The Blacklist Meets the Shortage

On 25 June, Apple raised prices on fourteen products without adding a single gigabyte of memory, and the shares fell. Around the same time, it emerged that Apple had approached Washington — the Commerce Department, then the White House — for clarity on whether it can keep sourcing memory from CXMT, a Chinese chipmaker the Pentagon has flagged.

That is how tight the memory market has become.

A shortage that stopped being a chart and became a price tag

The numbers are extraordinary. Contract prices for conventional DRAM rose roughly 90% in the first quarter of 2026 and another ~60% in the second.

This is not the 2020–23 pandemic shortage in a new coat. That one was broken supply chains; this is a structural reallocation of fab capacity toward the high-margin memory the AI build-out demands. One industry analysis (Kearney) expects it to persist to around 2030.

Apple is the visible face of the buyer-side pain. On 25 June it lifted prices across fourteen products — same hardware, higher price:

• MacBook Air — $1,099 → $1,299
• MacBook Pro — $1,699 → $1,999
• Across the line — +$30 (HomePod mini) to +$1,300 (a high-end Mac Studio)

No extra memory was added. Buyers simply pay more for the same chips inside; the iPhone was left untouched, for now. The memory tax stopped being an industry talking point and started showing up at the till.

The supplier nobody is supposed to want

Here is the part that makes the squeeze a story rather than a statistic. Apple approached the Commerce Department about a month ago, then escalated to the White House, seeking clarity on its procurement relationship with CXMT (ChangXin Memory Technologies).

At issue is whether CXMT could be added to the US Entity List, which would impose stiff licensing restrictions. Apple is not currently barred from using it; the open question is whether that stays true.

CXMT is not a neutral name. It sits on the Pentagon’s 1260H list of companies with alleged ties to the Chinese military — a designation about contracting and reputation, not an outright trade ban. The point is the direction of travel: a shortage is now testing a policy built to keep American technology and Chinese suppliers apart. Scarcity, not strategy, is reopening a door that policy tried to close.

The supplier in the headlines is the one the memory bulls were warned about

Step back and the irony sharpens. CXMT is the very name that careful memory analysts have flagged as the nearer clock on the memory cycle: the state-backed Chinese entrant scaling on the commodity-DRAM floor on which the whole memory multiple ultimately rests.

By most estimates CXMT runs around 265–300 thousand wafers a month of DRAM — roughly nine-tenths of a leading incumbent’s wafer scale — with a high-bandwidth-memory ramp underway, an HBM3 target for 2026, and a $4.2bn listing to fund it. The threat the bulls were told to watch is, today, supply the market is scrambling to secure.

Two-sided, plainly: CXMT is not yet a threat at the high-end (HBM). Independent teardowns put its HBM3 yields near 25%, it lacks the most advanced lithography, and stacking remains the binding obstacle — so the high-end memory where the incumbents’ real value concentrates is moated for years.

CXMT’s bite is on the commodity DDR5/LPDDR5 floor, not the high-end ceiling. The high end holds; the floor is where the pressure builds.

Scarcity flipped the leverage

Put the pieces together and you get an awkward feedback loop. The same shortage that hands the incumbent memory makers their pricing power is, at the same moment, pulling demand toward the disruptor — getting CXMT’s die qualified, designed into Western hardware, and its scale-up funded by the very scarcity it will eventually help to end.

The oldest law in commodities applies: the cure for high prices is high prices. Here, the cure happens to wear a blacklist.

The escape route that adds no fab at all

There is a way out of the squeeze that builds nothing: needing less of the scarce memory. The clearest sign it is real is that the company most exposed to high-bandwidth memory is hedging against it — Nvidia struck a $20bn deal for Groq’s inference chips, which run on fast on-chip SRAM and skip HBM altogether.

It is not alone. Cerebras runs an entire silicon wafer as a single chip, carrying its memory on-chip as SRAM rather than in HBM stacks, and is already in production with real customers. One floor down, AMD bought MEXT, a memory-software firm whose tool lets cheap flash stand in for costly DRAM.

Three companies, two layers of the memory stack, one idea: engineer the need down rather than wait for supply. None of this displaces HBM tomorrow — the GPU complex still consumes it voraciously. But if these routes mature on the same 2027–2030 horizon as the new fabs, the long-run cap on memory pricing power may come not from a Chinese commodity flood but from the architecture moving past scarce memory — with the margin migrating to whoever owns the road around the tax.

Two ways this resolves

The bull reading. The shortage is structural and long — possibly to the end of the decade. AI memory demand keeps the incumbents’ pricing power intact, and CXMT stays years away from the high-end (HBM) where the real margin lives. The moat holds; the cycle runs hotter for longer.

The bear reading. The same shortage is hand-feeding the disruptor on the commodity floor, and the policy contradiction is real: Washington cannot both wall the incumbents off from CXMT and leave the door open for the buyers who now depend on it.

The high end is safe. The franchise valuation may not be.

• CXMT’s status — an Entity-List listing is an escalation, a carve-out an accelerant.
• August quarterly prints — the first clean read on whether the commodity-DRAM price leg is still climbing or beginning to roll.
• The demand-side routes — whether SRAM-based inference (Groq, Cerebras) and flash-as-DRAM software gain traction at scale.

What to do with this — and what we are not doing

What to look for. Whether the memory price leg is still rising or topping in the August prints; the US decision on CXMT’s status, which sets the direction for the whole commodity-DRAM clock; and whether the AI build’s binding constraint stays physical.

What to be wary of. Reading a record shortage as an unambiguous positive for the memory makers — the same scarcity is funding the disruptor and the route-around; and conflating the protected high-end (HBM) with the contested commodity floor — two markets, two futures.

What we are not doing. This is general market commentary, not advice or a recommendation on any security. The companies named are described as illustration, not as a trade. Anything that touches your own portfolio is a suitability conversation — one that needs a person, not a page. We are always glad to have it.