Mapping the Memory Supply Chain

Who produces each layer, who genuinely belongs to the agentic thesis, and who merely resembles it

Part two of two. Part one, "Memory Is the New Compute," develops the thesis and its counterarguments.

In the first part of this series I argued that the binding constraint on AI infrastructure has shifted from compute to memory, and that an agent's memory is not one thing but three layers: the KV-cache in HBM, working DRAM, and persistent storage.

That framework has a practical use beyond the technical: each layer corresponds to a distinct set of manufacturers. Anyone thinking about this thesis solely in terms of Micron, SK hynix and Samsung is leaving out most of the map — and, in some cases, the least crowded segments.

| Layer | Who produces it | |---|---| | KV-cache in HBM | SK hynix, Micron, Samsung; TSMC (base die); BESI, Hanmi, ASMPT (stacking) | | Working DRAM | The same three; Nanya and Winbond in legacy DRAM; Rambus and Montage in the interface | | Persistent storage | SanDisk, Kioxia (NAND); Seagate, Western Digital (HDD); the software layer |

What follows is a link-by-link walk through the chain, with two warnings at the end worth reading before acting on any of these names.

1. NAND and persistent storage

SanDisk (SNDK) expanded its non-GAAP gross margin by 5,570 basis points in a single year, to 78.4%. It is also SK hynix's partner on HBF (High Bandwidth Flash), whose first open specification was published in August through the Open Compute Project: up to 512 GB per stack, 0.4 to 3.0 TB/s, using the standard UCIe chiplet interface. If HBF takes hold, it creates a new tier between HBM and SSD — precisely the gap that today forces agents to truncate context.

Kioxia (Tokyo: 285A) co-operates the Yokkaichi and Kitakami plants with SanDisk, the world's largest NAND capacity, with the joint venture extended through 2034. It reported ¥2.34 trillion in fiscal 2025 revenue, up 37% year over year, and is developing SSDs with Nvidia that would connect directly to the GPU.

In hard drives, Seagate (STX) and Western Digital (WDC) benefit from a counterintuitive effect. During the NAND price spike, the cost-per-terabyte gap between a 30 TB enterprise SSD and an equivalent nearline hard drive widened from roughly 6x to roughly 16x. Seagate estimates that replacing enterprise HDD demand with NAND by 2028 would require around $240bn of additional capital expenditure and some 35 new fabs, against roughly $1bn on the hard drive side. IDC projects that close to 80% of hyperscale storage will still sit on hard drives through 2028.

The general point: an agent's long-term memory does not live in HBM. It lives here, and this link trades on very different multiples.

2. Legacy DRAM: the gap the majors left behind

Nanya (TW: 2408) and Winbond (TW: 2344) are capturing the space the big three are vacating as they reallocate wafers toward HBM and DDR5. DDR4's share of wafer starts at Samsung and SK hynix is expected to fall to low single digits in the second half of 2026, and legacy memory prices have surged accordingly.

Nanya grew 690% year over year in the second quarter of 2026 and consolidated its position as the world's largest DDR4 supplier; Qualcomm named it among its first 35 partners for its AI data center push. Winbond posted NT$98.1bn in first-half revenue, up 139% year over year, already exceeding its full-year 2025 total.

CXMT in China grew 716% year over year — the fastest-growing DRAM supplier globally — and listed in Shanghai to fund expansion into HBM and LPDDR6. Here it is worth being blunt: for a Western investor, CXMT is more threat than opportunity. It is the mechanism by which this cycle most likely ends.

3. Equipment and packaging: the bottleneck inside the bottleneck

If HBM is the scarce link, stacking is the scarce link within the scarce link. A 16-layer HBM4 stack costs on the order of $20,000, and a 100-nanometer misalignment destroys it entirely. Production yield, not wafer capacity, is what constrains supply.

BESI leads in hybrid bonding — the transition from microbumps to direct copper-to-copper joining — and is expected to be Micron's sole TCB bonder vendor for HBM4, displacing the suppliers used in earlier generations. It has also flagged a significant increase in hybrid bonding system orders for advanced logic and HBM4 applications as customers execute their 2026-2027 roadmaps.

ASMPT (0522 HK) and Hanmi Semiconductor (042700 KR) compete for the same ground. Hanmi signed a 44.2 billion won contract with SK hynix in June for its TC Bonder 4.5 Griffin, designed specifically for HBM4 — equivalent to roughly 7.6% of its 2025 revenue, for equipment destined for the Cheongju facility.

TSMC deserves separate mention because almost no one counts it as memory exposure: it fabricates the HBM4 base die, with 3nm confirmed for HBM4E, alongside the CoWoS packaging that integrates the assembly.

4. Interface and controllers

Astera Labs (ALAB) designs PCIe, CXL and Ethernet connectivity semiconductors for rack-scale AI infrastructure — the plumbing linking GPUs, CPUs and memory. First quarter 2026: record revenue of $308.4m, up 93% year over year and 14% sequentially, with a 76.3% GAAP gross margin. The offset is valuation: it trades at multiples that leave no room for disappointment.

Rambus (RMBS) sells memory interface chips to all three DRAM manufacturers and develops HBM4E controller intellectual property and the SOCAMM2 chipset for LPDDR5X modules in AI servers. Its chief executive has explicitly pointed to agentic workloads as a driver of standard DDR demand, and to the shifting CPU-to-GPU ratio as inference scales.

Here, though, some skepticism about the narrative is warranted. Rambus product revenue grew 15% year over year in the first quarter and 22% in the second, against a memory interface chip market growing at roughly 29.5% annually. It is losing share to Montage Technology and Renesas, and trades at a premium to larger peers. Being in the right sector is not the same as winning in it.

5. The software layer: a real option, not a position

There is one further avenue: persistent memory infrastructure proper. Vector databases, context layers and KV-cache offload systems, with names such as Pinecone, Weaviate, Zilliz, MongoDB, Elastic, Redis and WEKA, plus specialized startups like Mem0, Zep and Letta.

It is worth being honest about this layer: it holds the greatest theoretical upside and the weakest evidence. Most of these companies are private. The market sizes in circulation come from research aggregators of widely varying quality — some publish internally inconsistent series. And there is still no technical consensus on which agent memory architecture will prevail.

It is a venture capital thesis wearing public-market clothes. For a public portfolio, the clean exposure remains in hardware; the software layer is, for now, an option to monitor rather than an allocation to size.

Two warnings before acting

First: not everything that is rising is the thesis. Nanya, Winbond and CXMT are rising because the big three abandoned DDR3 and DDR4, not because agents need their chips. It is a legacy memory price trade — real, but different, more fragile, and with a faster reversion mechanism. Confusing it with agentic inference exposure is a category error that shows up in the account when the cycle turns.

Second: the further down the chain, the higher the beta. Rambus fell roughly 28% in a month after peaking in June. The equipment and legacy DRAM names trade in Korea, Hong Kong and Taipei, with the access, liquidity, currency and governance frictions that implies for a US portfolio. Indirect exposure through diversified semiconductor vehicles rarely captures these names in meaningful proportion — but neither does it impose their volatility.

Conclusion

The value of thinking about this thesis in layers is not collecting names. It is recognizing that "memory" bundles together businesses with very different economics, cycles and risks: a three-player oligopoly in DRAM, a six-player scrum in NAND, a duopoly in hard drives, a handful of equipment suppliers with pricing power over production yield, and a software layer with no established winner.

The useful question is not "should I take memory exposure?" It is "which layer, on what economics, and at what entry price?" A good deal of the last twelve months' return has already been paid in the most obvious layer.

This content is for informational and educational purposes only. It does not constitute investment advice, an offer, or a recommendation to buy or sell any security, and it does not take into account the objectives, financial situation or particular needs of any investor. Investing in semiconductor equities involves significant risks, including possible loss of principal, and volatility substantially above that of the broad market. Investing in foreign securities adds currency, liquidity, information and corporate governance risks. References to specific companies are illustrative and do not constitute recommendations. Past performance does not guarantee future results.

Sources

• Kioxia and Sandisk — 9th and 10th generation NAND technology announcements (August 2026) and the HBF open specification via the Open Compute Project

• TrendForce — Second-Tier No More: Kioxia and SanDisk in the AI NAND Race, and coverage of the Taiwanese memory market (July 2026)

• Counterpoint Research — AI Demand Reshapes DRAM Rankings in Q2 2026

• Seagate and IDC — Cost-per-terabyte estimates and hard drive share of hyperscale storage

• Hanmi Semiconductor / SK hynix — TC Bonder 4.5 Griffin contract (June 2026)

• BESI — Commentary on hybrid bonding system orders for HBM4

• Rambus — Q1 and Q2 2026 results and SEC Form 10-Q

• Astera Labs — Q1 2026 results