Kai ChenIn Q2 2026, enterprise SSDs accounted for 48% of global NAND shipments, up from 26% a year earlier. At the same time, SK Hynix, projected to lead HBM4 in 2026, was expanding across DRAM, NAND, advanced packaging, flash specifications, and interconnects. The company at the front of AI’s hottest memory market was building around colder tiers.
Inference costs force operators to price every byte
OpenAI and Anthropic put inference costs at more than half of revenue in profitability projections shared with investors. At that share, model providers must assign every byte a cost. They must decide which model state, KV cache, retrieval corpus, and working context deserves scarce HBM, which can sit in DRAM, and which can wait in persistent storage without making the workflow uneconomic.
Agentic workloads sharpen those choices because they can keep working without a person waiting for each answer. Operators can trade some response latency for lower serving cost, but every additional step gives them another reason to retain context rather than reconstruct it. Model providers then spend more of the optimization budget on data placement, cache management, retrieval, reliability, and orchestration—the systems that turn an output into an action.
The shipment shift came with a revenue surge.
AI agents remain an unsettled source of demand. Vendors and customers still lack a common definition of an agent, while enterprises have largely deployed them to improve efficiency and reduce costs rather than generate new revenue. Suppliers cannot responsibly capitalize every agent demonstration as future memory demand. Even cost-cutting workflows retain more persistent state than isolated chatbot responses.
Vendors are filling the gap between HBM and SSDs
Traditional server diagrams put fast memory beside the processor and bulk storage at a polite distance. Memory vendors are now attacking the space between them.
Sandisk and SK Hynix published a High Bandwidth Flash specification for modules holding as much as 512GB and delivering between 0.4TB/s and 3.0TB/s. The specification gives system designers a NAND-derived tier with much more capacity than HBM and much more bandwidth than conventional storage. They can reserve HBM for the hottest state and move active context that cannot justify its cost into high-bandwidth flash.
Samsung is working from both ends. The company previewed zHBM, which vertically stacks HBM atop AI accelerators, and zNAND-O, which uses V-NAND. The first shortens the path between compute and HBM; the second makes NAND-derived storage more responsive.
Xcena raised $135 million at a $570 million valuation for an MX1 chip that performs data orchestration and KV-cache management inside memory modules. Micron, meanwhile, began mass production of PCIe 6.0 SSDs that reach 28GB/s, twice the speed of PCIe 5.0 drives.
Sandisk, SK Hynix, Samsung, Xcena, and Micron have chosen different insertion points, from vertical stacking to flash interfaces to cache logic inside the module. Their designs converge on the same architectural seam: accelerators lose economic value when systems cannot feed them the right state at the right price.
SK Hynix is buying options across the hierarchy
SK Hynix completed HBM4 development in 2025, and Nvidia allocated about 70% of its 2026 HBM4 demand to the company. Counterpoint estimated that SK Hynix would hold 54% of the global HBM4 market in 2026. Nvidia’s allocation and that projected share give SK Hynix an inside track with accelerator customers as it expands into adjacent tiers.
The company has committed $38 billion to new Korean capacity, including a roughly $24.7 billion DRAM facility in Yongin and a roughly $13.3 billion NAND fab in Cheongju. SK Hynix also broke ground on a $4 billion advanced-memory packaging facility in Indiana, where next-generation HBM mass production is scheduled for the second half of 2029.
SK Hynix Ventures adds another route into the surrounding stack. The investment arm covers AI computing, data centers, and optical interconnect technology—areas that determine how memory connects to accelerators and how quickly data moves across a cluster. Solidigm, already an SK Hynix subsidiary, gives the parent company an operating enterprise-storage business.
Solidigm is reportedly weighing a U.S. NAND flash factory, while SK Hynix is reportedly discussing memory production in Ohio with Intel. Either project would extend SK Hynix’s U.S. footprint beyond advanced packaging. SK Hynix said it had not confirmed a plan arising from the Intel talks. A rumor is not capex, despite the market’s recurring efforts to depreciate it on schedule.
CXMT is reportedly preparing to enter NAND from DRAM amid memory shortages, challenging SK Hynix, Samsung, and YMTC.
Long fab cycles force bets before demand settles
Kioxia and Sandisk plan to spend more than $31 billion in Japan through 2032, including $11.3 billion on a Kitakami plant. Micron expects its 2026 capital expenditures to exceed $25 billion. SK Hynix tied its Indiana packaging operation to a 2029 production schedule. These companies face a semiconductor capacity lag: model demand can change in months, while fabs, packaging lines, power systems, and supplier networks require years.
Fab operators must also secure electricity on terms that survive an entire cycle. KEPCO asked Samsung and SK Hynix to prepay approximately $18.7 billion for power supplied to planned chip clusters. Samsung and SK Hynix rejected the proposal, citing uncertainty about long-term chip demand.
By rejecting the prepayment, Samsung and SK Hynix set a limit on their own bullishness. Both companies are spending heavily, but neither considered demand certain enough to underwrite decades of power costs upfront. Memory manufacturers still remember what a commodity reversal looks like: NAND flash revenue fell 45% year over year in the fourth quarter of 2022. AI buyers do not repeal commodity cycles by ordering more drives.
Governments and suppliers nevertheless face the same physical constraint. An inference hierarchy needs wafers, packaging, power, and interconnects before operators can optimize data placement in software. SK Hynix, Kioxia, Sandisk, and Micron are making those choices years before customers can reveal how much capacity each tier will absorb.
Software determines which capacity earns a premium
SK Hynix’s projected HBM4 lead contrasts with its position in NAND. Samsung held 25% of global NAND shipments in Q2 2026, SK Hynix 22%, and YMTC 14%. Samsung has also made initial commercial HBM4 shipments, keeping the two companies in competition at both ends of the hierarchy.
SK Hynix gains system-level leverage only if model builders can use its tiers together. Hardware vendors must provide interfaces that move context predictably, controllers that place data without wasting accelerator cycles, and packaging that scales. Software must expose cost and latency trade-offs before added capacity turns into inventory.
AI operators now judge HBM and enterprise flash by the cost of keeping accelerators fed. Enterprise SSDs’ rise from 26% to 48% of NAND shipments makes that shift visible—and explains why SK Hynix’s HBM lead is pulling the company toward colder tiers.