The Memory Mirror: Two Worlds for HBM in 2029
Memory was the boring chapter of every 2023 semiconductor primer. In 2026, it is the chapter that decides whether the AI buildout actually works
In November of 2023, Eric Flaningam published a primer on the semiconductor industry that did something rare. It made the most complex industry on the planet feel legible. I keep coming back to it. Not because it was prescient about everything, but because of what it didn’t emphasize. Read it today and the memory section is almost an afterthought. Three big players. A cyclical market. A huge drop in revenue over the course of 2022. Two and a half years later, that section is the one that aged the most.
In 2023, memory was the boring chapter. In 2026, it is the chapter that decides whether the AI buildout actually pencils out.
Every memory cycle I have sat through in twenty-plus years of procurement has had the same shape. A setup phase nobody pays attention to. A peak phase where the narrative hardens into “this time is different.” And a correction phase that humbles whoever bought the narrative. The question worth asking, for anyone building product, sourcing components, or allocating capital, is which phase we are actually in right now, and what the next three years look like from here.
So let me sketch two worlds. Both are plausible. Both have receipts. The truth, I suspect, will be a blend, but the gap between them is wide enough to be worth living inside for a few thousand words.
Here is what the data actually looks like in May 2026.
SK Hynix overtook Samsung in annual profit for the first time ever, on the back of HBM3E leadership. SK Hynix and Micron joined the trillion-dollar market cap club last week. HBM now represents more than half of DRAM sales at the leading memory companies. DRAM prices are up 171% year-over-year. DDR5 spot prices have quadrupled since September. Samsung and SK Hynix have already raised 2026 HBM3E contract prices.
This is not the memory market Eric described. The cycle didn’t disappear. It got rewritten. Every gigabyte of HBM consumes roughly three times the wafer capacity of DDR5, which means the AI buildout is structurally cannibalizing the supply that used to keep your laptop, your car, and your phone affordable. PC prices are projected to rise 15 to 20% this quarter. Automakers are quietly warning of production disruptions on 2028 and 2029 model years as legacy DRAM production contracts. The memory market has become the binding constraint on the entire AI economy, and almost no one outside of supply chain saw it coming.
Now the two worlds.
World One: memory becomes the new oil.
In this world, HBM is not a chip. It is an infrastructure asset, priced and traded more like a long-cycle commodity with structural scarcity baked in. Three years from now, in May 2029, the memory market looks like this.
The three incumbents have held discipline. Yongin in Korea, Idaho in the US, and SK Hynix’s M15X have all come online, but they came online slowly and were absorbed by demand on contact. HBM4 and HBM4E are sold out through 2030. Margins haven’t collapsed. The supply curve never overshot because the leadership at all three companies remembers 2022, and remembers it well. They build for orders, not for forecasts.
The demand side keeps surprising to the upside. Inference, which was projected to be two-thirds of AI compute spend by 2026, ends up being closer to 80% by 2029. And inference is more memory-bound than training, not less. Custom ASICs from Broadcom and Marvell, Google TPUs, AWS Trainium, Microsoft Maia all stack HBM at higher ratios than Nvidia’s GPUs do. Sovereign AI buildouts in the Gulf, India, Europe, and Southeast Asia layer in on top of the hyperscaler base.
In this world, memory is the most strategically protected piece of the stack. The Korean government treats SK Hynix and Samsung the way Taiwan treats TSMC. Micron’s Idaho fab becomes the centerpiece of the second CHIPS Act. HBM is added to dual-use export control regimes. The companies that figured this out early, the OEMs who locked in three- and five-year supply agreements in 2025 and 2026, look like geniuses. The ones who treated memory as a commodity to be sourced on the spot market are gasping for air.
The big winner, oddly, is Micron. The patent data already hints at this. Micron filed 621 HBM-related patents between 2018 and 2026, nearly double SK Hynix’s 315. They were the late entrant who out-engineered the field, and in the optimistic world, that R&D investment compounds. They go from 10 to 21% market share to closer to 30%. America finally has a domestic memory champion that matters at the leading edge.
The deepest read in this world is that memory stops being cyclical. Not because cycles are abolished, but because the demand floor is set by AI infrastructure that has the same multi-decade visibility as electrical grid buildout. You don’t unbuild a hyperscale data center. You don’t unbuy the GPUs and ASICs already in racks. The memory underneath them needs to be refreshed, expanded, and upgraded forever. HBM becomes a subscription business in everything but name.
World Two: 2008 all over again, just with stacks.
In this world, the same forces that lifted everyone in 2025 and 2026 reverse hard, and the memory industry runs the same script it has run four times in the past twenty years. Three years from now, in May 2029, the market looks like this.
Samsung’s management, the ones already publicly worrying about a 2028 downturn, turn out to have been right but too cautious. Every memory company on earth, plus Chinese state-backed CXMT and YMTC, plus the Yongin cluster, plus three new Micron fabs, all came online inside an 18-month window between late 2027 and early 2029. Fab construction takes about two years. Everyone started building in 2025 and 2026. The capacity wall hits in 2028.
Simultaneously, AI demand growth doesn’t go negative. It just decelerates. The hyperscaler capex curve was running at plus 83% in 2026, eating 100% of their cash flows from operations. That number cannot stay there. Goldman’s own analysis shows the hyperscalers cut buybacks by two-thirds to fund AI, and there is a limit to how long you can do that before activist investors and ratings agencies take notice. Frontier model economics finally get scrutinized. The agentic AI revenue case turns out to be slower and messier than the narrative suggested.
NAND goes first, because NAND was always more fragmented. Kioxia and YMTC undercut everyone. NAND prices fall 40% in a single quarter in early 2028, taking Western Digital and SanDisk down with them. DRAM follows six months later. HBM holds longer because of technical complexity and packaging constraints, but by mid-2029, even HBM contract prices are off 25% from peak and falling.
In this world, the three-year hangover from 2026’s HBM gold rush looks a lot like the 2009 to 2012 DRAM bust. Marginal players exit. Samsung’s foundry business, already wounded by HBM market share losses, gets restructured. SK Hynix takes a writedown that erases the trillion-dollar market cap milestone people are celebrating today. Micron’s stock, having climbed three-fold from 2025 to 2027, gives most of it back. The Korean economy takes the hit, and the political conversation in Seoul turns sharply protectionist.
The OEMs who locked in long-term HBM contracts at 2026 prices in this world look foolish, not visionary. The ones who stayed on the spot market and waited get to source 2029 HBM at a fraction of what their competitors paid. The lesson, again, is the one the memory industry has taught for forty years. It is a cyclical business, and structural narratives about why this time is different are usually the surest signal that this time is exactly the same.
The CSCOs who are designed for both worlds are not running a memory strategy.
They are running a direct-materials strategy where memory is one of fifty categories under the same disciplined operating model.
Here is the part most memory-cycle commentary misses. The people sourcing HBM are also sourcing everything else that sits next to it on the bill of materials. Substrates. Interposers. Power management ICs. Connectors. PCBs. Mechanicals. Passives. The HBM shortage doesn’t live in a vacuum. It bleeds into wafer allocation for the rest of the BOM, into packaging capacity that the same OSATs are running for your other components, and into supplier behavior across categories that have nothing nominally to do with memory.
Every cycle I have watched, the people who survived weren’t the ones who picked the right view. They were the ones who built supply chains that could operate in both.
If you are a CSCO or CPO reading this in May 2026, the practical question isn’t which world will happen. It is which world you are designed for. Most procurement organizations are still designed for World Two. A cyclical commodity sourced on spot, managed in spreadsheets, with category strategies that get refreshed once a year. Almost none are designed for World One. The companies that hedge, and execute the hedge with discipline across the full BOM, are the ones who get to keep their options open when the direction changes.
Here is what I would put on the agenda for the next two quarters. The first move is memory-specific. The other four are operating-model moves that apply to every direct-material category you own, with memory as the most acute current example.
Re-segment your memory spend into three buckets, not one. Most procurement organizations treat memory as a single category run by one commodity manager. That model breaks in 2026. Split your buys into HBM and advanced packaging, where structural scarcity is real and the playbook is long-term capacity reservation. DDR5 and high-density server DRAM, where the cycle still rules and disciplined spot-plus-contract blending wins. And legacy DRAM and NAND for embedded, automotive, and consumer products, where the squeeze from HBM-driven wafer reallocation is the primary risk. Different buckets, different strategies, different risk profiles. Stop running them through one playbook.
Get to part-level cost and risk visibility across the full BOM, not just the noisy categories. Memory is what’s screaming today. Six months ago it was substrates. Eighteen months ago it was MLCCs. The pattern is always the same. The category nobody was watching becomes the one that determines whether you ship. The procurement organizations that absorb these shocks well aren’t the ones with the best memory analyst. They are the ones who have continuous part-level cost benchmarks, supplier risk signals, and BOM exposure mapping on every category that matters, refreshed faster than the market moves. If your team is still pulling that picture together in a spreadsheet every quarter, you are going to be late on the next category too.
Build supplier intelligence beyond your tier-one list. Two-supplier strategies dominated by the same handful of incumbents are not a hedge. They are concentration risk with extra steps. Memory is the obvious example. Micron is the only credible Western-domiciled leading-edge memory supplier, and the geopolitical premium on US-made content is going up, not down, regardless of which world we end up in. But the same logic holds across substrates, OSATs, power components, and the long tail of suppliers that sit two and three tiers down your BOM. Knowing who your sub-tier suppliers actually are, and having a working relationship with the credible alternates, is the single highest-ROI investment a procurement organization can make in 2026.
Renegotiate the commercial structure of your supply agreements, not just the price. The contracts most OEMs have today were written for World Two. They assume spot-market flexibility, short tenors, take-or-pay light. In a World One environment, that posture loses you allocation. And not just on HBM. The same dynamic is playing out across constrained categories from advanced packaging to specialty chemistries. Move toward longer-tenor agreements with capacity reservations, but build in volume flex bands, indexed pricing tied to published benchmarks, and explicit allocation language for the constrained sub-categories. The goal isn’t to lock in a price. It is to lock in a seat at the table.
Run scenario plans continuously, not annually. Every CSCO already runs scenario plans on tariffs, semiconductor lead times, and logistics disruption. The memory cycle is a reminder that the cadence on these has to change. World One and World Two are diverging fast enough that a once-a-year refresh is not good enough. The companies that get caught flat-footed in the next correction won’t be the ones who picked the wrong view. They will be the ones whose category strategies were still pointed at the world as it looked twelve months ago.
These five aren’t really five separate things. They are one operating model. A procurement organization that sees its full direct-materials picture in close to real time, runs disciplined hedges across every constrained category at once, and adjusts faster than the market does. Memory is the urgent case in 2026. The pattern is the bigger lesson.
This is the operating model we have been building toward at LevaData. Part-level cost and risk visibility across the full BOM. Continuous supplier intelligence beyond the tier-one list. Scenario planning at a cadence that actually matches the market. The reason I write about memory the way I do is because I see the same dynamics playing out across every constrained category our customers source, every week, in real time. Memory just happens to be the one that’s loudest right now.
The cycles haven’t gone away. They have just gotten bigger.
The 2023 primer was right about one thing it almost threw away in a single line. Memory is heavily cyclical. The peaks are higher now, the corrections are deeper, and the consequences for getting it wrong reach further into the operating model of every company that ships a product with a chip in it.
The companies that come out of the next three years in the strongest position won’t be the ones who guessed right. They will be the ones who built a direct-materials operating model that worked in both worlds, and started building it now.