The Memory Cycle Breaks — Micron Signs Take-or-Pay Contracts and Customer-Funded Capacity, and Oracle's $638bn Backlog Confirms the Buildout Went Off-Balance-Sheet

The Lead

Last issue’s “Week Ahead” said Micron’s print was the cleanest read in the complex — whether the HBM/DRAM pressure was a sector-wide squeeze or a margin gift. The 10-Q answered a bigger question than the one we asked. Micron isn’t just enjoying a tight cycle; it is changing the contract structure of the memory business so the cycle matters less.

Read directly from the filing: Micron’s strategic customer agreements “are structured as take-or-pay agreements, with binding commitments for specific volumes over the multi-year contract terms,” and “pricing for most agreements is either fixed, or is subject to mi[nimum bands].” Those customers are posting cash deposits — “customer deposits associated with strategic customer agreements for which revenue has not yet been recognised” — which Micron expects to keep receiving “under strategic customer agreements.” In plain terms: Micron has persuaded its largest buyers to commit to volumes and pre-fund the capacity, in advance, on contracts that fix or band the price. That is the memory industry’s holy grail — revenue visibility and a floor under pricing — and it is the structural opposite of the spot-market commodity that memory has always been. To pull supply forward, Micron also bought a fully-built fab from Powerchip in Tongluo, Taiwan, skipping the multi-year greenfield build. (It also buried a Section 301 trade-investigation risk in the risk factors — the one cloud worth tracking.)

The second filing of the week is the financing bookend. Oracle’s 10-K disclosed a remaining-performance-obligation backlog of $638bn — up from $138bn a year earlier, a 4.6× surge — built on capex of $55.7bn, roughly 83% of revenue. And the tell that closes a loop we opened in Issue 12: Oracle explicitly carries “lease commitments that have not yet commenced and were not included in our consolidated balance sheet,” tied to “long-term lease commitments with third-party data centre providers,” against $129.5bn of outstanding debt. The compute buildout is being financed off the income statement and off the balance sheet — exactly the migration forecast [2026-06-13-001] predicted. That forecast confirms (see below).

Two filings, one theme: the AI buildout is locking in its economics — Micron locking in demand and price, Oracle locking in capacity and financing — and in both cases the structures sit partly outside the income statement. The cycle isn’t being ridden; it’s being contracted away.

Threads

Micron breaks the memory cycle — take-or-pay contracts and customer-funded capacity de-commoditise DRAM/HBM

The whole memory thesis has always carried a cyclicality discount: demand is real but pricing is a commodity, so the multiple stays low because the next glut is always coming. Micron’s 10-Q attacks that discount directly. Take-or-pay means the customer pays for the committed volume whether or not it lifts it — the volume risk moves to the buyer. Fixed or banded pricing means the price risk is dampened too. And customer deposits mean the buyer is funding Micron’s capacity build ahead of delivery — customer-funded capex that reduces Micron’s external financing need and signals how badly the capacity is wanted. Stack those three and a chunk of Micron’s revenue stops behaving like a commodity and starts behaving like a contracted utility. The Powerchip fab (Tongluo, Taiwan) is the supply-side complement: buying finished cleanroom capacity to meet the contracted volumes years before a greenfield could. The convergence engine reads MU bullish and top of the board this week — and for once the structural filing detail supports the score rather than the tape driving it.

The open question the forecast below pins down: is this a Micron-specific coup, or the start of an industry shift? If Samsung and SK Hynix — equally supply-constrained — follow into contracted pricing, the cyclicality discount compresses across the whole sector, not just one name.

Oracle’s $638bn backlog confirms the off-balance-sheet thesis — 83%-of-revenue capex and data-centre leases that never touch the balance sheet

We have to be precise here, because precision is the product. Oracle’s verified disclosures: RPO of $638bn (from $138bn), capex $55.7bn (~83% of revenue), $129.5bn of outstanding debt maturing out to 2066, and — the load-bearing line — off-balance-sheet “lease commitments that have not yet commenced,” tied to long-term data-centre leases with third parties. The DeepSeek extraction also flagged a “$260bn off-balance-sheet lease” figure; we could not verify that number against the raw 10-K and have not used it. What is verified is the structure: Oracle is contracting enormous future data-centre capacity through leases that sit outside today’s balance sheet, funded alongside record debt and capex.

That is the precise pattern Issue 12’s forecast described — “the off-balance-sheet financing of compute proliferates… a customer compute-lease guarantee, or a delayed-draw / SPV facility explicitly tied to AI capex.” Between Oracle’s off-balance-sheet data-centre lease commitments and Micron’s customer-deposit-funded take-or-pay structures, two more large names disclosed compute financing that lives outside the income statement, inside the window. The read-through hits the hyperscalers Oracle is chasing (the engine flags AMZN/MSFT/NVDA off the RPO surge), but the durable point is the financing form, not the names: when the backlog is $638bn and the leases aren’t on the balance sheet, the risk — when it comes — will be hard to see from the outside. Issue 11 said follow the debt; Issue 12 said it went off the balance sheet; Issue 14 confirms it’s now standard practice.

The landlords scale up the stack — Digital Realty consolidates Teraco, buys a venture arm, and banks land

The physical layer kept integrating. Digital Realty’s 8-K bundled three moves: it lifted its stake in Teraco — Africa’s largest data-centre platform — from ~55% to 77% (triggered by a partner exercising a put right, i.e. partners monetising into a scarce asset); it acquired Columbia Capital, a digital-infrastructure VC/PE manager with ~$9bn in fund commitments, giving DLR an in-house venture arm and a pipeline; and it land-banked ~1,440 acres near Kansas City for $377.6m, with the seller taking partnership units rather than all cash. A REIT doesn’t usually buy a venture firm. Read together, DLR is no longer just leasing racks — it’s consolidating scarce platforms, building a capital-allocation arm, and warehousing developable land for the next leg. The data-centre landlord is becoming a data-centre financier, which rhymes with the Oracle and Micron threads: everyone in the stack is locking in capacity and the means to fund it.

Since Issue 13 (18 June)

• Forecast [2026-06-13-001] — CONFIRMED. Predicted: within 60 days, ≥1 more mega-cap or merchant-silicon name discloses a structured / off-balance-sheet compute-financing arrangement tied to AI capex. Oracle disclosed off-balance-sheet data-centre lease commitments “not yet commenced and not included in our consolidated balance sheet,” and Micron disclosed customer-deposit-funded take-or-pay agreements — two distinct off-income-statement financing structures, both inside the window, both verified against the primary filings. The thesis that the buildout’s financing is migrating off the balance sheet is now evidenced three issues running.
• Forecast [2026-06-18-001] — OPEN, with an honest caveat. The call (the convergence-bearish AI-hardware names — DELL, CRWV, CRDO, ALAB, MRVL, ANET — underperform TSM/AVGO over 60 days, resolving ~17 Aug) was built on the convergence engine’s then reads. This period we fixed four measurement bugs in that engine (a 1000× units error on 13F values, individual insider sells mis-scored as bearish, a mis-attribution join, and a transcript-coverage gap). The corrected engine now reads several of those names — DELL and CRDO — bullish, not bearish. The forecast is still graded on price at resolve_by, but its premise rested partly on signal that has since been corrected; we flag that rather than quietly let it ride. (Detail in the methodology note.)

Watchlist Updates

• MU tops the convergence board (bullish) and is the issue’s anchor — the take-or-pay structure is the most important single disclosure of the period. Watch whether the deposits and contracted volumes show up as margin stability next quarter.
• ORCL reads bullish on demand (the $638bn RPO) but the watch is the funding: 83%-of-revenue capex, $129.5bn debt, and off-balance-sheet leases mean the equity is a leveraged bet on the backlog converting on schedule.
• DLR (and EQIX by read-through) — the landlords are consolidating; the question is returns-on-capital as land and platforms get bid up.

The Week Ahead

The calendar quietens after the Micron/Oracle one-two. The catalyst worth pre-positioning is structural, and it’s this issue’s forecast:

Forecast [2026-06-27-001]: within 60 days, at least one of Samsung or SK Hynix discloses a take-or-pay / long-term contracted-pricing structure analogous to Micron’s. Confidence 0.55, horizon 60 days, event-judged. If it confirms, memory’s shift from spot to contracted pricing is an industry regime change and the sector’s cyclicality discount should compress broadly — not just at Micron. If 60 days pass with the peers still selling on spot terms, Micron’s coup is idiosyncratic and the de-commoditisation read is overstated.

Methodology and Disclosures

Threads are written from the SEC-filing database the system maintains; this issue covers the 19–27 June window (9 filings, 25 signals). Load-bearing figures were verified against the primary filings: Micron’s take-or-pay language and customer deposits, Oracle’s $638bn RPO, $55.7bn capex, off-balance-sheet lease commitments and $129.5bn debt. Where the automated extraction produced figures we could not verify — a “$18bn” Micron deposit total, a “$260bn” Oracle off-balance-sheet lease figure, SpaceX’s cash position — we omitted them rather than publish unverified numbers. Separately, this period four measurement bugs in the convergence/scoring layer were fixed (13F value units off by 1000×, insider sells mis-weighted, an outcome mis-attribution join, and transcript outcomes ungraded); the corrected engine changed several convergence reads, which is why this issue’s map differs from Issue 13’s and why we caveated forecast [2026-06-18-001]. The machine-graded calls ledger holds 47 open calls awaiting their resolution windows. Forecasts are pre-registered, dated and falsifiable, and graded in the following issue. Nothing here is investment advice.