AI memory trade broadens as Marvell rerates and SK Hynix tightens its grip

Two semiconductor stories on 2 June 2026 are two ends of the same AI value chain — the bottleneck is moving up the stack.

By Monexus Staff Writerglobal7-minute read3 Jun 2026☆ Save ↗ Share ⎙ Print

On 2 June 2026, two independent signals from the AI hardware stack converged on the same thesis: the value of the AI cycle is no longer sitting only with the GPU maker.

South Korean chipmaker SK Hynix consolidated its position as the leading supplier of high-bandwidth memory — the stacked DRAM modules that sit next to AI accelerators and feed them data at the rates the workloads require. Reporting from Nikkei Asia on 2 June framed the company's pivot to HBM as a deliberate escape from commodity semiconductor cycles. That escape, on the evidence of the share price and the customer rosters, has largely worked.

A continent away, custom-silicon designer Marvell Technology jumped roughly 30 percent in a single trading session on 2 June after Nvidia chief executive Jensen Huang, in remarks covered by CryptoBriefing, named the company as a candidate member of the trillion-dollar club. The endorsement, more than any product cycle, was the proximate trigger.

Read together, the two moves are not separate stories. They are the same story told from two ends of the supply chain. The bottleneck in AI infrastructure is migrating up the value chain from the accelerator itself to the supporting silicon — memory, networking, custom chips — and that redistribution is creating a new class of winners tethered not to consumer demand cycles but to the capital expenditure plans of a handful of hyperscale customers.

The bottleneck the GPU cannot clear

HBM is the narrowest bottleneck in the entire AI build. Training a frontier model requires terabytes of memory bandwidth that a standard DDR5 module cannot provide. HBM stacks die vertically and connects them through through-silicon vias, achieving bandwidths an order of magnitude higher per stack than commodity DRAM. Demand from Nvidia, AMD, and a long queue of hyperscalers building in-house accelerators has outrun supply through the current AI buildout. SK Hynix, by virtue of early process-leadership decisions and a tight working relationship with Nvidia's accelerator generations, took the dominant position in current HBM product cycles. Its main competitors — Samsung Electronics and Micron Technology — have been chasing the qualification window for the next generation, where the volumes are concentrated.

The Nikkei Asia framing matters. Memory has historically been a brutal cyclical business. The capital intensity of fab construction and the willingness of Asian incumbents to add capacity in downturns have repeatedly produced two-year busts that wiped out the gains of the previous boom. The bet SK Hynix is making is that HBM is different — that the technical specificity of stacking, the qualification timelines with one or two anchor customers, and the absolute scarcity of leading-edge packaging capacity make this memory cycle look more like a logic-foundry business than a DRAM business. Pricing power sticks. Margins hold. Returns on capital stop mean-reverting the way they used to.

That is also why the market is willing to award the company a multiple that would have been unthinkable for a memory name five years ago. The thesis is not that DRAM is a good business again. The thesis is that HBM is not really DRAM — and the marginal analyst is finally pricing the distinction.

The Marvell signal and the widening of the trade

The Marvell story is a different cut of the same fabric. Custom silicon — the application-specific chips that hyperscalers like Google, Amazon, Microsoft, and Meta now design internally for inference, networking, and storage workloads — has been one of the quietest multi-billion-dollar businesses of the AI cycle. Marvell is one of the two or three public companies that monetise that build-out through custom ASIC design, high-speed SerDes, and data-centre interconnect. When Huang called the company a candidate trillion-dollar enterprise, he was doing two things at once. He was paying a public compliment to a partner whose custom-silicon work reduces Nvidia's customers' dependence on Nvidia's own accelerators. And he was signalling, on the record, that the AI silicon market is large enough to support multiple trillion-dollar entrants, not just one.

The market reaction — a 30 percent intraday move on a single endorsement — is itself a piece of evidence. A move of that magnitude on a CEO's offhand remark from a peer is not the response of a market that thinks the AI buildout has peaked. It is the response of a market repricing the addressable market for AI infrastructure in real time, with the trigger being a respected insider's verbal confirmation that the ceiling is much higher than the consensus had penciled in.

Why this looks like a broadening, not a peak

Both stories sit inside a structural shift that the consensus narrative is only beginning to price. For most of 2023 and 2024, the AI trade was effectively a single-name trade. Nvidia carried it. The bull case was simple: spend is enormous, Nvidia has the best accelerator, Nvidia captures most of the spend. The bear case was equally simple: spend normalises, custom silicon eats into Nvidia's share, the cycle ends, and Nvidia's multiple compresses.

The Marvell move suggests the bull case is winning on its own terms but the path to the spend is wider than one ticker. The SK Hynix move suggests the spend is not just wider but more durable, because the HBM bottleneck cannot be cleared quickly. New HBM capacity requires new fabs, new packaging tools, and a multi-quarter qualification cycle with the anchor customer. None of that can be compressed by simply writing a larger purchase order; the fab economics and the qualification windows are immovable on the timeframe that matters for equity returns.

The corollary for the rest of the memory complex is uncomfortable. If HBM commands the kind of pricing and margin profile that logic has historically enjoyed, then commodity DRAM and NAND will continue to be valued as a residual — a buffer to be used when HBM tooling is fully booked. The cyclical returns of the pre-AI era are not coming back for the legacy product lines; the upside is concentrated in the leading-edge product lines, and the legacy business exists mostly to keep fab utilisation high and to provide downside optionality if the AI cycle peaks. That is not the kind of return profile that supports a high multiple on the parent company.

What remains contested

This framing also exposes a question that the source material does not yet answer. The Nikkei Asia report frames SK Hynix's strategy as an escape from commodity cycles. CryptoBriefing's report frames Marvell's rerating as an endorsement of a wider AI silicon stack. Neither piece, taken alone, gives a clear view of how the supply of HBM and the demand for custom silicon are likely to evolve together over the next four quarters. The two stories are correlated in the abstract — both depend on hyperscale capex — but the cadence of each company's revenue cycle is different, and a divergence between the two is plausible if hyperscalers pull back on accelerator orders while continuing to invest in custom silicon for inference, or vice versa. The two trades are not the same trade, even if they rhyme in the short term.

What is not plausible is that the AI buildout, viewed through either window, looks like it is contracting. A 30 percent move in Marvell on a Tuesday in early June is not the action of a market that has stopped believing in the spend. A Nikkei Asia profile of a memory company running at near-full utilisation on leading-edge product is not the framing of a sector in retreat. The bottleneck has not lifted. The demand has not softened enough to be visible in the share-price action. What has changed is the price of admission to the trade: it is no longer Nvidia, full stop. It is Nvidia and the names that make Nvidia's accelerators work, and the names that design the silicon that competes with Nvidia's accelerators, and the names that supply the memory and the interconnect that hold the whole thing together. The winners on a continuation of the cycle are SK Hynix, Marvell, the leading-edge packaging complex, and the hyperscalers themselves. The losers are the holders of legacy memory and the holders of a single-name thesis on the GPU maker. That is a broader trade, with more entrants, and arguably a more durable one.

How Monexus framed this: The wire coverage treated the SK Hynix HBM lead and the Marvell surge as separate stories. We treat them as two ends of the same value-chain shift, and read the 30 percent Marvell move as evidence that the AI trade is broadening, not peaking.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

https://t.me/s/NikkeiAsia
https://t.me/s/CryptoBriefing
https://en.wikipedia.org/wiki/High_Bandwidth_Memory
https://en.wikipedia.org/wiki/SK_Hynix
https://en.wikipedia.org/wiki/Marvell_Technology