brain/
sourcestock-market

Autoresearch: Did SK Hynix scrap or keep LTA DRAM/HBM price caps, and can 7–9 customer inventory weeks be corroborated?

Re-fetch of the Jul 22/29 2026 SK Hynix LTA-cap gate and the 7–9 week customer DRAM inventory claim; official language is variable pricing plus deposits, not scrap-or-keep; 7–9 weeks remains uncorroborated.

Source

Autoresearch: Did SK Hynix scrap or keep LTA DRAM/HBM price caps, and can 7–9 customer inventory weeks be corroborated?

Generated by /autoresearch on 2026-09-19. Synthesized across 3 rounds from 11 web pages (1 failed), anchored by the Grokipedia SK Hynix entry. See Provenance. Treat as raw material — review before promoting into a project or thread. Context: vault/projects/stock-market Research-only pass. This file is written outside the vault. No buy / sell / size.

Summary

SK Hynix’s official 2026 Q2 print and call are dated 29 July 2026, not 22 July. The 22 July primaries are a P&T7 packaging-plant investment 6-K and an Ohio-campus rumor denial — they do not discuss LTAs or price caps. On the 29th the company confirmed LTAs with around 10 customers, typically about five years, with deposits and non-uniform pricing “designed to address price volatility.” It did not say caps were scrapped, did not say caps were kept, did not say take-or-pay, and refused to disclose LTA sales coverage. No later 2026 SK Hynix earnings primary exists as of 19 September 2026. The 7–9 week customer DRAM inventory figure remains single-sourced to unverified useluminix.com; TrendForce HTML, Micron’s 10-Q, and SK Hynix IR do not print that number. SemiAnalysis 2027 HBM pricing sits behind a paywall and is not treated as confirmed.

Findings

The July 22 “Q2 report” gate is a date error

The live-threat gate treated 22 July 2026 as SK Hynix’s Q2 report date. The issuer’s own invitation and 6-K place the earnings release and call on 29 July 2026 at 09:00 Seoul (Q2 2026 earnings-call invitation; Form 6-K, 29 July 2026). The 29 July 6-K is numbers only: revenue KRW 79.318746T, operating profit KRW 60.542608T, net profit KRW 93.922593T, provisional K-IFRS, call materials pointed at www.skhynix.com IR. It contains no LTA, deposit, or price-cap language (Form 6-K, 29 July 2026).

What SK Hynix did file on 22 July is unrelated to the cap question: the board raised the P&T7 Cheongju advanced-packaging investment to KRW 7.0931T (5.88% of 31 December 2025 consolidated equity), citing AI-memory demand and an accelerated cleanroom opening (Form 6-K, 22 July 2026, P&T7). A same-day 6-K denied a Korea JoongAng Daily report that the company was in talks to buy Intel’s Ohio campus (search-visible 6-K d139467d6k.htm, 22 July 2026 — index listed; body not re-fetched beyond the search extract). Neither filing restates LTA pricing.

Official Q2 language: LTAs kept, cap status unsaid

The newsroom release of 29 July 2026 reports record results (revenue KRW 79.3187T, operating profit KRW 60.5426T / 76% operating margin, net profit KRW 93.9226T) and states that the company “has finalized Long-Term Agreements (LTAs) with around 10 customers, including key strategic partners,” with further talks ongoing, to secure mid-to-long-term supply stability (SK hynix 2Q26 financial results). HBM4 mass shipments began in Q2; M15X mass production was pulled forward; 2026 CapEx is discussed in the call, not in the short release. The release is silent on price caps, floors, take-or-pay, and LTA mix.

The call transcript (29 July 2026) is the load-bearing primary. President of Corporate Center Song Hyeon-jong said LTA talks with around 10 customers, including key customers, are “designed to secure mid to long-term supply stability”; “specific pricing structures vary depending on the customer and product characteristics” and “are designed to address price volatility”; “financial mechanisms such as deposits are incorporated to support contract fulfillment” (SK hynix Q2 2026 transcript, stockanalysis/Quartr). In Q&A he added: contract term “normally is around five years”; “our pricing structure will also not be uniform”; the firm is discussing “a range of pricing mechanisms that can better respond to price volatility” to reduce short-term uncertainty; besides volume commitments, agreements include “mechanisms such as deposits”; “we cannot say how much of our total sales will be covered by LTAs” and will keep coverage at an “appropriate level” (same transcript). Seoul Economic Daily’s same-day recap quotes the same non-uniform / deposit language and does not add a scrap-or-keep verb (SEDaily, 29 July 2026).

Words the call did not use: “price cap,” “ceiling,” “scrapped,” “removed,” “take-or-pay.” 2027 HBM volume and pricing talks were “progressing smoothly”; contractual terms and individual-customer prices were not disclosed; conventional DRAM’s recent rise “may also have some influence” on HBM talks, but HBM pricing is “not determined solely by conventional DRAM prices” because of wafer, TSV, and packaging intensity (same transcript). That is an open negotiation, not a confirmed 2027 HBM price path.

The “scrapped caps” claim is still a July 2 secondary, not a July 29 confirmation

The claim that SK Hynix removed LTA price caps predates the print. TrendForce’s 2 July 2026 news item — titled Reportedly Removes Price Cap — attributes the story to Green Economy News “citing industry sources,” with Global Economy News adding that SK Hynix “may now be the only major memory supplier not applying a price cap.” The same item says Micron’s SCA (via Green Economy News / ZDNet, not Micron IR) set an existing-product cap at 2Q 2026 market plus a term-long floor, with HBM / DDR6 / LPDDR6 priced separately (TrendForce news, 2 July 2026). That is hearsay stacked on hearsay. The 29 July official record neither confirms nor denies it.

Seven days after that rumor, TrendForce’s own research note of 9 July 2026 still treated LTAs as caps: “several U.S.-based CSPs have entered into multi-year long-term agreements (LTAs), which restrict suppliers from raising prices for these clients”; from 3Q26, “the primary source of server DRAM price increases will shift toward customers without LTAs, as well as incremental supply sold outside LTAs” (TrendForce press, 9 July 2026). The 3 July 2026 pricing note likewise said server-DRAM gains would moderate because a portion of procurement is governed by LTAs (TrendForce press, 3 July 2026). TrendForce’s 29 July recap of the SK Hynix print then relayed Reuters-citing-analysts that conservative price increases under LTAs helped the beat-miss versus Samsung — an inference that caps were still binding, not that they had been scrapped (TrendForce news, 29 July 2026). That is still secondary.

Later 2026 primary: no SK Hynix Q3 2026 earnings release had been posted as of 19 September 2026. TrendForce’s 7 September 2026 DRAM-revenue note (2Q26 industry revenue ~US$154.73B, +59.5% QoQ) says supplier inventories remain at historic lows and 3Q26 conventional DRAM contract prices should rise 13–18% QoQ; it does not restated scrap-or-keep of SK Hynix caps (TrendForce press, 7 September 2026).

Conditional implication (unconfirmed): if LTA price caps actually break, LTA-covered hyperscalers regain a price-driven reason to pull forward and the HBM book loses that insulation. Official 29 July language is variable / non-uniform pricing plus deposits. That is a softer seam than either “caps intact (TrendForce 9 July)” or “caps scrapped (Green Economy News via TrendForce 2 July).” It does not by itself prove the kill-switch fired.

Micron’s 10-Q names ceiling/floor bands — for Micron, not SK Hynix. Do not invent MU book-share.

Micron’s Form 10-Q for the quarter ended 28 May 2026 (filed 24–25 June 2026) is the only issuer primary in this pass that uses ceiling/floor words. Strategic customer agreements are take-or-pay with binding multi-year volumes. “Pricing for most agreements is either fixed, or is subject to minimum and maximum pricing.” “The largest agreements generally have a ceiling price for existing products that approximates the market price in the second calendar quarter of 2026, and a floor price through the term of the agreement.” “A minority of the agreements do not have any fixed pricing or price bands, as pricing for those agreements is subject to market conditions.” Remaining performance obligations under min-volume / min-price were approximately $5 billion as of 28 May 2026 ($422 million already in contract liabilities). Cash deposits and related financial commitments expected for agreements concluded to date: $22 billion, of which approximately $18 billion in cash deposits (Micron Form 10-Q, period ended 28 May 2026).

What the 10-Q does not disclose, and this pass does not invent: an HBM-versus-conventional-DRAM LTA book-share percentage; the share of DRAM bits under floor/ceiling bands; or a numeric split of the HBM book that is “insulated.” It reports product revenue (Q3 FY26 DRAM $31.328B, NAND $9.943B) and segment revenue (CMBU $13.769B of $41.456B that quarter — CMBU is defined as hyperscale-cloud memory plus HBM for all data-center customers, so it is not an HBM book-share). One unnamed customer was 10% of revenue for the first nine months of FY2026, primarily in CMBU (same 10-Q). Those are the disclosed figures. They are not SK Hynix LTA-cap evidence, and they are not a MU HBM-book percentage.

Micron’s HTML Q3 FY26 press page timed out in this pass (investors.micron.com); the 10-Q is the substitute primary. TrendForce’s 2 July item’s ZDNet/Green Economy News retell of Micron SCA bands is secondary to the 10-Q and should not be preferred over it.

Customer DRAM inventory 7–9 weeks: not corroborated

The 7–9 week OEM/distributor figure, the 3–5 week supplier figure, and the “cross 10 weeks” falsifier all still live on Luminix, “DRAM Cycle Mid-2026 Update” (useluminix.com is not in the vault reliability tracker and has no verified institutional track record). This pass searched TrendForce, SK Hynix IR, Micron IR/SEC, and named-desk recaps for those exact weeks.

Fetched TrendForce HTML does not print 7–9 customer weeks or a 10-week line:

  • 3 July 2026: PC OEMs still replenishing; smartphone brands more conservative; server DRAM inventory build-up into 2H26 as CPUs recover; no week count (TrendForce).
  • 9 July 2026: “gradual buildup of DRAM inventories at U.S. CSPs” from server-CPU shortage, plus incentive to build ahead of 2027 tightness; no week count (TrendForce).
  • 7 September 2026: “suppliers’ inventories remain at historic lows”; 3Q26 contract +13–18% QoQ; no customer week count (TrendForce).

TrendForce PDF DRAM Market Bulletins (29 July and 16 September 2026) were not fetched (PDF, host not on the *.gov whitelist). Search snippets of the 29 July bulletin are qualitative only: smartphone inventory weeks “elevated,” server OEM inventories “relatively low,” supplier inventories “extremely low.” That is not a 7–9 print and is not treated as fetched.

A TrendForce 2022 press note did use “7–9 weeks” for smartphone-brand inventory at end-3Q22 (TrendForce, 22 September 2022 — search-visible; not a 2026 figure). A December 2025 TrendForce-cited item used 8–10 weeks as a healthy smartphone-memory norm and below four weeks as the then-current stock (TrendForce news, 8 December 2025 — search-visible, not selected as a 2026 corroboration). Neither is a 2026 7–9 week customer-DRAM corroboration.

Micron’s 10-Q gives inventory dollars ($8.567B at 28 May 2026) and does not disclose inventory days or customer weeks (Micron 10-Q). SK Hynix’s 29 July materials do not give inventory weeks.

Verdict on the falsifier input: dram_customer_inventory_weeks > 10 cannot be scored from a corroborated 2026 number. Directionally, named industry press still says supplier stocks are very thin (TrendForce “historic lows”; search-visible KB Securities via trade press: Samsung/SK Hynix below 10 days — not fetched as a primary and not the customer-week series). That is the opposite tier from the 7–9 / 10-week customer line.

SemiAnalysis 2027 HBM pricing: paywalled, not confirmed

SemiAnalysis’s Memory Mania (6 February 2026) is marked Paid; the public lede asserts a deteriorating DRAM imbalance through 2027 but does not yield a fetchable 2027 HBM price table (SemiAnalysis / Memory Mania). The Memory Model product page advertises 2022–2027 HBM pricing as a subscription dataset (semianalysis.com/memory-model — search-visible, not treated as a fetched price). Per the brief: do not treat paywalled 2027 HBM pricing as confirmed. SK Hynix’s own 29 July answer is the governing public fact: 2027 HBM talks are open and undisclosed (transcript).

Grokipedia’s SK Hynix page is a company-history primer (HBM3E 12-high, ~36% DRAM share in H1 2025, Intel NAND deal) and does not cover 2026 LTA caps or inventory weeks (Grokipedia: SK Hynix).

Contradictions and open questions

  • Scrap vs keep vs variable. TrendForce research (3 July, 9 July) still describes LTAs as price-restricting. TrendForce news (2 July) reports SK Hynix removed caps, citing Korean trade press / industry sources. The 29 July official record says non-uniform mechanisms to address volatility and is silent on ceilings. Reuters-via-TrendForce (29 July) then attributes the estimate miss to conservative LTA pricing — which cuts against “caps scrapped.” These three cannot be collapsed into one verb.
  • July 22 vs July 29. The wiki gate treated 22 July as the Q2 report. The 22 July 6-Ks are P&T7 capex and an Ohio-rumor denial. The Q2 print/call/6-K are 29 July.
  • 7–9 weeks vs “historic lows” vs <10 days. Luminix 7–9 customer / 3–5 supplier weeks is uncorroborated. TrendForce HTML says supplier inventories are at historic lows and CSPs are building, without weeks. Search-visible KB Securities puts Samsung/SK Hynix finished-goods at <10 days. Those series are different tiers; this pass does not reconcile them.
  • Take-or-pay. Micron’s 10-Q uses the words. SK Hynix’s call does not. Do not back-propagate Micron legal terms onto SK Hynix deposits.
  • 2027 HBM price. Official: in discussion, undisclosed. Paywalled SemiAnalysis / secondary desk notes (Goldman, J.P. Morgan) are not confirmed in this pass.
  • No later SK Hynix primary after 29 July 2026 had restated LTA caps as of 19 September 2026. Next useful issuer evidence is deposit/RPO amounts, an explicit ceiling/floor restatement, or the Q3 2026 call — not another reprint of the 2 July rumor.
  • MU book-share. Not disclosed as an HBM-LTA percentage in the fetched 10-Q. Not invented here.

Provenance

Rounds run: 3 (full)

Sub-questions by round:

Round 1 (broad survey):

  1. What did SK Hynix officially disclose about LTAs, deposits, and price caps in the Q2 2026 report and 29 July call?
  2. Did any later 2026 SK Hynix / Micron / SEC primary confirm scrap vs keep of LTA price caps?
  3. Can customer DRAM inventory weeks (7–9 vs a 10-week falsifier) be corroborated from TrendForce, issuer IR, or SEC?
  4. What contract language (variable pricing, deposits, take-or-pay) exists vs the “scrapped caps” secondary claim?

Round 2 (drill-down):

  1. After 29 July, did any SK Hynix / TrendForce / SEC primary restate LTA caps as kept or scrapped? — targeting post-gate silence
  2. Can TrendForce HTML give customer DRAM inventory in weeks matching 7–9 vs the 10-week falsifier? — targeting the Luminix single source
  3. What do Micron IR / SEC filings actually say about SCA caps and inventory days? — targeting MU book-share invention risk

Round 3 (resolve remaining uncertainty):

  1. Does TrendForce’s 7 September 2026 note still treat LTAs as lock-ins, and does it publish customer inventory weeks? — targeting a later industry restatement
  2. Does the SK Hynix SEC 6-K add cap language beyond the newsroom? — targeting a parallel primary
  3. Can any named reputable source besides Luminix print 7–9 customer DRAM weeks? — targeting the falsifier input

Anchor source (Grokipedia, fetched before round 1):

  • SK Hynix — 3,017 chars extracted (245 citations on the live page) — company history and HBM leadership through 2025; no 2026 LTA-cap or inventory-week language.

URLs fetched (11 successful, 1 failed):

Round 1:

Round 2:

Round 3:

Inspected but not treated as corroboration:

  • Luminix DRAM Cycle Mid-2026 Update — unverified domain; sole source of 7–9 / 3–5 / 10-week falsifier
  • SemiAnalysis Memory Mania — marked Paid; 2027 HBM pricing not confirmed
  • TrendForce DRAM Market Bulletin PDFs (29 July 2026, 16 September 2026) — not fetched (PDF; host off *.gov whitelist)

Tools used: WebSearch, WebFetch, grokipedia.py (python3 .claude/skills/_lib/grokipedia.py fetch "SK Hynix" --max-chars 3000). X pass not enabled. Generated: 2026-09-19 20:45 UTC

Referenced by
Mechanisms
HBM pre-committed + CoWoS half-owned by Nvidia → DRAM spot price +158% → memory/packaging bottleneck downstream of foundry