Companies across Europe continue to invest heavily in artificial intelligence to accelerate software development, improve data analytics, automate business processes, enhance customer service, optimize supply chains, and support research initiatives.
According to research cited by McKinsey & Company, many organizations begin to see measurable financial returns from AI initiatives within one to two years, while the long-term impact can significantly improve productivity, operational efficiency, and business growth.
AI has the potential to reduce development costs, shorten product launch cycles, and increase organizational agility. However, the benefits are not distributed evenly. McKinsey estimates that approximately 75% of the economic value generated by generative AI comes from just four areas: software engineering, marketing and sales, customer operations, and research and development.
As a result, businesses are increasing capital investments in AI infrastructure and computing capacity. The goal is straightforward: deploy AI faster, integrate it into core business processes, and bring new products and services to market ahead of competitors.
Yet many organizations discover that securing the necessary infrastructure is far more complicated than securing budget approval. GPU accelerators remain difficult to obtain, delivery timelines are often unpredictable, and demand continues to outpace supply across multiple regions. In many cases, AI initiatives are delayed long before deployment begins—not because of technical challenges, but because the required hardware cannot be delivered on schedule.
We spoke with Dmitrij Avramov, CTO at ServerMall, about why GPU infrastructure remains one of the most constrained segments of the IT market, how procurement strategies have changed, and why timing has become a critical factor in AI projects.
— Dmitrij, many companies across Europe are investing in AI infrastructure. Is the GPU shortage still a real issue, or has the market started to stabilize?
— Unfortunately, the shortage remains very real. In fact, from an infrastructure perspective, the situation is often more challenging than many public discussions suggest.
The issue is no longer limited to GPU availability alone. Today, organizations are competing for access to the entire AI infrastructure stack: accelerators, high-density GPU servers, power capacity, rack space, and, in some regions, even data center availability.
What makes the market particularly challenging is that the largest technology companies continue to absorb a significant share of global supply. Hyperscale cloud providers such as Amazon, Google, Microsoft, and Meta, along with AI companies building large-scale training environments, purchase infrastructure in volumes that few enterprises can match.
These organizations negotiate directly with manufacturers and commit to long-term procurement programs. As a result, a large portion of future production capacity is effectively reserved well before the hardware reaches the broader market.
For everyone else, the challenge is not necessarily whether GPUs exist, but whether they can be obtained within the required project timeframe.
— How quickly does available inventory disappear? And why are organizations buying so much GPU capacity?
— In some cases, available stock lasts only a few days. Occasionally, it disappears within hours.
We regularly see situations where a shipment of GPU servers becomes available and is fully allocated almost immediately. Organizations understand that missing one procurement window may mean waiting several additional months for the next opportunity.
At the same time, semiconductor manufacturing capacity cannot be expanded overnight. Producing advanced AI accelerators requires highly specialized fabrication processes, and building new manufacturing facilities takes years and investments measured in tens of billions of euros.
As for demand, the role of GPU servers has fundamentally changed.
A decade ago, GPUs were primarily associated with scientific computing, engineering simulations, and rendering workloads. Today, they have become a strategic component of enterprise infrastructure.
Organizations are deploying AI assistants for software development, implementing generative AI applications, automating internal workflows, improving cybersecurity operations, and accelerating data analysis. Demand is growing simultaneously across finance, manufacturing, healthcare, telecommunications, retail, logistics, research institutions, and the public sector.
In other words, GPU infrastructure is no longer a niche technology. For many organizations, it is becoming a foundational layer of digital transformation.
— So does speed of decision-making now become the most important factor?
— Absolutely. Today, speed has become one of the key competitive advantages in AI infrastructure projects.
Large-scale deployments based on high-end accelerators such as NVIDIA H200, B200, or B300 are often decided within days, and sometimes even within hours. Companies that need modern GPU infrastructure cannot always follow traditional procurement cycles with multiple rounds of negotiations and lengthy internal approvals.
If an organization waits too long, the issue is not only that the price may change. The bigger risk is that the required hardware capacity will simply no longer be available within the required timeframe.
Companies need to adapt their procurement approach: faster technical validation, faster budget approval, and faster decision-making. Otherwise, the infrastructure reserved for one AI project may be allocated to another customer.
— You mentioned that projects can fail before they even start. Why does that happen?
— Because the market has changed. Organizations that are best prepared are the ones that secure infrastructure first.
In the past, enterprise customers could receive several offers, spend weeks comparing vendors, run extended procurement procedures, and return with a decision later. This approach worked well for traditional IT infrastructure projects.
AI infrastructure is different. When GPU availability is limited, a company that spends several weeks evaluating options may return to the market and discover that the required systems are no longer available or that delivery timelines have moved significantly.
Let me give you an example from our experience. In 2025, before the latest wave of memory price increases, we received a request from a customer looking for six GPU servers equipped with NVIDIA H200 accelerators.
We prepared a commercial proposal, but the customer decided to take additional time to evaluate alternatives. They considered other suppliers because they believed the initial pricing was too high.
Approximately two months later, they returned with the same project. However, by that time, the market conditions had changed significantly. Due to rising GPU and memory costs, the total project price had increased by roughly two times compared with the original proposal.
This time, the customer made a decision within four days.
Companies need to accept that GPU infrastructure procurement in 2026 is no longer a market where organizations can spend months comparing offers and negotiating every detail. It is becoming closer to a time-sensitive allocation process: either you secure the required capacity when it becomes available, or another organization will take it.
Initially, this was a major adjustment for enterprise customers. Large companies are traditionally built around structured approval processes and long procurement cycles. But AI infrastructure requires a different mindset — closer to the speed of technology startups, where decisions can move from initial concept to execution in a matter of days.
— So the challenge is not only the price? And why do these projects remain economically viable despite rising costs?
— The biggest challenge is availability. Price is often just a consequence of limited supply.
As I mentioned earlier, hyperscale companies and large AI infrastructure providers reserve significant amounts of GPU capacity. But there is another factor as well: speculative demand.
Sometimes real customers are unable to access available hardware because systems are purchased by intermediaries who plan to resell them at a premium. These transactions may offer immediate availability, but they also create additional risks.
Unlike established infrastructure providers, unofficial resellers often cannot provide long-term support, lifecycle management, warranty coordination, or replacement guarantees. For experimental projects this may be acceptable, but for production AI environments and business-critical systems, these risks are difficult to justify.
Regarding profitability, many AI infrastructure projects remain economically attractive even with higher hardware costs.
Companies increasingly calculate ROI not only based on the initial capital expenditure, but also on the cost of delaying innovation.
If an AI cluster allows a company to develop new models faster, automate operational processes, improve customer experiences, or launch a product six months earlier than competitors, the financial impact can outweigh the additional infrastructure investment.
Of course, businesses still need to evaluate projects carefully. But many organizations now understand that the biggest risk is not only paying more for infrastructure — it is losing time in a market where technological advantages are measured in months, not years.
— How significant is the price difference compared with official manufacturer pricing?
— This is one of the most interesting aspects of today’s GPU market because the same hardware can effectively have several different price levels depending on availability and delivery timeline.
The first option is the manufacturer-level price. This is the lowest price and usually the closest to the figures announced during vendor presentations. However, customers often need to wait 10–12 weeks just for production capacity to become available, followed by transportation, integration, and deployment.
The second option is a confirmed allocation through an approved project pipeline. In this case, delivery times are usually shorter — around 6–8 weeks — but the price is already significantly higher because capacity has been reserved and allocated through a specific supply chain.
The third option is purchasing equipment that is already available from stock. This provides the fastest deployment timeline, but prices can be dramatically higher due to limited availability and market premiums.
Nevertheless, companies continue to purchase even at these levels because, in many AI projects, the cost of waiting can be higher than the cost of the hardware itself.
— How has the supply chain changed? Has logistics become more complicated?
— Significantly. The global supply chain for advanced computing hardware has become much more complex.
Previously, companies could rely on established distribution channels, predictable transportation routes, and standard procurement processes. Today, AI infrastructure requires much more detailed planning because manufacturers, distributors, and customers all need to manage increased demand, limited production capacity, and stricter compliance requirements.
We work with international partners across different regions, including major technology markets in Asia, but today the entire ecosystem has become much more controlled and structured.
Manufacturers themselves have strengthened their supply chain requirements. For example, leading GPU vendors increasingly verify end customers, project details, and intended use cases to ensure compliance with export regulations and internal policies.
In some large-scale deployments, customers may need to provide additional documentation, confirm installation environments, or demonstrate that the equipment has been deployed according to the original project requirements.
This creates a completely different procurement environment. A single mistake in documentation, unclear project information, or compliance concern can delay or stop a shipment.
In some cases, AI infrastructure can travel through multiple countries and logistics hubs before reaching its final destination — not because the technology is unavailable, but because the entire supply chain has become significantly more controlled and complex.
— Dmitrij, what are the biggest risks businesses face when purchasing server infrastructure today?
— It is difficult to point to only one risk because companies are dealing with several challenges simultaneously: access to the latest technologies, pricing uncertainty, shortage of qualified specialists, infrastructure complexity, and long-term support.
But if I had to choose one factor, I would say that selecting the right technology partner has become more important than ever.
A lot depends on the supplier: technical expertise, understanding of AI workloads, access to hardware, logistics capabilities, and the ability to support customers throughout the entire infrastructure lifecycle.
In the past, when enterprise customers had straightforward access to official hardware and software channels, the main concern was usually choosing the correct architecture — ensuring high availability, selecting the right storage and networking design, and matching hardware specifications to expected workloads.
Today, companies also need to consider whether their infrastructure partner can actually deliver the required systems, provide lifecycle support, and help manage a much more complex technology environment.
— And the final question: looking ahead, what is your main advice for companies investing in AI infrastructure over the next few years?
— If AI infrastructure and GPU computing are genuinely important for your business, the most important thing is to start planning early and work with a reliable technology partner.
Companies should optimize their internal planning and procurement processes to reduce the time between identifying an AI opportunity and securing the necessary infrastructure.
The organizations that benefit most from AI will not necessarily be those that simply buy the most powerful hardware. They will be the ones that successfully integrate AI into their existing business processes.
When implemented correctly, AI can significantly improve productivity and operational efficiency. For example, Toyota has reported thousands of hours in annual productivity improvements through AI-powered solutions across engineering and business operations.
These results do not appear instantly, and not every company will achieve the same level of impact. However, organizations that approach AI strategically — combining the right infrastructure, data strategy, and business processes — can generate substantial returns.
The biggest mistake businesses can make today is waiting too long. In an environment where AI capabilities evolve every few months, delaying infrastructure decisions can mean losing valuable time compared with competitors who are already deploying these technologies.
If you enjoyed the interview, make sure to follow us on LinkedIn so you don’t miss the release of new materials.
