Nvidia First Customer for HBM4 as Memory Stock Reportedly Locks 70%

Nvidia will be the first customer for HBM4, chief executive Jensen Huang has said, according to reports from Yahoo Finance and The Motley Fool — and a single AI memory stock has reportedly locked up roughly 70% of those next-generation memory orders. The claim extends a theme Huang established at CES 2026, where he told the audience that memory has become the biggest bottleneck in AI. Since those remarks, Yahoo Finance notes, memory specialists Micron and SanDisk have outperformed Nvidia’s own shares — a striking reversal in a market that has long treated the GPU designer as the default AI trade.

Jensen Huang says Nvidia will be the first customer for HBM4

Both Yahoo Finance and The Motley Fool report that Huang has said Nvidia will be the first customer for HBM4, the next generation of high-bandwidth memory that will feed the company’s future AI accelerators. The reports pair that remark with a second claim: one AI memory supplier has reportedly locked up around 70% of those orders.

The material reviewed for this article does not include order volumes, contract values or delivery schedules, so the scale of the commitment remains unconfirmed. Even so, the direction of travel is clear. Nvidia’s accelerator roadmap is increasingly gated by how much memory bandwidth and capacity it can attach to each GPU, and being first in the queue for HBM4 is a statement about where the company believes its next competitive edge will come from. For a business whose products already command extraordinary demand, securing the newest memory generation before rivals matters as much as any architectural improvement on the GPU die itself.

Memory is now the biggest bottleneck in AI, Huang told CES 2026

The HBM4 comments build on the message Huang delivered at CES 2026, where — as Yahoo Finance reported — he said that memory has become the biggest bottleneck in AI. That framing matches what engineers have observed for years: raw compute has scaled faster than the bandwidth needed to keep it fed. Modern accelerators can execute enormous numbers of operations per second, but only if data arrives from memory quickly enough — and increasingly, it does not.

The constraint bites hardest in inference. Serving a large language model means holding its weights, plus a growing key-value cache for every active conversation, in fast memory close to the processor. As context windows stretch and the models tracked in our AI models database keep growing, capacity and bandwidth — not raw compute throughput — determine how many users a single GPU can serve and at what latency. Anyone who has sized a local deployment with a free VRAM calculator will recognise the pattern: the memory runs out long before the compute does. Huang’s remark simply elevated that engineering reality into an investment thesis.

The AI memory stock reportedly holding 70% of Nvidia’s HBM4 orders

Neither headline names the supplier outright, and the reports reviewed here attach no figure beyond the roughly 70% share. As industry context rather than a claim drawn from the reports themselves, the description most closely fits SK hynix. The South Korean manufacturer has been Nvidia’s principal high-bandwidth memory partner through the HBM3 and HBM3E generations, and it is widely regarded as the frontrunner for the first wave of HBM4 supply.

The wider HBM market is effectively a three-way contest between SK hynix, Samsung and Micron — the only DRAM makers with the packaging and stacking expertise the technology demands. That concentration is why a reported 70% share of the launch customer’s orders would be so significant. HBM is co-designed and qualified alongside the GPU package it attaches to, so switching suppliers mid-generation is slow and expensive; early allocation tends to persist for the life of a product cycle. Whoever dominates the first HBM4 shipments to Nvidia is well placed to dominate the generation.

Investors should nonetheless treat the 70% figure as reported rather than confirmed. Until the companies involved disclose supply agreements or volumes, the precise split remains an industry estimate circulating through financial media.

Micron and SanDisk have outperformed Nvidia stock since CES 2026

Yahoo Finance reports that Micron and SanDisk have both outperformed Nvidia’s stock since Huang’s CES 2026 comments. The report does not quantify the gap, and past performance is no guide to what comes next, but the rotation is telling. Nvidia stock has been the market’s default expression of AI enthusiasm for three years; the fact that memory suppliers have led it since January suggests investors now see scarcity — and therefore pricing power — migrating down the supply chain from GPU design to the components that feed it.

The two outperformers sit in different corners of the memory market, which reinforces the breadth of the trend. Micron is one of the three DRAM manufacturers capable of producing HBM, while SanDisk specialises in NAND flash storage. The AI build-out, in other words, is straining several categories of memory at once: the high-bandwidth stacks bonded directly to accelerators, the conventional DRAM in servers, and the storage tiers holding training data and model checkpoints.

What HBM4 changes for next-generation AI accelerators

High-bandwidth memory stacks DRAM dies vertically and connects them to a processor over an exceptionally wide interface, delivering far more bandwidth than conventional memory modules. As general industry background — these figures come from published JEDEC standards rather than from the cited reports — HBM4 doubles the per-stack interface width to 2,048 bits, compared with 1,024 bits for HBM3 and HBM3E. A wider interface means substantially more bandwidth per stack, which translates directly into faster training steps and higher inference throughput.

Memory configuration is already the headline differentiator between accelerator cards — it is the first specification we weigh in our guide to the best GPUs for AI — and HBM4 will widen that gap further. If memory is the bottleneck, as Huang argues, then the accelerators that ship with HBM4 first will hold a practical advantage that raw compute comparisons understate.

What the HBM4 supply race means for AI builders and buyers

For developers and infrastructure buyers, the immediate implication is about supply rather than specifications. If one supplier really has locked up around 70% of the first customer’s orders, HBM4 capacity will be tight at launch and priced accordingly, and that memory cost will flow through to accelerator prices and, ultimately, cloud GPU rates. Those economics land on anyone paying for tokens or renting compute — a dynamic our AI price-performance index tracks across providers.

Tighter memory supply also shifts the buy-versus-rent calculus for teams weighing their own hardware, the trade-off captured in our self-hosting vs API calculator. When next-generation accelerators are scarce, waiting lists lengthen and secondary prices rise, strengthening the case for renting capacity in the short term. None of this is confirmed detail — the story remains a reported one — but the direction is consistent with Huang’s own framing: the scarce resource in AI is no longer the processor, it is the memory attached to it.

Frequently asked questions

What did Jensen Huang say about HBM4? According to Yahoo Finance and The Motley Fool, Huang said Nvidia will be the first customer for HBM4, the next generation of high-bandwidth memory. The reports reviewed did not include timing, volumes or pricing.

Which company has reportedly locked up 70% of Nvidia’s HBM4 orders? The reports centre on a single AI memory stock without naming it in the headlines reviewed. As industry context, the description most closely fits SK hynix, Nvidia’s lead HBM partner through the HBM3 and HBM3E generations — though the 70% figure remains reported rather than confirmed.

Why is memory the biggest bottleneck in AI? Compute performance has scaled faster than memory bandwidth and capacity. Serving large models requires holding weights and growing key-value caches in fast memory, so memory — not raw compute — increasingly limits throughput and latency.

What is HBM4? HBM4 is the next JEDEC-standard generation of high-bandwidth memory. It doubles the per-stack interface width to 2,048 bits versus HBM3E, delivering significantly more bandwidth for AI accelerators.

Have memory stocks outperformed Nvidia stock in 2026? Yahoo Finance reports that Micron and SanDisk have outperformed Nvidia’s shares since Huang’s CES 2026 remarks about the memory bottleneck. The report did not quantify the difference.

The bottom line

The reported story is straightforward: Nvidia intends to be the first customer for HBM4, and one memory supplier has reportedly captured about 70% of those orders. The significance is larger than either headline. Huang has spent 2026 arguing that memory, not compute, is the binding constraint on AI — and the market has responded by re-rating the companies that make it, with Micron and SanDisk outpacing Nvidia’s stock since CES, per Yahoo Finance. What to watch next is confirmation: formal supply agreements, capacity disclosures on earnings calls, and the first HBM4-equipped accelerators reaching customers. Until then, the 70% figure should be read as reported, not settled — but the shift in where AI’s scarcity sits looks increasingly real.

Sources: news.google.com. Reported July 19, 2026.