Autoresearch: HBM + CoWoS bottleneck — 2026 corroboration

Generated by /autoresearch on 2026-06-01 (via /explore-chain hbm-cowos-as-binding-bottleneck). Synthesized from recent (late-2025 / 2026) web sources to corroborate or refute each step of the HBM/CoWoS binding-bottleneck mechanism, which until now rested on a single May-2026 source. Treat as raw material — review before promoting. Context: vault/projects/stock-market

Summary

The HBM/CoWoS binding-bottleneck thesis is strongly corroborated by independent 2026 sources — in several places the constraint is more extreme than the original synthesis claimed. HBM is sold out through 2026 at all three makers; Nvidia holds ~60% of TSMC CoWoS (more than the "50%+" original); the memory price surge has accelerated far beyond the cited +158%. One stated falsifier has fired: Samsung cleared Nvidia's HBM qualification and is now shipping to Nvidia at scale — which moderates (does not break) the "SK Hynix exclusive leverage" sub-claim in step 4.

Findings

HBM is sold out / pre-committed through 2026 (step 1 — confirmed)

All three HBM makers have sold out 2026 capacity to AI customers. SK Hynix: "We've sold out our DRAM, NAND, and HBM capacity for next year" (Notebookcheck). Micron's full 2026 HBM (HBM3E 12-Hi + some HBM4) capacity is sold out (TrendForce, Tom's Hardware). Samsung's 2026 HBM is sold out, preorders exceeding planned output (KED Global).

Falsifier check — "HBM bit-supply growth >40%/yr would relax the lock-in": NOT fired. 2026 bit-supply growth is constrained: ~20% for DRAM/NAND (The Next Platform); IDC sees 2026 DRAM/NAND supply growth at 16%/17%, well below historical norms; TrendForce notes 2026 capex favors process upgrades/hybrid bonding over capacity expansion, keeping bit-supply growth limited even as HBM demand grows ~70% YoY (TrendForce). Supply cannot keep pace — the lock-in tightens rather than relaxes.

Nvidia holds ~60% of TSMC CoWoS; capacity doubling but still booked out (step 2 — confirmed, refined up)

Morgan Stanley's CoWoS allocation analysis: Nvidia secures ~60% of 2026 CoWoS demand (~595k wafers, ~510k at TSMC for CoWoS-L / Rubin), AMD ~11% (~105k), Broadcom ~15% (~150k, much of it Google TPU / Meta / OpenAI), Google TPU ~90k (Astute Group, wccftech). The original "50%+ Nvidia" is corroborated and refined upward to ~60%.

Falsifier/contradiction check — "TSMC CoWoS expansion eases the bottleneck": partially valid, does not break the chain. TSMC is doubling CoWoS (~$56B capex; ~130–150k wafers/month by late 2026, ~4× the ~35k/month of late 2024) — but Nvidia's advanced-packaging lines are booked through 2027 and Nvidia still holds ~60%, so expansion is being absorbed by demand rather than easing the constraint (FinancialContent, FinancialContent / through-2027).

Memory pricing surge — supply-constrained, accelerating (step 3 — confirmed; original figure now conservative)

The decoupling from the normal DRAM cycle is confirmed and more extreme than the cited $3.76→$9.71/GB (+158%): DRAM contract prices rose ~90–95% QoQ in Q1 2026 and a further ~58–63% QoQ in Q2 2026 (NAND +70–75%) (Tom's Hardware, TrendForce). AI will consume ~20% of 2026 DRAM output; the big three (95%+ of DRAM) redirected capacity to HBM/DDR5, squeezing legacy supply (S&P Global). Lead times 32–40+ weeks; constraints projected through 2027+.

Falsifier check — "DRAM spot < $7/GB without a capacity surprise": NOT fired. Prices are at record highs and still climbing — the opposite of the falsifier.

Beneficiary stack + the Samsung-qualification update (step 4 — confirmed beneficiaries; SK-Hynix-exclusivity sub-claim weakened)

All three memory makers benefit (all sold out). Share (Counterpoint, Q3 2025 HBM revenue): SK Hynix 57%, Samsung 22%, Micron 21% (Korea-market coverage). SK Hynix remains Nvidia's lead HBM4 supplier.

Falsifier FIRED — "Samsung HBM finally clears Nvidia qualification at scale → would dilute SK Hynix's leverage." Samsung passed Nvidia's 12-layer HBM3E qualification (Sept 2025), began Nvidia HBM3E shipments in Q3 2025, and sold out 2026 HBM (KED Global); its HBM4 reportedly passed Nvidia's tests (SamMobile) and it shipped industry-first HBM4E samples (TechTimes). The original chain's tension ("Samsung has historically struggled with Nvidia qualification") is now resolved in Samsung's favor — so SK Hynix's exclusive leverage is moderated, though all three still benefit from the sold-out market. This is a calibration-worthy update.

Packaging + memory as a bottleneck parallel to foundry (step 5 — confirmed)

The implication holds: even as leading-edge wafer supply gets alternatives (Intel/Samsung foundry), CoWoS advanced packaging (TSMC-Nvidia-dominated, booked through 2027) and HBM (sold out, supply-constrained) remain binding, independent constraints — a packaging+memory story running parallel to the foundry re-rate, not redundant with it (synthesis of the CoWoS + HBM findings above).

Contradictions and open questions

• HBM share ordering (minor): Counterpoint puts Samsung (22%) ahead of Micron (21%) in Q3-2025 HBM revenue, while some coverage says "Micron overtakes Samsung" on certain metrics — the #2/#3 ordering is contested and metric-dependent; SK Hynix's lead (~57–62%) is not.
• Samsung qualification (resolved, against the chain's framing): Samsung is now Nvidia-qualified on HBM3E and reportedly HBM4 — the "SK Hynix benefits from Samsung's qual struggles" angle no longer holds. SK Hynix still leads; the dynamic is now more competitive.

Provenance

Rounds run: focused single-pass corroboration across 4 chain sub-questions (HBM pre-commitment, CoWoS allocation, memory pricing, beneficiary/qualification), driven by /explore-chain.

URLs (search + fetch): Notebookcheck (SK Hynix sold out), TrendForce (Micron sell-out; bit-supply-limited; Q2 contract prices), Tom's Hardware (HBM roadmaps; DRAM +63% Q2), The Next Platform (HBM supply can't meet demand, ~20% bit growth), Astute Group (Nvidia 60% CoWoS; SK Hynix HBM share), wccftech (Nvidia books packaging through years), FinancialContent (TSMC $56B CoWoS doubling; booked through 2027), S&P Global (AI squeezes DRAM), KED Global (Samsung passed Nvidia qual + 2026 sold out), SamMobile (Samsung HBM4 passed Nvidia tests), TechTimes (Samsung HBM4E first samples), Counterpoint via TweakTown (HBM share 57/22/21).

Tools used: WebSearch, WebFetch. Generated: 2026-06-01.