SK hynix and SanDisk Just Opened a $1 Trillion Memory Race-HBM May Not Rule AI Forever

HBF matters because AI still has a memory wall to clear

The market is still focused on HBM. The more interesting development is what sits beside it: a potential new memory tier for AI inference.

AI data centers are already running into the memory wall: memory has emerged as one of the biggest bottlenecks in AI infrastructure, and the memory shortage could last until 2030. That keeps HBM critical, but it also creates the opening for HBF.

Why this mattered

The signal was not a product launch. It was a specification release. Sandisk and SK hynix advanced the HBF spec through the Open Compute Project, about six months after the consortium began work. Google and Tenstorrent also joined the effort, which makes the project look more like an industry architecture discussion than a single-vendor pitch.

HBF is positioned as a new memory layer between HBM and SSD. It is not trying to replace HBM outright. The goal is more practical: give AI inference systems more capacity closer to the compute core, while improving bandwidth, power efficiency, and total cost of ownership.

HBF does not need to beat HBM at speed

What matters now is whether HBF can be good enough for inference, not whether it can outspeed HBM. The latest update is that HBF is moving from concept toward adoptable standardization. Sandisk and SK hynix released the HBF technical specification through the Open Compute Project, and Google and Tenstorrent joined as consortium members. That makes it closer to a platform effort than a vendor demo.

Why standardization matters

When memory is proprietary, customers face lock-in and integration risk. An open spec gives designers a shared framework for adding HBF into AI inference systems and accelerators. In practical terms, that should make adoption easier across multiple vendors.

This also matters because progress has been fast. This was the first HBF standard showcased within six months of consortium launch, and Sandisk has outlined a path for customer sampling targeted for 2026. For investors and engineers, that is the first bridge from narrative to proof.

Why inference is the likely starting point

HBF is aimed at the gap between HBM and SSDs. That placement fits inference, where systems need more near-compute capacity and higher bandwidth without adding too much power or cost. SK hynix also highlighted a 2.5x improvement in power efficiency with its tenth-generation 375-layer 4D NAND, which matters because any new memory tier near the GPU has to fit inside a tight thermal and power budget.

If that positioning holds in real systems, the payoff is straightforward: system builders get another way to expand near-compute memory without making every extra gigabyte a power and cost problem.

What could slow adoption

There are still real engineering questions. 3D NAND's limited endurance could matter if flash is asked to behave like a memory tier over long write cycles. Software also may prove harder to adapt than proponents expect, especially when systems have to manage different memory tiers on one device.

What could be mispriced if HBF gains traction

That moves the story from architecture debate to portfolio positioning.

The mid-layer opportunity

The market is still focused on HBM scarcity, but the next opportunity may sit in the mid-layer. HBF is being built as a new memory layer between HBM and SSD, and the open standardization path through OCP should make that layer easier for others to adopt. If that happens, demand could emerge before the revenue impact is obvious in financial statements.

SanDisk stands out for optionality. It has customer sampling targeted for 2026, and the work is already spanning NAND design, ASIC design, and packaging. If HBF becomes a real inference tier, SanDisk may help define more than just the flash part of the stack.

Who to watch if the tier matures

• Tier architects as well as memory suppliers: The opportunity maps to tiered memory, so winners may be companies that can combine capacity, power efficiency, and system integration.
• Ecosystem partners: Google and Tenstorrent joining the consortium matters because system and accelerator designers helping shape the spec can reduce adoption friction.
• Packaging and interface developers: Memory architecture does not win on cell chemistry alone. The companies that solve interface, control, and packaging may matter just as much.

What would confirm the thesis

• Customer sampling moves from roadmap to real customer evaluation.
• Consortium participation broadens beyond the founding members.
• Tiered-memory language starts showing up in platform designs, not just event presentations.
• OCP standardization keeps expanding the potential buyer base.

What could invalidate the thesis

If sampling slips well beyond current plans, or if 3D NAND's limited endurance becomes a system blocker rather than an engineering challenge, HBF may remain niche. That would not end the HBM story. It would simply mean the tiered-memory market is taking longer to form.