Introducing the AI Compute Supply Chain bottleneck scoreboard.
Every dollar of hyperscaler AI capex flows through an eight-layer production stack — silicon substrates, fabrication equipment, foundry, advanced chip packaging, high-bandwidth memory, compute silicon, optical interconnect, and grid power. The public market has aggressively re-rated three of those layers over the past year. The other five carry the next leg of multiple expansion. We name the firm that owns the binding bottleneck at each layer, score them on a four-dimensional criticality scale, anchor them to forward price-to-sales, and publish a concentrated positioning shortlist that refreshes each capex cycle.
When NVIDIA reports earnings, the public market reads it through a small set of names: NVIDIA itself, TSMC, ASML, and the four U.S. hyperscalers buying the chips. That set is correct — and incomplete. Every Blackwell GPU is the output of an eight-layer production pipeline. The market has aggressively re-rated upstream supply over the past year — SK Hynix up 900%, optical interconnect names up more than 1,000%, Japanese wafer-fab-equipment specialists up 200–300%. NVIDIA itself is up only 13% year-to-date in 2026.The catch-up trade is rotating into high-bandwidth memory second sources (Samsung +160% YTD, Micron +120%), power infrastructure (GE Vernova +71%, Vertiv +115%, Eaton +26%), and Chinese optical (Innolight +90% YTD). The structurally under-rerated layer is still silicon substrates — Shin-Etsu Chemical is up only 40% over the trailing year while its downstream customers tripled or more. This paper partitions the stack into eight layers, names the binding-bottleneck owner at each, scores them on a four-dimensional criticality scale, anchors them to forward enterprise-value-to-sales, and refreshes pricing live from public reporting.
Three regimes. Each carries a different forward question.
The 2024-2025 trades are largely played out. Forward leverage from here is execution, not multiple expansion.
Catch-up trades: HBM second sources, power infrastructure, optical interconnect second leg. Where 2026 capital is rotating.
Headline names + nuclear premium have cooled. Substrates the layer most under-rerated; sub-rack interconnect IC has digested.
From sand to intelligence. Click any layer.
The left column maps the eight-layer production stack from substrate inputs at the bottom to hyperscaler intelligence at the top. Power and cooling sits at Layer VIII because every layer above it depends on grid-connected megawatts; substrates at Layer I because every layer above it consumes silicon. Click any layer to drill into the firms that own its binding bottleneck. Filters above narrow by listing venue, layer status, criticality score, or shortlist tier.
Click any layer to drill down. Left rule color = severity (light = moderate · dark = binding). Right pill = pricing status. Pulled values match latest public reporting; per-firm sources in each card.
Wafer fabrication equipment
The densest cluster of monopolists in the cohort. ASML priced; Japanese WFE cluster (Lasertec +222%, Advantest +295%, Disco +147% 1Y) has rerated hugely. Forward question is whether the rerate is durable into 2027 — Tokyo Electron near 52-week high, Lasertec backlog at record.
What the live pricing changes about the thesis.
The eight-layer supply chain was framed as a structurally under-priced opportunity for U.S.-dollar investors: foreign-listed firms in Japan, Korea, Hong Kong, Shenzhen, and Stockholm owning the binding bottlenecks at multiples consistent with cyclical-commodity businesses rather than AI-cycle structural shortages. Live pricing across the cohort shows most of the first-leg rerate has already happened. SK Hynix is up nine-fold over the trailing year. Lumentum and Innolight are up roughly twelve-fold. Lasertec and Advantest tripled. Tokyo Electron is at a 52-week high. The Japanese wafer-fab-equipment cluster is no longer obscure.
The second-leg opportunity is different in character.First, the high-bandwidth memory trade has rotated to the second sources — Samsung Electronics is up 160% year-to-date on HBM3E qualification at NVIDIA and more than 30% of HBM4 share secured for 2026; Micron is up 120% YTD with high-bandwidth memory sold out through 2026. Both names lagged SK Hynix through 2024–2025 and have now compressed the gap. Second, the power infrastructure layer is the active 2026 trade — GE Vernova is up 71% YTD with a backlog accelerating from $163B toward $200B; Vertiv is up 115% YTD on a $15B backlog (+109% year-on-year); Eaton lags at +26% YTD despite parallel hyperscaler exposure. Third, the optical interconnect layer has a clearer next leg in 1.6T transceivers and silicon photonics — NVIDIA's $2B investment in Coherent in March 2026 is the signal of where the second leg comes from.
Where the original under-pricing thesis still holds.Silicon substrates is the cohort's most under-rerated link — Shin-Etsu Chemical is up only 40% over the trailing year while its downstream wafer-fab-equipment customers tripled or more. Towa, which supplies compression molding for advanced chip packaging, sits near a 52-week high but at a market cap that is structurally small for the role it plays in the Chip-on-Wafer-on-Substrate (CoWoS) packaging cycle that TSMC runs. ASMPT's Q1 2026 bookings of +72% year-on-year have not been fully reflected in the multiple given the Hong Kong liquidity discount.
The surprise.Constellation Energy is down 14% YTD despite a Q1 beat and more than 5,650 MW of hyperscaler contracts. The nuclear premium has cooled materially after the 2024–2025 hype cycle; the market is now skeptical of restart timing and power-purchase-agreement economics. Astera Labs is down roughly 53% from its late-2025 peak on AI-capex digestion fears. NVIDIA itself is up only 13% year-to-date, the smallest YTD gain of any active large-cap AI name in the cohort.
What exists, what it doesn't do, and where this slots in.
| Source | Output | Public | Layer-resolved | Forward-looking | Per-firm scored |
|---|---|---|---|---|---|
| Sell-side semiconductor research | Single-firm coverage notes | No (paywalled) | No | Single-name | No |
| SemiAnalysis | Subscription deep-dives | Partial (paywalled) | Yes | Process-focused | No |
| TechInsights / IDC / Gartner | Market-sizing reports | No (paywalled) | Partial | Reactive cycle | No |
| The Information / Bloomberg | Episodic reporting | Yes | No | Single-event | No |
| Hyperscaler 10-Ks (capex disclosure) | Demand-side signal | Yes | Partial | Quarterly | No |
| NVIDIA / TSMC earnings transcripts | Bottleneck narrative | Yes | Single-firm name-drops | Quarterly | No |
| Forward Indicators — VI | Layer-resolved bottleneck scoreboard | Yes | Yes — all 8 layers | Yes — forward multiples | Yes — per firm |
Sources: sell-side coverage notes (multiple), SemiAnalysis, TechInsights, IDC, Gartner, hyperscaler 10-Ks, NVIDIA / TSMC investor materials. Compiled May 2026.
What distinguishes this program from the six reference programs above: layer-resolved partitioning across all eight supply-chain layers with the bottleneck owner named in each, a public scoreboard with bottleneck-criticality scored on four dimensions per firm, and a forward-multiple anchor tied to FY1/FY2/FY3 consensus rather than the trailing twelve-month print that screens default to.
Five moving parts. Each is versioned. Each ships.
Layer taxonomy
Eight layers, ordered from substrate inputs (Layer I) to power infrastructure (Layer VIII). Every dollar of hyperscaler AI capex flows through all eight, making the partition operationally exhaustive. Layers are defined by the production step they own, not by SIC code. The taxonomy excludes the demand-side hyperscalers (Microsoft, Alphabet, Amazon, Meta, Oracle) from the cohort proper — they appear as the demand-side anchor above the stack but are not the analytic target.
Firm selection
For each layer the program identifies firms that meet four conditions: operational criticality (the firm supplies a production step the layer above cannot proceed without, with greater than 50% share or no current second source); public listing on a recognized exchange with audited disclosure; disclosed AI exposure (segment, backlog, named customer, or tool-shipment count); and computable valuation reference (market cap above USD 1B, daily liquidity, derivable enterprise value). Firms meeting criticality but not disclosure enter as qualitative tracking entries only.
Valuation convention
Forward enterprise-value-to-sales: market capitalization plus net debt minus cash, divided by forward consensus revenue (the next full fiscal year where available, trailing twelve months otherwise). Where firms operate diversified mixes, AI-exposed segment revenue is used if separately disclosed; full-firm revenue otherwise, with the diversification flagged on the firm row. Forward multiples are chosen over trailing because the central claim concerns the 2026–2028 order-to-revenue lag — trailing multiples systematically understate the exposure the program is measuring. Foreign exchange uses filing-date central parity for the renminbi and spot for the won, yen, Taiwan dollar, Hong Kong dollar, and euro.
Bottleneck-criticality scoring
Each firm carries a bottleneck-criticality score from zero to twelve, built from four dimensions, each scored zero to three:
- Supplier concentration — how few firms can supply this step.
- Substitution lead time — how long a downstream buyer would need to qualify a second source.
- Customer concentration — how dependent the firm is on a small number of buyers.
- Order-to-revenue lag — the gap between booked orders and recognized revenue, which is when the public market gets to see the bottleneck.
The cohort's central analytic operation is to regress forward enterprise-value-to-sales on bottleneck-criticality. The residuals — under-priced firms with high criticality, over-priced firms with low criticality — are the program's primary positioning output.
Editorial guardrails
Public filings only. Where a single-source claim has not yet been independently corroborated, the cell is marked as a snapshot pending verification. The program does not transact in price targets, does not issue buy or sell ratings, and does not assign portfolio weights.The output is a layer-resolved scoreboard with bottleneck-criticality scores and forward multiples on a single grid. Readers form their own positioning. The register is descriptive — a measurement instrument, not a recommendation desk.
What we will and will not publish.
Public documents only. Every firm in the cohort resolves to a recognized exchange listing with audited disclosure. Bottleneck-criticality scores resolve to disclosed customer concentration, named-supplier disclosures, and lead-time press disclosures. We do not use leaked decks, anonymous channel checks, or supplier-side rumour.
Forward multiples, conservatively anchored. All multiples use FY1 consensus where available; firms with FY1 missing fall back to TTM. Where FY1 consensus is materially above TTM (the power-layer norm), the program uses the lower of the two and discloses the gap rather than chasing the forward number.
No price targets, no recommendations.The program does not transact in buy/sell ratings, does not assign portfolio weights, and does not publish target prices. The output is a layer-resolved measurement instrument. The positioning shortlist names the cohort's clearest residuals; the inference is the reader's.
Listing-venue symmetry. When the program flags a foreign-listed firm as under-priced, it applies the same disclosure-quality bar to the US-listed comparable. Foreign listings do not earn a free pass; they earn the same scrutiny on supplier concentration, customer concentration, and order-to-revenue lag as the US peers above them in the layer.
No-fabrication rule. Where a multiple, share count, or criticality score is missing from the public record, the cell reads as pending verification rather than being filled with an estimate. This is the binding rule for the dossier and for the scoreboard.
Versioned cohort. Layers, firms, and methodology cut are stamped. The founding cut as of 31 May 2026 reflects live pricing pulled from public reporting. When a firm is added, removed, or re-scored, the change appears in the changelog with the source that drove it.
Notes
- NVIDIA Q4 FY2026 earnings call, 26 February 2026, naming HBM die supply as the single largest near-term constraint on Blackwell delivery cadence; CoWoS-S advanced packaging capacity at TSMC named as the second-most-cited constraint. Transcript via NVIDIA investor relations.
- SK Hynix Q1 2026 earnings release, 24 April 2026: revenue ₩52.6T (+60% QoQ), operating margin 71.5%. HBM revenue contribution disclosed above 40% of DRAM segment with HBM gross margin materially above blended DRAM margin. ASML $7.9B EUV order — largest single ASML order ever per ASML investor communications.
- ASMPT Q1 2026 earnings release, 5 May 2026: revenue +32% YoY, bookings +72% YoY, Q2 guidance +37% YoY. TCB and hybrid bonder orders tied to HBM4 ramp. TCB TAM disclosed as targeting USD 1B+ by 2027 with ASMPT positioning for 35-40% share. Source: ASMPT Hong Kong Stock Exchange filings.
- Tokyo Electron Q4 FY2026 earnings release, 28 April 2026, disclosing coater/developer market share at EUV layers above 91% (CY2025). FY27 coater/developer sales expected +50% YoY. WFE market guidance USD 150-170B for CY2026 and CY2027 each. Source: Tokyo Electron Investor Relations, FY2026 annual filing.
- GE Vernova Q1 2026 release, 22 April 2026: orders $18.3B (+71% organically), revenue $9.3B (+16%), net income $4.7B (50.9% margin). Gas turbine backlog 83GW → 100GW, targeting 110GW by year-end 2026. Total backlog $163B, targeting $200B by 2027 (pulled forward from 2028). Electrification segment data-center orders $2.4B in Q1 alone (more than all of 2025).
Full sourcing for the cohort lives in the project research dossier.