How to mitigate the impact of the memory shortages

AI data centre expansion has led to high DRAM demand, hyperscaler stockpiling, and wafer capacity reallocation to high bandwidth memory (HBM) all of which has thrown the supply chain into disarray. HBM stacks chips vertically to deliver more bandwidth and higher capacity with lower power consumption and a smaller footprint, making it the foundational component of modern AI systems. This has seen supply diverted to a more lucrative new market, with some manufacturers (or ‘fabs’) signing over their entire consignment to AI for the year. 

The resulting shortages have seen prices rocket. A spike in prices in Q4 2025 has continued to be felt through into this year and there’s no real prospect of them levelling out until 2027. In fact, the three top producers of DRAM are sold out into next year, even though they’re operating at full production, while new production lines have yet to come online. This is seeing lead times extended to beyond 40 weeks for large DRAM orders, meaning many will have to abandon FY 2026 plans. 

Some may choose to wait it out but given that there really is no way of knowing when the market will self-correct, it makes more sense for IT teams to look to limit the impact of memory shortages, and that means contingency planning. But where and how should they begin and what are the most effective steps they can take to ride out the drought?

Taking stock

Approximately 70% of the memory used today is located in datacentres which means most are going to have to juggle capacity. To start with, it’s necessary to understand what workloads you have as a business by assessing what’s running, where and how much is currently being utilised. Right-sizing is important because it reduces the risk of capacity overspend or the propensity to panic purchase, neither of which you want to do when capacity is priced at a premium.

Once the IT team has an understanding of where capacity is being utilised, they can decide on what’s critical and prioritise those workloads, because what you don’t want is a low priority workload monopolising memory and capacity. But it’s not a decision that can be taken in isolation. IT will need to engage with business leaders to ascertain the importance of specific workloads and to enable the business to pursue its goals and objectives. 

Secondly, seek to optimise server configurations so that they use memory efficiently. By collaborating with IT architects, these configurations can be aligned with validated DRAM modules. Optimisation also allows the business to avoid excessive memory densities that can exacerbate lead-time risks. And look to see if you can leverage the solutions being offered by vendors to offset memory issues. VMWare, for instance, allows its customers to align their infrastructure plans with the company’s emerging standard five-year licensing agreements which allow IT teams to use memory tiering to reduce memory demands.

Look at all the options

Thirdly, look to adjust plans to allow for a phased deployment. By splitting up deployment across the fiscal year, prioritising critical workloads and deferring those deemed non-critical to a later date when pricing stabilises, it’s again possible to reduce risk. For instance, the business could choose to acquire servers this year with half memory capacity and plan for increases in 2027 to better align memory consumption with a three-to-five-year lifecycle.

Another thing to consider is GPU accelerated workloads. These can allow the business to reassign infrastructure that was earmarked for AI for non-AI use cases. Other workarounds include examining current cloud allocation to see if there is any bandwidth there that can be utilised. The business could, for example, add capacity in the cloud this year to allow it to invest in on prem capacity next year. Or if a CAPEX purchase is now out of the question, OPEX models such as leasing or financing might be a viable option. 

Among the customers we are talking to today, most are either going to make an initial upfront purchase if they have the capital or they're shutting off workloads to give them the capacity to meet their plans for the year. None are standing still and waiting it out until 2027. That just isn’t an option in today’s market and frankly nobody knows when this issue will be resolved.

Future risks

There are numerous macroeconomic factors at bay affecting price and availability. The situation could worsen if the major cloud providers who have also been stockpiling memory in anticipation of massive AI growth have created a backlog that the fabs cannot clear it fast enough. We also need to see the fabs rebalance the market by no longer prioritising HBM production as this has reduced the supply of traditional DRAM for server RDIMMs. And then there’s the issue of time – the longer this goes on the more unmet business demand there will be.

On the other hand, there are factors that could improve supply. It’s been reported that the fabs are asking for more information from their customers when taking orders to reduce the risk of stockpiling. There have been objections to AI datacentre builds at the local level and concerns have been raised that the grid will not be able to supply sufficient power to some sites, both of which could curb demand. And additional fab production lines are expected to come online, increasing throughput. 

But crucially, businesses also have a part to play. If IT teams can redesign their data centre architectures to use memory more efficiently, demand will also reduce and if they continue to optimise their usage that reduction could even be permanent. That would help the market recover but it also has additional benefits in because it provides a sustainable way for the business to stabilise, manage costs and achieve its technological goals without being at the mercy of the market.