Memory shortages, cost unpredictability, and the case for compute-as-a-service in enterprise AI

At Lenovo Tech World Hong Kong, conversations around what the current hardware constraints actually mean for enterprise AI buyers, and what vendors can realistically offer dominated discussions.

With RAM shortage and pricing an increasing concern globally, the topic was definitely on the agenda at Lenovo Tech World Hong Kong. While the focus was on AI, it was inevitable that the memory shortage discussion was part of the discussion among attendees as most of them now find themselves caught between rising component costs and constrained supply, as cloud providers hoover up available capacity

Beneath a lot of the confident AI investment data, the gap between what enterprises want to deploy and what the hardware supply chain can actually deliver to them, at a price they can budget for was discussed at a roundtable session during the summit in Hong Kong.

Linda Yao, vice president and general manager for hybrid cloud and AI Solutions at Lenovo, addressed the concerns without deflection. "We have a lot of enterprise customers coming to us with a similar problem," she said. "Can I do a buy-ahead to secure my supply? Can I do a price lock so I can have better cost predictability? When we distill down how the memory shortage impacts enterprise buyers, it comes down to two things: cost predictability and supply visibility."

Yao further added that it is not just about the cost. “It's more about cost predictability as they plan their budgets and number two is supply visibility.”

Lenovo's response, as Yao described it, sits within what the company calls its TruScale model–compute delivered as a service, with the supply chain visibility and financial flexibility that outright hardware purchasing increasingly cannot guarantee.

For enterprise customers not yet ready to deploy but wanting to lock in pricing and availability, Lenovo is offering what amounts to a forward contract on hardware: a price lock, with deployment flexibility built in for customers who need six to twelve months before they are ready to roll out.

The asset recovery angle was also notable. Yao described how Lenovo is helping customers offset the upfront cost of new AI-capable infrastructure by crediting the value of existing devices being replaced–effectively allowing organizations to partially self-fund a hardware refresh through the depreciated value of what they already own.

"We can offer them the value of that device on credit," she said. "It's a combination of different tools in our toolkit to help our customers address the root cause of this industry-wide issue."

For industry players, the conversation points to a services opportunity that goes beyond the hardware transaction itself. The memory cycle that Yao described is not unique to any one vendor, it is a market-wide dynamic that creates planning complexity for IT buyers across the region.

Partners positioned to help customers model infrastructure requirements, structure phased deployment timelines, and navigate the trade-offs between on-device and cloud-based AI inference are in a position to add substantial value that a straightforward hardware sale does not capture.

Art Hu framed the underlying trade-off clearly. The most capable AI models currently live on the public cloud, but they are also the most expensive. As smaller, distilled models improve rapidly, running credibly on devices with a few billion parameters–the calculus shifts. "Every time you charge this device or run it, there's no incremental cost," Hu said.

"As Lenovo's CIO, I want to model what is the right task to run at the right model. I want the right combination of cost, latency and security." For many enterprise workloads, that combination increasingly favors the edge or on-premises infrastructure over a cloud call, which is precisely where Lenovo's hybrid architecture argument finds its most commercially grounded footing,” Hu added.

The memory shortage will ease as supply cycles normalize. But the underlying dynamic–cost predictability, supply visibility, and the need for flexible consumption models in enterprise AI infrastructure–is structural, not cyclical.

For vendors and partners that can build planning and financing capability around those constraints, the current period of hardware tightness is as much an opportunity as it is a challenge.