Jensen Huang hints NVIDIA will be first major customer for HBM4 as next-gen AI chips move toward production

NVIDIA’s push into next-generation AI hardware is drawing attention not just for what it will put on its own chips, but for the memory that must sit alongside them. In a market report published July 18, Jensen Huang suggested that NVIDIA will be the first major customer for HBM4, the next generation of high-bandwidth memory designed to provide much higher data throughput for AI training and inference.

The report also claims that orders for HBM4 have been “locked up,” with a figure of roughly 70% cited for the share of orders attributable to the first wave of customers. While the exact sourcing and who those orders are with is not detailed in the report itself, the takeaway is that early demand for HBM4 appears to be pre-negotiated before broad rollout, reflecting the criticality of memory bandwidth in modern AI systems.

HBM, or high-bandwidth memory, is a specialized memory architecture stacked close to the logic that needs it. Compared with more conventional memory approaches, HBM is built to deliver far higher bandwidth, which matters because AI models increasingly bottleneck on moving data to and from the compute units. HBM4 is being positioned as the next step in that progression, which is why securing supply for it is likely to be treated as strategically important by companies producing accelerators.

For NVIDIA, the memory question is inseparable from the company’s platform approach. NVIDIA sells AI accelerators and the software stack around them, and in practice customers measure system performance end-to-end, not just on compute speed. If memory capacity or bandwidth lags, total throughput can stall even if the chip itself is ready, making the timing of HBM4 availability a real operational constraint.

The report frames Huang’s comments around NVIDIA’s upcoming chip launches, implying that next-gen NVIDIA platforms are aligned with the arrival of HBM4. In that framing, being an early customer would allow NVIDIA to validate designs, manage qualification timelines, and reduce risk that memory availability could limit shipments.

NVIDIA has long benefited from the broader “AI infrastructure” buildout, but memory supply can introduce a different kind of pressure than chip availability. HBM is a complex product, and the amount of memory required per system scales quickly as data sizes grow and as customers demand faster iteration cycles. That is why early allocation and long-lead procurement for HBM generations can become a headline issue when companies move to new acceleration generations.

Still, key details are not disclosed in the report. It does not specify the exact terms of Huang’s remarks, the identities of the memory suppliers referenced, which specific NVIDIA products will use HBM4, or whether the “locked up” figure relates to planned capacity, firm purchase orders, or forecast commitments. Without those specifics, the “70%” figure should be treated as a reported industry estimate rather than a confirmed accounting number.

Going forward, investors and customers will likely watch for confirmations on HBM4-related supply commitments, any mentions of HBM4 in NVIDIA product disclosures, and whether the memory ramp keeps pace with NVIDIA’s accelerator shipment expectations. Because HBM4 is central to system-level performance, even small timing shifts could ripple through AI server production schedules and lead to further market speculation about future availability.