Vespertec warns AI growth will hit hard data centre limits

Vespertec has forecast that physical infrastructure limits will shape the next stage of artificial intelligence deployments in 2026, with constraints around chip ecosystems, private cloud adoption, memory supply and data-centre cooling set to influence procurement and design decisions.

The Manchester-based data-centre infrastructure specialist said firms will face a market where access to accelerators and associated software remains concentrated, while shortages in DRAM and NAND storage affect server builds and expansion plans.

"2025 has been the year where AI hype meets with reality," said Phil Kaye, Co-Founder and Director, Vespertec.

Chip ecosystems

Vespertec expects hyperscalers to continue investing in in-house silicon. It pointed to Google expanding TPU deployments and said Meta has evaluated the use of Google's TPU platform.

It does not expect those efforts to weaken Nvidia's influence for most enterprises and service providers that buy and deploy systems outside hyperscaler environments. The company said organisations tend to favour platforms that fit existing workflows and integrate across mixed vendor estates.

Vespertec highlighted software tooling and developer familiarity as factors that affect buying decisions. It said established stacks reduce friction when teams deploy new infrastructure alongside existing systems.

The prediction matters for IT buyers that want optionality on accelerator hardware. It also matters for data-centre operators that need to plan power, space and networking around the types of servers that customers can acquire in volume.

Private cloud

Vespertec also expects a renewed focus on private cloud for AI deployments in 2026. It referenced AWS and Nvidia introducing Private AI Factory as a signal that large providers see demand for environments that customers control more directly.

The company said firms want closer control over workload behaviour and performance governance. It linked that preference to the need for predictable scaling and more consistent outcomes as AI workloads move from pilot programmes into production use.

Private cloud in this context does not exclude public cloud usage. It changes where some workloads run and where data sits, especially for organisations that weigh latency, governance and operational predictability alongside raw compute availability.

Memory shortages

Vespertec warned of a tightening supply of DRAM, with DDR5 facing a "severe shortage cycle". It said fabrication capacity has shifted towards high-bandwidth memory, including HBM3, for GPU production.

That shift, it said, reduces the output of traditional RDIMMs at the same time as AI server demand increases. It also said large buyers could absorb a significant share of available output.

"DRAM, and especially DDR5, is heading into a severe shortage cycle," said Kaye.

The company expects buyers to face allocation-based supply and higher prices. It also expects procurement lead times to become a more prominent planning factor for IT teams that have relied on late-stage upgrades and rapid expansions.

Vespertec made a similar point on NAND flash. It said demand for GPU-dense servers increases the need for high-capacity SSDs in data-heavy AI deployments. It expects ongoing price firming and gradual increases through 2026.

Cooling choices

Vespertec said cooling will move into a more selective phase. It expects immersion cooling to mature beyond proof-of-concept deployments, but it does not see it becoming mainstream in 2026.

It said direct liquid cooling will remain the primary approach for the most powerful AI systems. It cited compatibility limits for immersion in some configurations and said certain high-performance platforms still depend on direct liquid cooling.

It also pointed to increasing rack densities. It said deployments that exceed 70kW per rack raise airflow management challenges and place pressure on conventional cooling designs.

Vespertec said more vendors now support immersion where it fits, and it expects more interoperability testing and more developed tooling and service models. It expects that shift to make immersion a more common topic in design discussions, even if most deployments still use other approaches.

"Immersion will not be mainstream in 2026, but it will feature more often in design discussions as organisations acknowledge that traditional cooling approaches will not always match future compute ambitions," said Kaye.