

The AI infrastructure market presents an apparent contradiction. Memory is in critically short supply. TrendForce reports that conventional DRAM contract prices increased by roughly 93% to 98% in the first quarter of 2026, as suppliers gave priority to server applications and cloud providers fought for available capacity. High-bandwidth memory shortages may persist until 2030, according to a warning from SK Hynix.
At the same time, announced data center demand increasingly exceeds what can realistically be delivered. In Britain, applications for grid connections rose from 41GW to 125GW in less than a year, largely because of data center projects.
One market is defined by physical scarcity. The other contains more proposed capacity than the system can accommodate. Together, they demonstrate why declared demand is a poor substitute for irreversible investment.
A forecast made in concrete
During the 2023 memory downturn, manufacturers cut production as prices plummeted. SK Hynix alone reported an annual operating loss of 7.73 trillion won. When AI demand accelerated, suppliers entered the boom with a recent reminder of how quickly shortages can turn into gluts.
Samsung, SK Hynix and Micron account for around 90% of the DRAM market. They have responded to AI demand by directing more production towards high-margin server memory and HBM. This has tightened the supply of conventional memory used elsewhere, even as smaller buyers struggle to compete with hyperscalers for allocations.
Memory and storage companies are still investing. Micron, for example, has outlined more than $250 billion in US manufacturing and research investment through 2035. But new capacity arrives slowly. Fabrication plants require years of construction, equipment installation and qualification. TrendForce expects most near-term supply growth to come from process improvements rather than major increases in wafer capacity.
A new fab is effectively a demand forecast made in concrete. To approve one, a manufacturer must believe that today’s demand will survive several generations of chips and AI products. It must also accept that improvements in model efficiency or changing architectures could reduce the amount or type of memory required before the plant reaches volume production.
The companies closest to the physical bottleneck are expanding, but in stages, and their caution sits uneasily beside the AI industry’s largest demand projections.
Announced capacity is not installed capacity
Data center developers face different incentives. Land, power and equipment must be secured long before a facility can begin operating. When each resource is scarce, applying early and applying for more than ultimately required can appear rational. The problem emerges when those applications are added together and treated as a reliable forecast of future demand.
Britain offers a clear example. Ofgem says data center projects accounted for at least 80GW of the recent increase in grid connection applications. The regulator is now proposing commitment fees and additional milestones, requiring developers to demonstrate financial and commercial readiness, including procurement activity.
The reform recognizes a basic distinction: reserving capacity is not the same as building it. A project can have a site, a public announcement and a place in the grid queue without having a committed customer, completed financing or a realistic route to energization.
We should not assume deception. Developers are responding to a system in which waiting can mean losing access altogether. However, the aggregate pipeline becomes misleading when every request is interpreted as independent, funded demand.
Infrastructure planners should focus on how much capacity has secured power, financing, equipment and customers, and how much ultimately reaches operation.
When investment also becomes demand
AI financing creates another difficulty when measuring the market. Chipmakers increasingly invest in companies that are also current or prospective customers. The most prominent example was Nvidia’s proposed investment of up to $100 billion in OpenAI, linked to the deployment of at least 10GW of Nvidia systems. In December 2025, Nvidia said the definitive agreement had not yet been completed and that the potential OpenAI purchases were not included in its existing chip bookings.
Supplier financing is not inherently improper, as strategic investment can help new technologies reach commercial scale. But it complicates how demand should be read.
When a hardware supplier invests in an AI company that then uses part of that capital to buy computing infrastructure, the same relationship generates both an investment asset and potential product revenue. Demand may still be genuine, but it is not fully independent of the company supplying it.
Technology leaders assessing the market should therefore distinguish between revenue funded by end customers and demand supported by capital circulating within the AI supply chain. These two indicators provide different evidence about whether infrastructure will remain economically productive.
Measure commitments, not announcements
CIOs, cloud providers and enterprises planning AI capacity should remember that a real memory shortage can coexist with speculative data center pipelines. Scarcity today does not prove that every projected workload will materialize, just as abandoned projects do not make the current shortage imaginary.
The most useful indicators are the commitments that are expensive to reverse: binding customer contracts, non-refundable deposits, secured power, equipment already procured and facilities that have entered construction. Once a data center becomes operational, utilization and revenue from external customers provide stronger evidence than its announced capacity.
Infrastructure plans should follow the same logic. Capacity can be added in stages as workloads, utilization and returns become measurable. Long-term commitments should be matched to demand that has already demonstrated some durability.
The defining contradiction of the AI boom is that the industry is over-ordering and under-building at the same time. The businesses taking the longest-term physical risk appear less certain than the headline numbers imply. For anyone making infrastructure decisions, that gap may be the most revealing evidence in the market.
We've featured the best AI tool.
This article was produced as part of , our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit
Chief Technology Officer at EXANTE.