Silicon Analysts

HBM Pricing Dataset — Method Notes

hbm-pricing-method v1 · Last updated: 2026-07-28 ·

How the hbm-pricing series is constructed — what each point’s type means, the arithmetic behind the modelled values, the two measurement bases that must never be compared, and the gaps we left empty on purpose. Per-point sourcing travels with the data itself; this page covers how the series was built.

Get the data

  • HBM pricing reference page — the human-readable tables and charts
  • GET /api/v1/market-data/hbm-pricing — JSON; recent points are open, full history is Pro
  • GET /api/v1/export?dataset=hbm-pricing&format=csv — the full series as flat CSV (Pro)
  • API documentation

The hbm-pricing dataset (data/market-data-config.ts) was rebuilt 2026-07-28 from an adversarially-verified deep-research pass: 6 research agents (per generation/regime + derivation inputs), every candidate point verified by an independent skeptic against its source and unit arithmetic, then synthesized. 48 candidate points → 41 published.

The one structural finding

No public HBM spot market exists — in any generation, in any period. DRAMeXchange's live spot list carries no HBM line item; TrendForce's HBM coverage is contract-only; and capacity is pre-sold via annual/multi-year LTAs, so no uncommitted inventory reaches brokers. The prior hbm3e-spot series ("spot ≈2x contract in 2023") could not be verified in ANY external source — every search hit repeating it resolved to siliconanalysts.com's own pages (a circular echo of our own database). The spot series was retired. The only open-market evidence found: one 2026 broker ask ~$77/GB (6-8x contract — an outlier, not a market) and documented but unpriced gray-market channels into China.

Re-adding a spot series requires primary external evidence (a broker sheet, a DRAMeXchange line, a named-source transaction print) — not analyst narrative.

Point typing

  • data_type: 'Confirmed' — the source states the number (claim tier: stated).
  • data_type: 'Estimate' — derived (arithmetic from disclosed figures) or modelled (spread/anchor applied); the derivation inputs are named in each point's source_note so any point is recomputable.
  • data_type: 'Projection' + is_projection — forward-looking, incl. superseded forecasts kept deliberately (dated forecast reversals are data).

Method notes (from the synthesis)

  • STATED ANCHOR SERIES: TrendForce 'Projection on HBM Prices, 2022-2025' (Ellie Wang, FMS, Aug 5, 2025) — blended HBM ASP $1.40/Gb (2022), $1.38 (2023), $1.49 (2024), $1.80F (2025) = $11.20/$11.04/$11.92/$14.40 per GB. All-generation blends: used ONLY as anchors for modelled points, never emitted as generation points.
  • Generation back-solve method: blended ASP + TrendForce bit-mix (2023 actual: HBM2e ~50%, HBM3 ~39%, other ~11%) + generation premium k = 1.25-1.5x solves per-generation $/GB. The 2023 HBM3 mix note was corrected in verification from a misquoted '60% HBM2e' to the actual 50/39 split; result robust.
  • HBM3 launch level (H2 2022): 1.2-1.45x premium over the $11.20/GB 2022 blend — sole-supplier (SK hynix) H100 launch era.
  • HBM3 2024: Mirae Asset SK hynix all-HBM ASP ~$1.7/Gb minus the HBM3E premium (20-30%, ramped Mar 2024) isolates HBM3 at ~$1.5-1.6/Gb; direction stated by SK hynix (KED Global, May 2024).
  • HBM3 2025/2026: stated percentages on modelled bases — +5-10% (TrendForce May 2024 stated 2025 hike) then -10% to -20% (Goldman/TrendForce mature-product decline). The Dec 2025 ~20% hikes were HBM3E-specific; no source claims legacy HBM3 participated. Monitor 2026 actuals for DDR5-crowding upside.
  • HBM2E 2020-2022 rests entirely on the TechInsights Dec-2019 $120/16GB (pre-packaging) anchor plus the 2022 blended-ASP back-solve; verification raised the 2021/2022/2023 mids ($8.5→$10, $8→$10.8, $7→$9.8) because the packaged level, not the pre-packaging anchor, is what supply agreements price.
  • Per-stack conversions use 12-hi HBM3E/HBM4 = 36GB, 8-hi HBM3E = 24GB, HBM2E = 16GB, HBM2 = 4/8/16GB as stated per source. Korean per-unit figures carry unresolved 8-hi/12-hi capacity ambiguity — flagged in conversion_note wherever it moves the number.
  • Revenue/bits derivations divide TrendForce/Goldman/Micron revenue statements by TrendForce bit chains. Known ~30% inconsistency between TrendForce's growth-chain bits and share-based bits is carried as the low/high range rather than hidden.
  • PER-STACK vs REVENUE-IMPLIED CONFLICT (~1.7x): Korean-press per-stack figures (HBM3E 12-hi ~$300-350 = $8.3-9.7/GB) vs Goldman/revenue-implied ~$14-15/GB. Both retained as separately labeled points; do not average them.
  • 2026 HBM3E cell merges three modelled variants (hike on $350 base = mid, hike on $300 base = low, hike on Goldman $15 base = high) into one point per the duplicate-merge rule; Samsung renewal and superseded Goldman projection kept as separately labeled points.
  • Forecast-history preservation: Goldman's Nov 2025 -28% 2026 HBM3E projection is retained labeled '(proj., superseded)' — the Jul-to-Dec 2025 regime whipsaw (double-digit-drop view → ~20% hikes on H200-China/ASIC demand) is itself a datum.
  • SEED CORRECTIONS (rule 4): HBM4 2026 headline $14 → $15.5 (band $13.9-16.7) on the verified $500→$560→>$600 per-stack progression; HBM3E blended 2026 $11 → $12 (band $10-14) after the verified Dec 2025 hikes superseded the decline view. All other seed points carried unchanged.
  • SPOT SERIES CAVEAT: extensive external search found NO public HBM spot market in any generation or period (DRAMeXchange spot list carries no HBM line; TrendForce HBM coverage is contract-only; capacity pre-sold via LTAs). The seed hbm3e-spot and hbm3-spot points are carried but downgraded to modelled/Low and flagged internal-unverified — every external echo of the 'spot 2x contract' narrative traces back to siliconanalysts.com itself. Recommend retiring or relabeling the spot series as 'gray-market estimate (internal)'.
  • CIRCULARITY EXCLUSION: siliconanalysts.com pages now rank for these queries and third parties echo our figures; all points here were sourced externally, and search results resolving to our own pages were excluded as corroboration.
  • Claim-type convention: per-stack prices stated by a source with $/GB computed by us = 'derived' (applied consistently; two HBM4 points adjusted stated→derived in verification). Quarterly granularity finer than half-years is not emitted for contract series: HBM prices reset via annual LTA negotiations (~12+ months ahead), so quarterly contract series would be fictional precision.
  • is_projection semantics: only 2026 forward-looking points carry '(proj.)' labels; 2026 points based on already-negotiated contract settlements (SK Hynix $560 agreed, Samsung renewal) are actuals for 2026 delivery, not projections.

Deliberate gaps (honest holes beat fabricated points)

  • HBM2 spot, HBM2E spot (2020-2026): left empty — no HBM spot market ever existed; DRAMeXchange/TrendForce never published HBM spot quotes; HBM sold only via negotiated supply agreements. The only spot-like HBM2E activity (Dec 2024 China bulk-buying) is captured as a blended derived point, not a spot series.
  • HBM3 spot beyond Q1 2023: left empty — targeted search found no external evidence of broker/gray-market HBM3 transactions in any period; the lone seed Q1-2023 print is retained but flagged internal-unverified.
  • HBM3E spot 2026: left empty — no spot market emerged; the only open-market ask found (broker, $77/GB) sits 6-8x above contract and is an outlier, not a market. HBM3E moves ~100% via LTAs.
  • HBM4 spot (all periods): left empty — no HBM4 spot market exists; all volume under annual/multi-year LTAs with NVIDIA/AMD/hyperscalers; no independent broker quote found anywhere.
  • HBM2 2022-2026: left empty — end-of-life product; TrendForce stated 2023 price declines but published no level, and the demand base was vanishing; a modelled point would be fabrication.
  • HBM2E 2025-2026: left empty — demand collapsed after the Dec 2024 US export controls closed the China channel; no post-2024 pricing anchor of any kind exists.
  • HBM3 2020-2021: left empty — product did not exist (SK hynix HBM3 mass production began Jun 2022).
  • HBM3E before Q2 2023: left empty — product not yet in negotiation/production (first shipments Q1 2024; seed series starts at the Q2-2023 negotiation era).
  • HBM4 before 2026: left empty — sampling only, no commercial pricing existed; first reported price event is the Aug 2025 initial quote for H1 2026 supply.
  • 8-hi vs 12-hi $/GB differential (all generations): deliberately not modelled — no credible public source quantifies it as of Jul 2026; ambiguity flagged in conversion notes instead.
  • Quarterly contract granularity finer than half-years: intentionally not emitted — annual LTA resets make it fictional precision; vendors disclose no quarterly HBM bit shipments, so quarterly implied $/GB is not publicly computable.

Basis discipline (added Jul 2026, after external adversarial review)

The dataset carries HBM3E/HBM4 contract prices on two measurement bases that differ by ~1.7×, and they must never be compared across series:

  • Revenue-implied / analyst basis (hbm3e-contract): the total economic value exchanged under LTAs — capacity guarantees, yield underwriting, engineering support. Goldman-class estimates and revenue÷bits derivations live here (~$13–17/GB for 2024–26 HBM3E).
  • Per-stack press basis (hbm3e-contract-stack, hbm4-contract): Korean-press prints of per-stack negotiation figures ÷ stack capacity — closer to a factory-gate component price (~$7–14/GB). Excludes the burdens above.

Rule: cross-series comparisons must be same-basis. An external adversarial review pass (Jul 2026) caught a violation we introduced: the 2026 HBM3E point (revenue-implied, $15.6) briefly sat above the 2026 HBM4 projection (per-stack, $15.5) — an artificial generational inversion. Fixed by adding the per-stack 2026 HBM3E blend ($12.2 [11.1–13.9]: SK $350×1.2= $420/36 + Samsung $500/36, both from this dataset's own verified prints), which restores the same-basis HBM4 premium of +27–37%. A regression test now pins the invariant.

Relocated cross-check: the HBM4 revenue÷bits floor

A former hbm4 blended point ($12.5/GB [12–13.1]) was removed from the chartable dataset as orphaned: Goldman's 2026 HBM TAM ($45B) ÷ TrendForce bits (>30B Gb = 3.75B GB) = $12.0/GB, with a Micron-implied variant at $13.1 — but both TAM inputs predate the Dec-2025 renewal hikes and the HBM4 premium, so the figure reads as a stale floor, not a market level. It remains a useful sanity floor: any modelled 2026 blend materially below ~$12/GB would contradict even the pre-hike TAM arithmetic.

The rejected fix

The review's alternative proposal — normalizing HBM4 to the revenue-implied basis at $15.6 × 1.30 ≈ $20.3/GB — was rejected: it is a two-hop model (modelled base × assumed premium factor) with no external source near that level. The dataset does not publish revenue-implied HBM4 until an external anchor exists.

Circularity guard

Silicon Analysts pages now rank for HBM-pricing queries and third parties echo our figures back. Research passes MUST exclude siliconanalysts.com (and unattributed reproductions of our numbers) as sources — this pass did, and future refresh agents should too.