Global memory shortage throws wrench into IT pros’ budgets, planning

Feeling forgetful?

You’re not alone; the IT industry is in the midst of a memory shortage that’s adding to budget pressures, delaying projects, and creating a whole lot of stress for IT pros.

The culprit is familiar: AI data centers, where major manufacturers are allocating chip production capacity to meet demand for the high-bandwidth memory these facilities need. Enterprise organizations’ IT departments are left in the lurch, scrambling to secure enough memory to keep AI and other digital projects on track.

Gartner expects PC shipments to fall more than 10% globally this year, ITPro reported, and for DRAM and solid-state drive prices to grow 130% by year’s end. One of the likely outcomes is enterprises holding off on device upgrades.

“In my 25-year career, we’ve never seen anything like this,” Bob Bogle, VP of devices at solutions integrator Insight Enterprises, told Morning Brew. “And it’s not a short-term problem. We for sure think this will go out through the end of 2027. So there’s some real strategy that has to be going into fleet management, device management, memory supply chain management.”

Bots enter the chat

Never ones to let a good crisis go to waste, scalpers are capitalizing on memory shortages by using bots to scrape RAM product pages, scoop up limited inventories, and resell them at a premium.

In a March blog post, DataDome’s VP of threat research, Jerome Segura, noted that “the explosive growth of AI is driving unprecedented demand for DDR5 RAM, and scalping bots are currently hitting DDR5 RAM product pages six times more often than legitimate traffic.”

DataDome’s Galileo threat research team discovered that bots were checking the inventory of some RAM kits every 6.5 seconds; DataDome blocked more than 10 million scraping requests. The bots are reportedly targeting both consumer and B2B RAM.

“By rapidly snapping up the limited DDR5 memory inventory for profitable resale, these bots further deplete the consumer supply, effectively boxing out legitimate customers and driving market prices even higher,” Segura wrote.

Working smarter

Val Cook, chief software architect at AI computing company Blaize, told us that the issue “is definitely affecting all of us.”

Blaize’s answer to this is “a hybrid architecture in which heterogeneous devices work cooperatively, applying the appropriate level of compute and requisite memory bandwidth to each processing domain,” according to Cook. One such hybrid approach involves Blaize’s chips running lower-cost, lightweight AI models that can determine if a more complex model is needed.

By hybridizing, Cook said, “you get a modern, complete, highly detailed experience at the cost of a much lower solution.”

But as workflows across business functions require more sophistication involving many AI models, “memory becomes a real problem,” he said.

“I don’t want to take a large GPU with a large amount of memory, pin that model to that GPU, and then park it there and use it 5% of the time,” Cook explained. Instead, he argues for hybrid architecture that supports switching between various contexts on small models—essentially, pairing the work with the right kind of chip.

Trevor Harries-Jones, COO and board member at decentralized GPU rendering platform Render Network Foundation, said in an email that it’s common for IT teams to run into issues with “procurement, availability, and timing” due to the legacy infrastructure models they operate within. However, he sees the issue as less about acquiring hardware and more about needing to access it more efficiently.

“What this moment really highlights is whether unused compute capacity (which already exists) can be harnessed and shifted toward shared and distributed compute models,” he said. “As AI demand continues growing, I think the conversation will increasingly move from ‘How do we buy more hardware?’ to ‘How do we better utilize the compute that already exists?’”

Sweat the fleet

Insight Enterprises helps clients navigate this challenge by implementing a multi-OEM strategy for enterprises that may have relied on one or two manufacturers before.

Brennen Smith, CTO of cloud GPU service Runpod, which has tens of thousands of GPUs across dozens of data centers around the world, emphasized the importance of a diversified supply chain; Runpod works with over two dozen providers.

One of the most tangible effects of not having enough memory is that it limits an enterprise’s ability to scale, Smith said. It throws into question what and when products and services can be launched, which necessitates paying close attention to financial metrics like internal rate of return and ROI on server investments.

“It becomes even more important when looking at the P&L at a product unit economics level,” Smith said.

He also recommended that IT pros analyze their qualified vendor list for memory to maximize flexibility, and standardize purchases for a certain type of memory at a given speed across the tech stack to enhance component interoperability.

Meanwhile, Bogle highlighted a strategy he called “sweating the fleet”—basically, making hardware last longer. Where IT pros might have previously swapped out PCs on a schedule every few years, they’re now giving users new devices only when they need them. This requires more meticulous performance management of the hardware.

“You’re borrowing from budgets to survive…and so you gotta sweat more out of your fleet,” Bogle said. “And then most clients are taking a forward position. So when they understand what their annual forecast is, they have to lock in pricing now. Because what we know and what we’re seeing is, even with pricing fluctuations, the less expensive moment is still right now.”

Smith said that his company is running more on the infrastructure they have available, and consolidating some of their servers: “We’re running denser, and then we’re able to spread those across multiple units.”

This comes with trade-offs: more time and resources spent on operations, taking servers offline in some circumstances, and paying close attention to thermal management.

“For IT professionals who are looking at doing this within their own stack, keep in mind if you are increasing the density—not so much of the DRAM but as a secondary effect of bringing other resources together—it’s going to change the thermal profile of your rack, potentially your data center, and certainly of your server,” Smith said. “And that’s something that I’ve seen bite people across the industry as people are looking at densifying across the board.”

He, too, is seeing forward purchasing become the norm, which requires close consultation with the finance side. IT leaders are facing the question of whether to place those orders now and try to negotiate flexibility into contracts, or, as Smith said, “Do you just grin and bear it, and you accept a 300% increase now? Because realistically, in about five months, I could see it being upward of 1,000% increase. Demand is not going down and supply is getting more constrained every single day.”

Insight is also seeing much longer lead times than usual, according to Bogle, in some cases requiring IT departments to place orders quarters in advance. And where it used to be possible to place orders on a just-in-time basis, Insight is now working with clients on their 2027 needs.

“We are operating in an environment where every week is as dynamic as a quarter-end or a year-end used to be,” Bogle said.