The question this piece answers: who owns the 2027–2028 memory supply, and what did they sign away to get it? For thirty years, DRAM was the textbook commodity: spot prices, quarterly repricing, a boom-bust cycle everyone could recite. Between June and October 2026, all three large suppliers — Micron, SK hynix and Samsung — converted a structural share of their future output into long-term, take-or-pay, deposit-backed contracts. Every number in this piece is taken from the companies' own earnings materials or press releases, cross-checked against at least two independent reports; where sources conflict, the piece says so instead of picking a favorite.

What did the three memory makers actually sign?

Start with the raw disclosures, because the interpretation only works on top of them. The three companies report on different calendars and disclose at different depths, so the table mixes hard filings with earnings-call statements — and marks which is which.

MetricMicron (FQ4-26, rep. Sep 30)SK hynix (Q3-26, rep. Oct 28)Samsung (Q2/Q3-26 calls)
Long-term agreements26 SCAs, up from 16 in FQ3 (+62.5%); take-or-pay; some into 2031; three automotive SCAs signed July 2026 (Ford on Jul 6, GM on Jul 1 — both on Micron's own IR site)14 customers, including leading cloud providers; typically five-year terms, with deposits10 signed, 10 in final talks; base term ~5 years + 1-year extensions
Coverage>35% of estimated revenue through 2030 (from ~25% one quarter earlier)>50% of 2027 and 2028 production capacityTarget: 60–70% of DRAM+NAND capacity by 2030
Price mechanism~three-quarters of SCA revenue under a defined framework — mostly floor/ceiling price bands; rest market-adjustedPrices "move within a range agreed with customers" (per the company's call, as reported)Not disclosed; prepayments required as part of terms
Money on the tableRPO ≈ $150B at minimum prices (from $100B, +50% QoQ); customer commitments $32B, mostly cash deposits; $12.7B deposits on the balance sheetDeposits included; amount not disclosedNot disclosed
Supply picture>75% of calendar-2027 output committed; vast majority of 2027 HBM bits under agreements at "significant year-over-year price increases"; supply constrained through calendar 2028Demand requests = 2× supply capacity; DRAM+HBM effectively sold out through 2029 (signed plus in-negotiation)2027 HBM fully sold out

One verification note before anything else: the SCA count, the coverage percentages and the price-band structure come from Micron's earnings call and prepared remarks, not from the press release itself — the press release confirms the financials (revenue, margin, the $12,747 million of customer contract-liability deposits in the cash-flow statement, and the CEO's SCA statement) but the deal detail is disclosed on the call. Zacks's two independent reports of the call, Yahoo Finance's analysis, The Motley Fool's and the Taiwanese press (工商时报, citing Seoul Economic Daily) all agree on the numbers to the dollar and the percentage point. That is five independent readings of one primary source, and none disagrees.

Why is the spot price no longer the price?

Because the market is being re-plumbed. TrendForce's 7 October 2026 DRAM bulletin — an industry price tracker whose whole business is watching this — describes the new structure: suppliers and cloud providers "have signed multi-year server DRAM LTAs, placing many of its bit supply under price-band mechanisms," with the rest still negotiated quarterly — and it characterizes the trade-off in one line: "trading some price upside for a smoother, more predictable cycle." Read that slowly. The contract price moves inside a band; the spot price applies to whatever bits were not contracted. When more than half of future capacity is inside bands — SK hynix's >50% for 2027–28, Micron's >35% of revenue through 2030, Samsung's 60–70% target — the spot quote stops being the market price and becomes the clearing price for the residual.

This is the LNG playbook. Liquefied natural gas spent decades moving from spot cargo to long-term destination contracts with take-or-pay clauses and price review mechanisms, and the result was that "the LNG price" meant the contract price, with spot as a thin secondary market. Memory is doing the same conversion at compressed speed, for the same reason: a supply-side product that requires years of capital to expand, bought by customers whose own capex cycles are measured in years, cannot discover its price one quarter at a time without both sides screaming. The difference is that DRAM did in six months what LNG took a decade to do (analysis, ours).

The demand side shows why it happened now. SK hynix reported that customer demand requests reached twice its available supply capacity — a ratio at which no market-clearing price is meaningful, because at 2× demand the price is simply rationed upward until someone drops out. When demand runs at twice supply, the efficient response for both sides is to replace the auction with a queue: contracts, deposits and dated allocations. That is exactly what got signed.

What does "even at floor prices" mean?

It means the downside protection is struck at the top of the cycle — and that sentence deserves unpacking, because it is the most consequential disclosure of the season. Murphy's prepared remarks (quoted by multiple outlets covering the 30 September call) state: "even at floor prices, we expect margins meaningfully above any prior cycle peak margins." Two facts sit inside that line. First, the contracted floor — the minimum price buyers committed to pay — is set at a level where Micron's margins would still beat every prior peak quarter. Second, Micron's current quarter is not a peak clinging to history: 87.0% non-GAAP gross margin in FQ4-26, guided to ~86.25% in FQ1-27 and described by the CFO as the floor for fiscal 2027, with margins expected to rise from there.

Now the reading (ours). A price floor is a put option the buyer sells to the supplier, and every option has a strike. In past cycles, long-term contracts failed buyers because the strike was set near then-current prices — when the cycle turned, the contract was underwater and both sides renegotiated. The 2026 structure prices the floor against record-cycle economics: even if the market turns, contracted volume keeps clearing at prices that would have been unimaginable peaks in any prior cycle. The buyer is paying for certainty at a strike set at the top. That does not make the buyers irrational — $150 billion of contracted revenue counted at minimum prices, covering 35%+ of revenue through 2030, is what lets Micron double capex (first-half fiscal-2027 capex of ~$25 billion, against $27.37 billion for all of fiscal 2026) without financing cycle risk. But it does mean the "risk transfer" everyone is celebrating is mostly one-directional: the seller's downside is capped by the floor, and the buyer's upside on price is capped by the ceiling.

Who is financing whom?

Follow the deposits, because they tell you the contracts are not just purchase orders — they are financing instruments. Micron's FQ4-26 cash-flow statement shows $12,747 million of proceeds from customer contract liability deposits in fiscal 2026 (against zero the prior year), and the balance sheet carries $12.895 billion of noncurrent customer contract liabilities. On the call, the company put total customer financial commitments at $32 billion, mostly cash deposits — up from $22 billion a quarter earlier. SK hynix's agreements include deposits. Samsung requires prepayments. And CXMT (ChangXin Memory), the Chinese entrant now being pulled into the same structure, disclosed at its 30 October briefing that contract liabilities grew to ¥7.036 billion from ¥447 million at the end of 2025 — a roughly fifteen-fold increase in customer prepayments within nine months (as reported by two independent outlets from the company briefing).

Run the ratio (ours): $32 billion of commitments against $150 billion of contracted RPO is a ~21% down payment on the contracted book. The deposits on Micron's balance sheet alone — $12.7 billion — equal 46% of Micron's entire fiscal-2026 net capex of $27.37 billion. In plain terms: the customers are pre-funding nearly half a year of the supplier's fab construction. That is not a purchase; that is a capital contribution wearing a procurement contract. It is also the answer to "why would buyers accept floors at record prices": because in a 2×-oversubscribed market, cash is the only allocation mechanism left, and the party willing to prepay is the party that gets bits.

The same logic explains the capex acceleration everyone flagged as euphoric. Micron's first-half fiscal-2027 capex of ~$25 billion nearly equals all of fiscal 2026's $27.37 billion; the full-year plan, as reported in the Taiwanese and Chinese trade press from the same call, exceeds $50 billion. No board approves that on spot prices. They approve it because more than 75% of calendar-2027 output is already committed — the customers have, in effect, underwritten the fab. The LTA is the collateral for the capex loan the supplier takes from its own balance sheet (ours).

What does this mean if you are the one buying memory?

Three concrete translations. First: your negotiation is about allocation, not price. When SK hynix's reported demand runs at 2× supply and Micron has 75% of 2027 committed, the quoted price matters less than whether you are in the queue at all. If your supplier can offer you a dated slot with defined pricing, that is the deal; if they can only offer "market price plus availability," you are buying the residual, and the residual is where the volatility lives. Second: the 2028 conversation has already started. Micron's management said discussions increasingly center on 2028 demand; SK hynix's signed-plus-negotiating book reaches 2029. Buyers who wait for "the market to settle" will be negotiating against people who locked allocation in 2026. Third: expect the two-tier market to become visible in your quotes. TrendForce's September 23 bulletin named the consequence already: "LTAs shrink negotiable supply as new fabs lag" — contracted bits do not respond to spot signals, so the uncontracted remainder must carry the entire adjustment burden. A price spike in the spot quote is no longer a signal about the whole market; it is a signal about the sliver that is still unspoken for.

And for the home-lab reader — this forum's other beat — the practical version is simpler: the consumer segment is the tail that wags for nobody. Samsung's disclosure distinguishes explicitly: standard DRAM and consumer NAND still run on normal order-allocation, while server DRAM and HBM are locked years forward. The DDR5 stick you buy at retail is priced off the residual market that the contracts have made thinner and jumpier. Nothing about that changes the buying rule — buy when you need it, size your build around the 1,920-watt branch we documented earlier — but it explains why retail prices stopped making sense in 2026.

What are the caveats?

Deal detail is call-disclosed, not filing-disclosed. The SCA count, coverage percentages and price-band structure come from earnings calls and prepared remarks; the audited press release confirms the financial statements (including the $12.7B deposits line) but not every contract statistic. We treated the numbers as reliable because five independent reports of the same call agree exactly — but they are management statements, not audited figures. Samsung's disclosure is the thinnest. The 10-signed/10-talking count and the 60–70% target come via earnings calls as reported by Seoul Economic Daily, TrendForce and two Taiwanese outlets; Samsung discloses no RPO equivalent, so its column is structurally softer than Micron's. HBM-specific price forecasts are single-source. TrendForce's view that HBM contract prices could rise 70–140% in 2027 is one house's forecast; we attribute it and do not treat it as fact. Percentages measure different things. Micron's 35%+ is share of revenue through 2030; SK hynix's 50%+ is share of production capacity for two specific years; they are not comparable, and this piece never averages them. And nothing here is investment advice — it is a reading of public earnings materials, offered because the same contract structure is now deciding what every stick of memory costs.