1 week ago
45
Hewlett Packard Enterprise has a problem most companies would love to have: too many customers, not enough parts. The company exited fiscal Q3 2026 with a $7.6 billion AI backlog, up 14% from the prior quarter, driven by enterprise and government clients racing to deploy AI infrastructure at production scale.
The bottleneck isn’t demand. It’s DDR5 DRAM and NAND flash memory, the critical components that power the high-bandwidth memory modules AI servers devour. Those supply constraints are expected to persist into 2027.
The numbers behind the surge
HPE’s quarterly revenue hit $12.2 billion, a 34% jump year-over-year. AI systems orders alone reached $2.4 billion for the quarter, climbing more than 30% sequentially.
Networking wasn’t far behind. Orders in that segment rose 36% on a normalized basis, with AI-specific networking orders totaling $700 million in Q3. That figure got a boost from HPE’s acquisition of Juniper Networks, which plugged a gap in the company’s networking portfolio at exactly the right moment.
To keep pace with the incoming orders, HPE has been stockpiling. Inventory swelled to $11.82 billion, a deliberate move to buffer against the memory shortage and ensure the company can convert backlog into revenue as components become available.
Perhaps the most eye-catching number came after the quarter closed: HPE announced a $3.5 billion inferencing agreement with a hyperscaler. By mid-2026, cumulative AI systems bookings had crossed $16 billion. The company responded by raising its full-year revenue growth guidance.
Who’s buying, and why it matters
The composition of HPE’s order book tells a more interesting story than its size. Over 60% of cumulative AI orders have come from enterprises and sovereign clients, not the hyperscale cloud giants that dominated early AI infrastructure spending.
Sovereign AI, where national governments build domestic AI compute capacity for security and data residency reasons, has become a particularly active segment. HPE’s hybrid cloud heritage and its relationships with government IT departments give it an advantage that pure hyperscale-focused competitors can’t easily replicate.
The memory problem isn’t going away soon
DDR5 DRAM is the choke point. AI servers require significantly more memory per unit than traditional servers, and the transition from DDR4 to DDR5 has strained manufacturing capacity across the semiconductor industry. NAND flash faces similar pressure as AI workloads demand faster storage for model training and inference.
HPE’s management has responded with two levers. First, multi-year supply agreements with memory manufacturers, essentially locking in capacity before competitors can claim it. Second, pricing adjustments that pass some of the increased component costs through to customers.
The company expects quicker backlog conversions in the coming quarter, though that’s contingent on component availability rather than any change in customer appetite.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·