Durantic Launches to Tame AI's Fragmented Infrastructure

LONDON, UK – May 20, 2026 – As the artificial intelligence boom shifts from a race for hardware to a struggle for operational control, a new company, Durantic, has emerged from stealth to offer a solution. Launched publicly this week, the London-based startup, founded by veteran infrastructure engineers from Meta and Hudson River Trading, aims to become the essential operating layer for the increasingly chaotic world of AI compute.

Durantic is tackling a problem that has become a critical bottleneck for AI developers, from frontier labs to large enterprises: managing the complex and fragmented infrastructure that underpins modern AI workloads.

The New Bottleneck in AI

For the past several years, the primary constraint in AI development was access to powerful GPUs. Now, as supply chains ease and a variety of compute options become available, a new, more complex challenge has emerged: operationalization. The AI infrastructure landscape is no longer a simple choice between on-premise and a single cloud provider. Instead, it's a fragmented ecosystem of hyperscalers, specialized GPU clouds like CoreWeave and Fluidstack, colocated bare-metal servers, and sovereign AI clusters.

This fragmentation creates a daunting operational matrix. Companies find themselves managing mixed generations of GPUs, running different orchestration systems like Kubernetes and Slurm side-by-side, and wrestling with custom high-performance network fabrics. The mundane but critical tasks of hardware acceptance, burn-in testing, and managing RMA evidence collection have transformed from routine IT work into major operational hurdles that can stall development.

Compounding the issue is a severe talent gap. The specialized expertise required to run high-performance, bare-metal infrastructure at scale is rare. Many enterprises investing in "private AI factories" are discovering their existing IT teams lack the depth to manage these environments effectively. Even the most advanced AI labs are visibly hiring for hyperscaler-level infrastructure operators, signaling a widespread shortage of the necessary skills. The result is that AI companies are being forced to become infrastructure experts, diverting precious time and resources away from their core mission of building intelligent models.

An Operating System for Fragmented Compute

Durantic's answer to this chaos is a managed service underpinned by a proprietary, bare-metal-native control plane. The company offers to take on the entire operational burden of the infrastructure layer, allowing customers to focus on their AI applications.

"AI companies are being forced to become infrastructure operators, and most don't want to be," said Ivan Diachenko, Durantic's founder and CEO, in the launch announcement. "We write the software that operates AI infrastructure, and we operate it for customers directly. That combination is what produces operational leverage at scale."

The company's model is straightforward: customers bring their compute—whether it's owned, leased, or colocated—and Durantic operates it. The service begins with a 30-day pilot on a defined slice of the customer's GPU fleet. If successful, it transitions to a recurring monthly fee that scales with the number of GPUs and servers under management.

Underneath the managed service is the company's core technology: a control plane designed from the ground up for bare-metal environments. An agent running on each machine handles the full lifecycle, from initial provisioning, network configuration, and storage coordination to telemetry collection and hardware maintenance. This integrated approach means every fleet Durantic operates contributes to sharpening the software, creating a powerful feedback loop between the product and the service.

Hyperscale Expertise Meets the Talent Gap

The credibility of Durantic's ambitious promise rests heavily on the deep technical expertise of its founding team. CEO Ivan Diachenko was a founding member of Meta's Traffic Disaster Recovery team, where he built systems to ensure network resilience for over three billion users. He later spent four years at the high-frequency trading firm Hudson River Trading, building and managing low-latency bare-metal infrastructure for global trading operations—an environment where performance and reliability are paramount and cloud abstractions are often a liability.

This experience with massive, high-stakes, bare-metal systems is complemented by co-founder and CTO Dmitrii Skokov. As the former Infrastructure Platform Lead at Replika, Skokov was responsible for building the MLOps and GPU provisioning systems that powered the AI companion service at scale. His background provides the crucial link between raw infrastructure and the specific needs of machine learning workloads.

Together, their experience represents the exact blend of skills that companies now find so difficult to hire: systems software engineering, platform engineering, and hands-on infrastructure operations. Durantic effectively productizes this rare expertise, offering it as a service to bridge the talent gap for companies that need hyperscale-level operations without building a hyperscale-sized team.

The Next Great Infrastructure Play?

Beyond solving an immediate technical problem, Durantic is making a strategic bet on the future structure of the AI market. Diachenko posits that the market is bifurcating into two distinct layers: a capital-and-capacity layer and an operating layer.

The capacity layer consists of the entities that procure and provide the physical compute, from neoclouds like CoreWeave and Crusoe to national sovereign AI initiatives. This layer is increasingly defined by capital-intensive activities like GPU procurement, data center construction, and securing anchor tenant contracts.

Durantic argues that the operating layer—which handles provisioning, networking, orchestration, and hardware lifecycle management across all this heterogeneous compute—currently has no incumbent leader. This is the space the company intends to dominate.

"The pattern rhymes with VMware over commodity servers, Datadog over cloud visibility, Red Hat over Linux," Diachenko stated. "Operating layers above commoditizing capacity are historically the most durable infrastructure businesses. AI infrastructure is producing the same dynamic now."

This thesis positions Durantic not merely as a services company, but as a potential foundational element of the AI stack, analogous to the operating systems and management platforms that brought order to previous eras of technological disruption. By combining software that gets smarter with every deployment and a service that scales with its customers' fleets, the company aims to build a durable, high-value business on top of the commoditizing world of AI compute.

With its headquarters in London, Durantic is now focused on building out its team and expanding its pilot program pipeline across the United States, United Kingdom, and European markets, beginning the journey of turning its ambitious vision for the future of AI infrastructure into a reality.