build AI systems becomes strategically important,” posits Hein. “Product roadmaps increasingly emerge from observing successful implementation patterns rather than simply adding new infrastructure capabilities.”

The platform evolves by helping customers move more efficiently from intent to execution. This is one of the more significant shifts currently occurring within the AI industry.

From Self-Service Infrastructure to Self-Service Outcomes

Cloud computing transformed software development because it dramatically reduced friction. A developer could enter a credit card, provision resources and begin building immediately. The infrastructure became available on demand, eliminating procurement cycles, hardware purchases and much of the operational overhead associated with traditional IT.

AI cloud providers are extending this concept. The first generation delivered self-service infrastructure. The next generation is delivering self-service outcomes. Rather than asking customers to manually configure models, inference endpoints, observability systems, security controls and orchestration frameworks, the platform increasingly takes responsibility for assembling those components.

Recent efforts to incorporate agents directly into cloud platforms provide an early glimpse of this future. Instead of requiring users to navigate hundreds of APIs, cloud services and deployment decisions, an intelligent layer can understand the customer’s objective and configure the environment accordingly.

Nebius recently introduced Nebius Echo, an AI assistant built directly into its cloud platform that allows customers to interact with their infrastructure using natural language. Rather than serving as another chatbot, Echo begins shifting cloud operations from manual configuration toward intent-driven execution, where the platform can explain services, inspect resources and perform infrastructure operations with appropriate user approval.

The customer describes the desired result, then the platform manages execution. Infrastructure remains essential but becomes more of the foundation on which the customer expertise adds value, rather than the end-result.

The Real Inflection Point

The common assumption is that AI infrastructure providers are competing to build the largest clusters. To some degree, they are. Demand for compute continues to grow and there is little doubt GPUs will remain foundational to the industry for years to come.

But focusing exclusively on infrastructure risks missing the larger transition underway. Even under the baseline scenario in the aforementioned whitepaper from Tirias Research, practical infrastructure deployment begins falling behind projected demand around 2028, leaving roughly 72 quadrillion tokens of annual inference demand unserved by 2030 despite assuming no major geopolitical or supply-chain disruption.

That shortfall is not simply an argument for building more datacenters. Given the current limitations on real estate, power and cooling, that might not be realistic even if someone wanted to go that route. Instead, it highlights that AI is evolving into a systems problem. As workloads become longer-running, more autonomous and increasingly interconnected, the value shifts from raw compute capacity toward the platforms that can orchestrate models, memory, tools, data and workflows efficiently.

The first phase of AI cloud was about access to compute. The emerging phase is about abstraction and focusing on reducing the expertise required to use that compute effectively. Agentic AI is what makes this transition urgent. Agents do not simply consume compute. They consume orchestration, judgment and workflow. A single model answering a question is relatively straightforward to deploy. An agent that plans autonomously, manages memory, calls external tools and executes multi-step workflows exposes every gap in the stack. Infrastructure alone was never designed to close those gaps.

That is the real inflection point. Not a shift away from AI cloud, but a maturation of it. As Tirias Research’s Hein best puts it, “The most successful AI cloud platforms will not be those that expose the most technology. They will be the platforms that allow customers to think about technology the least.”

The cloud era taught organizations how to consume infrastructure. The next phase of AI cloud is teaching infrastructure providers how to consume customer intent. When that happens, the defining characteristic of an AI cloud will no longer be the hardware it operates. Instead, it will be measured by how effectively it transforms customer goals into working AI systems—and how little customers need to think about everything in between.

Tirias Research tracks and consults for companies throughout the electronics ecosystem from semiconductors to systems and sensors to the cloud. Members of the Tirias Research team have consulted for IBM, Nvidia, Qualcomm, AMD and other companies throughout the data center, AI and Quantum ecosystems.