Here is the contradiction almost everyone gets backwards. The $94 billion in long-term memory agreements now spanning Samsung, SK Hynix, and Micron is being sold as a safety net—a stabilizer that finally tames memory's brutal boom-bust cycle. It is nothing of the kind. It is a rationing regime, and the fine print reveals who it protects.
Start with Andy Jassy's confession, because it was a confession, not a forecast.
On Amazon's second-quarter call, Jassy said the company would spend "approximately $220 billion in cash capex in 2026," then added the line that should have stopped the room: "The higher cost of memory is pushing this number up from our prior estimate of about $200 billion."
That is $20 billion of increase that Jassy attributed to the higher cost of memory.
Then came the part that turns an accounting note into an organizing fact: "Even at that amount, we will not have enough capacity to meet all the demand we have in 2026, and I believe this dynamic will also be true in 2027."
Read that carefully. The largest cloud buyer on earth is telling you that raising capex by 10% still does not buy enough of the one input it cannot substitute. Money has stopped clearing the market.
The largest cloud buyer on earth is telling you that raising capex by 10% still does not buy enough of the one input it cannot substitute. Money has stopped clearing the market.
Last week we treated memory scarcity as the theme; this week it has a price tag and a contract structure. The $94 billion figure is the scale of the "safety net" that DRAM makers have written with their customers, with agreements reportedly extending through 2030. The delta hyperscalers pay for it is roughly 30% on AI data-center capex, per Bernstein's math. Both numbers deserve scrutiny, because they describe the same transfer from two sides.
Here is the mechanism. These LTAs are not demand contracts in any traditional sense—they are supply-defense instruments, and their terms systematically strip the buyer of the leverage memory buyers have always enjoyed.
Samsung's agreements reportedly cap the quarterly price decline at 5% for standard storage products—with no upside cap—and the company has collected 25% as a prepayment deposit. Sit with the asymmetry: prices can only fall slowly, but they can rise without limit.
SK Hynix went further, reportedly removing price ceilings on some contracts and linking them to spot prices. In a shortage, spot-linked with no ceiling is not a hedge. It is an open-ended claim on the buyer's wallet.
Micron, meanwhile, has reportedly secured roughly $22 billion in cash deposits and related commitments across multiple five-year strategic contracts, with about $18 billion actually received. Management says profitability "exceeding the peak levels of past cycles" is possible on the guaranteed pricing structure.
Notice what every one of these terms does. It converts the memory maker's historic vulnerability—price collapse in a glut—into a floor, while leaving the buyer fully exposed on the ceiling. The volatility didn't disappear. It got reassigned.
That is why calling this a safety net is a category error. A net protects whoever falls into it. This one is stretched under the sellers.
Now the rationing reality, which is where the two-tier economy hardens into place.
DIGITIMES, citing industry sources, reports that 2027 DRAM and HBM capacity at Samsung, Micron, and SK Hynix is fully allocated—and that customers receive only 60% to 70% of their initial requests. A signed long-term agreement does not guarantee your volume. It guarantees your discount to a shortage you must still fight over.
Micron's Sumit Sadana made the constraint blunt at KeyBanc's forum: "The number one constraint for customers is DRAM." Not power, not real estate, not data-center space, not logic wafers.
He then delivered the sentence that defines the year: some data-center customers "are unable to secure more than half of the memory volume they need, even at elevated prices."
Even at elevated prices. Price is no longer the allocation mechanism. Access is. And access flows through prepayment and contract seniority.
That is the two-tier structure in one line. Amazon, Microsoft, Alphabet, Meta, and Oracle, collectively projected to spend over $700 billion in 2026 capex, get supply and predictable cost inflation. Everyone else gets the 30% capex tax with no volume guarantee.
The physics make the rationing worse, not temporary. Producing 100 bits of HBM3E reduces DDR output by roughly 300 bits, a three-to-one trade. Sadana warned HBM4E could push that toward four-to-one. Every AI accelerator built starves the standard DRAM pool that everyone else needs.
So the shortage isn't a queue that clears. It is a structural conversion of general-purpose memory into accelerator-bound memory, and the LTAs decide whose queue gets served.
Which brings us to the asymmetry that is the whole point. Follow the margin.
UBS's Timothy Arcuri, ranked second of more than 12,000 analysts on TipRanks, calls Micron's earnings power a "structural reset" and models EPS of $74.13 for 2026, $184.89 for 2027, and $265.65 for calendar 2028—all well above consensus. Bernstein separately, per its reported forecast, pegs its Samsung, SK Hynix, and Micron 2027 EPS estimates 25% to 40% above the street.
Those numbers are not a demand story. They are a terms story. They are what happens when cost pass-through flows one direction only.
Arcuri named the shift explicitly: memory has moved "from a commodity product into a strategic asset," with demand now showing high price inelasticity, cemented by long-term agreements that cover roughly half of Micron's revenue. Inelastic demand plus no price ceiling plus prepaid capacity equals the seller capturing the surplus.
Consider the pass-through chain Bernstein traces. HBM is only about 5% of an Nvidia VR200 rack's price, but if HBM rises 2 to 2.5x and Nvidia marks that up roughly 4x to hold its 75% gross margin, the rack price climbs about 24%. Stack conventional DRAM and NAND on top, and hyperscaler AI capex swells the full ~30%.
At every link—Micron to Nvidia, Nvidia to Amazon—the markup compounds. The one party who cannot mark up and pass along is the hyperscaler at the end, whose customers ultimately absorb it or whose margins don't.
This is the answer to the question everyone in the AI trade keeps asking: who captures the margin? For years the assumption was the cloud giants, sitting on scarce compute. The contract terms say otherwise.
The 30% capex delta is not a cost of doing business. It is a toll. It is the price hyperscalers pay to convert their own volatility risk into someone else's guaranteed profit—and they are paying it because, as Jassy admitted, the alternative is not having capacity at all.
There are cracks worth watching. Foreign investors reportedly net-sold nearly 6 trillion won of Korean chip stocks in a single early-August week, wary the AI cycle won't hold. Apple is reportedly seeking White House approval to install CXMT memory, testing whether the three-maker cartel can be diversified around. Morningstar warns that new fabs take two to three years, and that peak output meeting cooling demand is the classic setup for a glut.
But the glut, if it comes, arrives with a defense that no prior cycle had: Samsung's 5% quarterly floor, Micron's prepaid billions, SK Hynix's ceiling-free spot linkage. That is precisely the point of the $94 billion. It was written to make sure the next downturn lands on the buyers.
For now, the confession stands. The largest cloud company on earth raised its budget by $20 billion for memory alone, and still expects to come up short—in 2026, in 2027, with 2028 demand already "striking." The people who wrote the contracts made sure that shortfall is not their problem. It is the toll, and it is due at delivery.