AI infrastructure demand looks contradictory right now. Memory is in short supply, yet announced data center projects far exceed what can actually be built. Together, the two markets show why AI infrastructure demand on paper is a weak guide to investment that can’t be undone.
Memory is scarce
TrendForce reports that conventional DRAM contract prices climbed by about 93% to 98% during the first three months of 2026. Suppliers gave priority to server applications while cloud providers competed for what was available. SK Hynix has warned that shortages of high-bandwidth memory could last until 2030.
Samsung, SK Hynix and Micron make around 90% of the world’s DRAM. They’ve pushed more production toward high-margin server memory and HBM, which squeezes the conventional memory used elsewhere. Smaller buyers struggle to win allocations against the hyperscalers.
The makers are still investing. Micron has laid out US plans worth over $250 billion for manufacturing and research through 2035. But new capacity is slow: fabs need years for construction, equipment installation and qualification, and TrendForce thinks the next supply gains will mostly come from better processes rather than more wafers.

Why fabs move cautiously
Memory makers remember the 2023 downturn. They cut production as prices fell, and SK Hynix by itself booked a 7.73 trillion won operating loss for the year. Having seen a shortage turn into a glut so recently, they’re expanding in stages.
Every new fab is a bet on future demand, poured in concrete. Approving one means betting that today’s demand survives several chip generations. It also means accepting that better model efficiency or new architectures could change how much memory is needed, or which kind, before the plant reaches volume production.
Grid queues tell a different story
Data center developers face other incentives. Sites, electricity and hardware all have to be locked in well ahead of opening day. When each is scarce, applying early and for more than you need can look rational.
Britain shows the effect. Grid connection requests jumped from 41GW to 125GW within a year, driven largely by data centers. Ofgem says those projects account for at least 80GW of the increase. The regulator now proposes commitment fees and extra milestones, so developers must show financial and commercial readiness, including procurement activity.
That reform separates reserving capacity from building it. A developer can hold land, a press release and a queue slot while having no signed customer, no completed financing and no realistic route to energization. Developers aren’t necessarily being deceptive, since waiting can mean losing access entirely. The problem comes when the applications are added up and read as independent, funded demand.
When investment becomes demand
Financing adds another wrinkle. Chipmakers increasingly invest in companies that are current or prospective customers. The most prominent case was Nvidia’s planned stake of up to $100 billion in OpenAI, linked to at least 10GW of Nvidia hardware. By December 2025 the final agreement still wasn’t signed, and Nvidia said any OpenAI orders sat outside its booked chip sales.
Supplier financing isn’t improper in itself, and strategic investment can help new technology reach scale. But when a chip vendor funds an AI lab and the lab spends part of that cash on compute, one relationship produces both an investment asset and product revenue. The demand may be real, but it isn’t fully independent of the company supplying it.
Which AI infrastructure demand numbers to trust
For CIOs, cloud providers and enterprises planning capacity, a real memory shortage can coexist with a speculative data center pipeline. A shortage now doesn’t prove every projected workload will appear. Abandoned projects, likewise, don’t make the shortage imaginary.
The most telling signals are commitments that cost a lot to walk away from:
- binding customer contracts and non-refundable deposits
- secured power and equipment already procured
- sites where construction has begun
After a data center opens, utilization and revenue from outside customers say more than its announced capacity. Demand backed by end customers also deserves separate treatment from demand supported by money circulating inside the AI supply chain, because the two show different things about whether the infrastructure will pay for itself.
Building in stages
Planning around AI infrastructure demand works better when spending follows proof.
Capacity plans can follow the same logic. Grow in steps as usage and returns can actually be measured, and match long-term commitments to demand that has already proved durable.
The industry is ordering too much and building too little, both at the same time. The companies carrying the longest-term physical risk, the memory makers, sound less certain than the big headline numbers imply.
Source: TechRadar