The Death of the Cold Aisle

From Buying Hardware to the OPEX Model: How AI Infrastructure Hit a Physical Wall in 2026

Intro We used to perceive artificial intelligence as the magic of software code—something ethereal existing solely in digital space. However, 2026 mercilessly shattered that illusion. It turned out that behind every generated image or demand forecast lie not just algorithms, but the very tangible laws of physics, economics, and geopolitics.

The digital revolution has collided with harsh reality: electricity is not infinite, silicon has a density limit, and state borders still exist on the map. Old methods—buying new servers or expanding classic clouds—no longer work. They are too slow, too expensive, and physically butt against the ceiling of power grid capabilities.

To understand how the industry plans to exit this deadlock, we must descend from the level of code to the concrete floors of data centers and silicon wafers. In the following sections, we will trace how global business is forced to radically rethink its approaches. We will see why owning compute «metal» has turned into a heavy burden, how data learns to process itself without leaving the disk shelf, why the future of edge AI is linked to light transmission technologies, and how geopolitics forces corporations to literally tear their IT systems apart to meet contradictory requirements of different countries. This is the story of how a completely new, hybrid, and pragmatic world of high technology is being born from the wreckage of old server architecture.

From Buying Hardware to the Operating Model The AI industry is undergoing a fundamental shift, similar to the transformation of the classic server market a decade ago, but at a multipled pace. The concept of owning compute clusters (CAPEX) has finally given way to a consumption model via subscription (OPEX). The global GPU rental market has surpassed the $9 billion threshold, showing over 40% annual growth, while in Russia, this segment reached 17 billion rubles by the end of 2025 with a stable doubling dynamic.

The CAPEX Deadlock Traditional IT investment logic, calculated on a 3-to-5-year equipment depreciation, shatters against the reality of the neural network race. The lifecycle of a graphics accelerator has shrunk to 12–18 months. Each new revision of Nvidia chips (the transition from H100 to B200 architecture and preparation for Rubin) offers a multiple increase in energy efficiency.

Companies buying equipment today face a double blow:

Depreciation: In a year and a half, their multi-million assets lose up to 70% of their liquidity on the secondary rental market.
Talent Shortage: Operating a cluster requires maintaining a staff of engineers capable of supporting extreme liquid cooling regimes and optimizing low-level software.

A private GPU park has turned from an asset into a heavy financial burden where the cost of ownership outweighs any benefit of independence.

The Strategy of Giants The $17.4 billion contract between Nebius and Microsoft (with potential expansion to $19.4 billion) became a marker of the new reality. The paradox is that Microsoft, one of the few players capable of building mega-data centers on its own, consciously chooses the role of a tenant.

This choice is dictated by pragmatism:

Bypassing Infrastructure Barriers: Erecting a data center for B200 chips requires solving the most complex power supply and logistics issues, delaying launch by two years. Nebius guarantees the provision of dedicated racks in its Vineland (New Jersey) node by the end of 2026.
Guaranteed Silicon: In a global deficit, access to the latest accelerators is the main market currency. Nebius, having received $2 billion in direct investments from Nvidia in 2026, holds priority delivery quotas, while regular clients wait in line for months.
Narrow Specialization: Unlike universal hyperscale clouds, Nebius has built a vertical exclusively for AI. Microsoft buys not just processor time, but a ready, perfectly calibrated production cycle for training foundation models.

Hypergrowth Paradoxes and the Economics of Deficit Nebius’s financial results for Q1 2026—684% revenue growth—look like a statistical anomaly. However, this is a consequence of a «low base» and perfect timing: the company completed its transformation and built supply chains at the exact moment demand for compute exceeded global power.

At the same time, the GPU-as-a-Service business model contains an internal contradiction: to generate stable operational rental income, the provider must first freeze colossal sums in buying deficit iron. Analysts estimate the potential of this market at $260 billion by 2030 with a 35% CAGR.

The AI industry has finally transitioned to the phase of infrastructure utilitarianism. Compute has become a utility. Owning physical hardware remains a privilege of only a few state hyperscalers and defense structures. The entire commercial AI sector has bet on the operating model, accepting the risk of dependence on a narrow circle of specialized power suppliers in exchange for speed, scalability, and financial flexibility.

From Storing Bits to Data Activation The acquisition of the DataPelago startup by NetApp symbolizes the final point in the history of classic storage systems (SCS). Traditional IT architecture, built on the principle of «Dump terabytes from the lake, clean them in a separate cluster, and only then send them to the GPU,» can no longer withstand the economic load of the AI era.

The Penalty for Moving Companies are massively buying clusters based on Nvidia B200, but face a paradox: expensive accelerators are idling. The culprit is the infrastructure gap. While gigabytes of data are copied over the network from storage to the compute node, the GPU server rental meter keeps ticking. In the cost structure of model training, up to 80% of time may be spent not on mathematical calculations, but on the simple waiting for data packets. In the OPEX model, every extra minute of network wait is a literal burning of the operating budget.

DataPelago Architecture DataPelago proposed a solution the industry calls «Zero-Copy.» Instead of building a separate farm for data preprocessing, their DataPelago Nucleus technology integrates the compute core directly into the storage layer.

The startup, which emerged from stealth in late 2024, developed a Universal Data Processing Engine (UDPE) that embeds between application software (like Apache Spark or Trino) and physical hardware. The system uses DataVM—a specialized virtual machine with its own instruction set. It analyzes available resources in real-time: if the SCS controller has free GPU tensor cores, the task goes to them; if only CPU—optimization happens on the processor. This allows filtering and aggregating data 4–10 times faster than standard solutions, as information physically never leaves the disk shelf.

Economics of Activation For NetApp, this deal is not just buying promising software, but a paradigm shift. As CEO George Kurian emphasizes, modern enterprises need infrastructure that matches the power and intelligence of the AI models themselves.

Integrating DataPelago into the NetApp portfolio (Intelligent Data Infrastructure) eliminates the heavy ETL (Extract, Transform, Load) stage. Data is «activated» the moment it hits the disk. For business, this means a shift from reactive to proactive analytics: fraud is detected in the microsecond of a transaction, and equipment failure is predicted by the primary sensor impulse, without the need to gather data in a central lake.

Redesigning the Data Center Integrating compute into SCS forces data center engineers to revisit basic design principles:

Thermal Profile of SCS: Disk arrays are no longer «cold.» The presence of built-in GPUs and FPGAs in storage controllers requires implementing liquid cooling directly into SCS shelves, which were previously cooled by general airflow.
Erasing Network Boundaries: The classic separation into Storage (SAN) and Compute (LAN) networks loses meaning. Data center topology becomes a flat compute pool where an AI task can be performed on any available accelerator, whether in a server or a disk shelf controller.
Vectorization at Ingestion: Data is indexed and turned into vector representations (embeddings) at the moment of writing. The storage stops being a library of dead files and becomes a living factory for preparing information for neural networks.

The NetApp and DataPelago deal closes the last gap in the AI stack. If players like Nebius solve the deficit of physical power (compute), NetApp solves the problem of feeding that power. Infrastructure ceases to be a passive archive. It becomes an active participant: data is prepared, managed, and enriched with context the very second it is created, completely eliminating the idle stage of expensive GPU clusters.

Moving Away from GPU Monoculture to Optical Chips The race for power, expressed in the endless scaling of GPU clusters, has hit an insurmountable wall—the laws of thermodynamics and the capacity of power grids. Data centers for artificial intelligence have become the main consumers of electricity, and in Russia, connecting new megawatts to the grid has become the main infrastructure barrier. Further scaling by old rules became economically and physically impossible.

Division of Responsibility The industry responded to this deadlock with fragmentation. A clear dualism emerged: heavy training cycles of foundation models remain in the hands of global cloud providers like Nebius, who continue to operate deficit GPUs. However, the practical application of these models (inference) is rapidly moving to the edge.

Robotics, unmanned transport, and smart city systems cannot afford to send every video frame to the cloud—network latency (latency) will kill autonomy, and traffic bills will ruin the business. Compute must happen on-device. But you cannot fit a classic hot GPU into a drone or a camera. Next-generation Edge AI enters the scene, where electronics is replaced by photonics.

The Anatomy of an Optical Processor Optical (photonic) chips use the movement of photons instead of electrons to perform calculations. Since light moves without electrical resistance, it generates no excess heat and faces no «traffic jams» in copper conductors.

The key advantage manifests in matrix multiplication—the most resource-intensive operation in neural networks. Photonic cores perform these tensor computations orders of magnitude faster and more efficiently. If a classic GPU spends energy heating conductors, the optical chip spends it almost exclusively on useful work. This allows creating compact compute units that consume dozens of times less energy and do not require bulky cooling systems.

Strategic Advantage for Russia For the Russian IT industry, the transition to optical compute is not just an architectural shift, but a matter of technological sovereignty. Replicating the topology of modern GPUs (3–5 nanometers) is currently nearly impossible. However, photonic chip production relies on different technological chains that are much easier to localize.

This opens a path to creating independent peripheral infrastructure. Compact optical accelerators allow building «Sovereign AI» for industrial robots, drones, and security systems, completely removing dependence on importing deficit video cards. As noted in the industry, this perfectly fits the import substitution trend: a cold, fast, and autonomous chip that can be produced domestically.

The New Topology of Reality AI infrastructure has finally ceased to be centralized. We are witnessing the formation of a two-circuit model:

Super-dense Core: Giant, energy-hungry GPU factories in the cloud, where liquid cooling systems boil for training models.
Cold Periphery: A distributed network of autonomous devices based on optical chips that make decisions in real-time without turning to the cloud.

The future of compute is no longer defined solely by the number of transistors on a die. It is defined by the speed of light. The era of monolithic GPU clusters is being replaced by a hybrid era, where heavy intelligence lives in the global cloud, and instant reaction is provided by optical neurons distributed across the physical world.

Sovereignization and Standardization AI infrastructure has split into two parallel dimensions: rigidly regulated national perimeters and flexible transnational resource pools. Geopolitics has transformed from news background into a fundamental system design parameter, standing on par with chip architecture and bandwidth.

Law as a Driver of Engineering Thought In Russia, the infrastructure crisis caused by limited access to Nvidia solutions catalyzed the process of technological self-determination. The market took the path of forced diversification: establishing parallel import schemes for critical tasks and a mass transition to alternative accelerators from the PRC (specifically Huawei Ascend and Biren lines) for standard loads.

However, the main factor of transformation became legislation. The Federal Law on AI (No. 243-FZ), coming into force on September 1, 2026, introduced the official concepts of sovereign and national models. As Oleg Sazhin emphasizes, the state shifted focus from the formal origin of code to technological control and data jurisdiction.

Sovereign status mandates using solutions from Russian legal entities with full localization of compute and storage on RF territory.
National status allows the use of Open Source components if the possibility of independent modification and operation is preserved.

This regulatory framework did not become an insurmountable barrier—a long transition period until 2032 is provided for existing systems, and radical requirements (e.g., training only on local data) were softened. Nevertheless, legal compliance now dictates network topology. When designing IT systems, architects are required to factor in the physical location of servers (according to FZ-152) and the legal status of the AI model as basic equation variables.

Transnational Intermediaries and the Multi-Cloud Era While nation-states erect regulatory barriers, global capital seeks ways to bypass them through infrastructure flexibility. The market’s response to fragmentation was multi-cloud—a strategy where a company distributes load between different providers to avoid dependence on a single jurisdiction or vendor.

The $17.4 billion Nebius and Microsoft deal illustrates the birth of a new type of player—a transnational infrastructure broker. Nebius, having absorbed the Yandex engineering school but registered in the Netherlands, acts as a buffer. It accumulates access to deficit iron and provides it to giants like Microsoft, relieving them of the need to independently solve logistics problems with GPU shipments. Deploying power in New Jersey for the American giant, Nebius proves: in 2026, owning a physical server is less important than owning engineering expertise and contracts for chip supply.

The Architecture of Hybrid Reality To survive between the requirements of Sovereign AI and the need to use global models, corporations have switched to hybrid schemes that have become the de facto standard:

Cloud for Heavy Intelligence: Resource-intensive training (Training) is offloaded to specialized international platforms (Nebius and analogs) where clusters of deficit accelerators are concentrated. Here, the laws of global economy and physics apply.
Local Perimeter for Sensitive Data: Inference, work with personal data, and critical business processes are locked inside the country. Here, models that have acquired sovereign status are operated, and traffic does not leave the protected perimeter.

The connecting layer is provided by orchestration platforms. They ensure seamless movement of workloads: data is anonymized and goes to the global cloud for model enrichment, and updated weights return to the local data center for user operation.

The monopoly of one supplier (Nvidia) and one cloud (AWS/Azure) is dead. Modern AI infrastructure is a complex construct assembled at the junction of law and engineering. It is transnational where massive compute is needed, and extremely localized where it comes to data storage and national security. Market leaders are not those with the most racks of video cards, but those whose architecture can seamlessly exist simultaneously in several legal and technological realities.