— C+

MHP · Index 3 of 4 · AI Build-Out Family · monthly cadence · 11 tickers · 4 layers

Memory / HBM Pulse

HBM is the bottleneck. Every H100, H200, B200, and MI300X needs 4-12 stacks of high-bandwidth memory glued onto the logic die. Three companies make HBM at scale — SK Hynix, Samsung, Micron — and when their capacity sells out, the AI accelerator ceiling is set. MHP tracks the three layers that move first (test-equipment orders, advanced packaging, HBM-supplier momentum) plus the demand-pull layer that confirms the read.

Composite 81 / 100 Accelerating
Data month May 2026
Snapshot taken May 26, 2026
Universe 11 tickers across 4 layers · HBM3 / HBM3E volume coverage
Korea-exports macro signal pending (free public endpoints unreliable — MVP2 will add Korea Customs unipass API)
Lead-vs-Coincident Divergence

Bearish setup. HBM Big-3 (score 100) is running 20+ points hotter than the memory test/equipment layer (score 43). Test orders normally lead HBM revenue by 2-4 months — when Big-3 rallies without the lead layer confirming, the next leg may already be priced in.

History · momentum Building history

History starts 2026-05 — the series fills as each day is archived.

1 point · since 2026-05

Layer breakdown

Each layer captures one stage of the HBM value chain. Score is the percentile rank of average 21d return vs the layer's rolling 24-month history (first months use a deterministic heuristic). Weights reflect distance from the actual HBM revenue print — Big-3 closest, demand-pull furthest.

HBM Big-3 100 40% weight
21d avg: +47.0% tickers: 3

SK Hynix (HBM #1, ~50% share), Samsung (~35%), Micron (~15%). The three companies that actually make HBM. When they say "sold out 2027", that is the AI capex ceiling.

Memory Test / Equipment 43 25% weight
21d avg: -2.9% tickers: 3

Advantest (memory test market leader, ~50% share), Teradyne, Aehr Test Systems (HBM burn-in specialist). Test-equipment order books lead HBM revenue by 2-4 months — bullish divergence when this layer rises while Big-3 lags.

HBM-Adjacent Packaging 79 25% weight
21d avg: +11.8% tickers: 3

TSMC (CoWoS provider that mates HBM stacks to logic die), ASE Technology, Amkor. The packaging bottleneck — CoWoS capacity is one of the constraints on HBM-on-GPU output.

HBM Demand-Pull 100 10% weight
21d avg: +30.5% tickers: 2

NVIDIA, AMD. The actual buyers of HBM. Their order behaviour validates whether the supply-side signals are translating into accelerator volume.

Per-ticker breakdown 11 tickers · sorted by layer

Ticker Company Layer HBM capability 21d return
MU Micron Technology HBM Big-3 HBM3E, HBM4-sample +55.9%
000660.KS SK Hynix HBM Big-3 HBM3, HBM3E, HBM4-sample +54.8%
005930.KS Samsung Electronics HBM Big-3 HBM3, HBM3E, HBM4A-sample +30.3%
6857.T Advantest Memory Test / Equipment — +2.4%
AEHR Aehr Test Systems Memory Test / Equipment — -0.5%
TER Teradyne Memory Test / Equipment — -10.6%
3711.TW ASE Technology HBM-Adjacent Packaging — +33.5%
2330.TW TSMC HBM-Adjacent Packaging — +11.6%
AMKR Amkor Technology HBM-Adjacent Packaging — -9.8%
AMD AMD HBM Demand-Pull — +53.1%
NVDA NVIDIA HBM Demand-Pull — +7.9%

HBM4 readiness MVP2 trigger

MVP1 universe is the HBM3 / HBM3E generation — what is actually shipping in revenue in 2026. HBM4 is sample-only as of writing: SK Hynix shipped first HBM4 samples Q1 2026, Samsung's HBM4A variant follows in H2 2026, Micron's HBM4 samples land Q3 2026. Mass production targets late-2026 / early-2027.

MHP MVP2 will add an HBM4-transition sub-layer when SK Hynix or Samsung discloses an HBM4 revenue line for the first time. At that point the cross-generation pricing ratio (HBM3E / HBM4) becomes the leading derivative — directly analogous to CCST's H100 / B200 ratio, with the same regime-shift-detector behaviour. Every ticker in the MHP universe already carries an hbmCapability tag so MVP2 plug-in is mechanical, not a rebuild.

AI Build-Out Index Family 4 indices · supply-side AI hardware tracking

CCST · TAISP · MHP · STP — four independent indices, each with its own data source, equity basket, and equity-gap signal. Read them together: when one moves before the others, that's the rotation signal. (Agentic-demand-side tracking lives in the separate Agentic Demand Index.)

72.8 / 100 Tight

AI Build-Out Index — equal-weighted average of the 4 live indices below (4 of 4 reporting). Each index counts once; one with no current reading drops out rather than being imputed.

Methodology

MHP is the third of the Closelook AI Build-Out Index Family — a 4-index suite tracking the AI hardware supply chain (Compute Spot Tightness · Taiwan AI Supply Pulse · Memory/HBM Pulse · Semicap/Test Pulse). Each index has its own data sources, equity basket, and equity-gap signal. Closelook Agentic Usage Pulse (AUP), tracking demand-side AI adoption, is a separate standalone index in its own track.

Equity layer: 21d returns from the Closelook Lake API for each of 11 tickers across 4 layers. Per-layer score = percentile rank of average 21d return vs rolling 24-month history. Composite = layer-weighted blend (HBM Big-3 40% / Memory Test 25% / HBM-Adjacent Packaging 25% / HBM Demand-Pull 10%), with an optional 25% allocation to a Korea memory-exports YoY layer when the public macro-stats endpoints are reachable.

Korea-exports layer attempts OECD SDMX and FRED public endpoints in MVP1. Both are best-effort — when unreachable, composite re-normalises across the four equity layers without dropping. MVP2 will add Korea Customs Service unipass.customs.go.kr integration for the actual memory-segment-disaggregated first-20-days release that drops the 21st of every month.

The lead-vs-coincident divergence flag is the actionable signal: Memory Test layer leads HBM Big-3 by 2-4 months historically. When test runs 20+ points hotter than Big-3, the supply chain is preparing for delivery the Big-3 haven't booked yet (bullish). When Big-3 runs 20+ points hotter than test, the next leg may already be priced in (bearish).

Index reading is a diary entry on our process. Closelook publishes research, not investment advice.