AI Supply-Chain Investment Map: Public Companies and Potential 10X Winners

Research date: June 4, 2026 | OSINT-based equity research across the full AI value chain

Important disclaimer. This document is OSINT-based research and educational analysis, not investment advice. I am not a financial advisor or a registered investment professional. Nothing here is a recommendation to buy or sell any security. All figures are drawn from public sources (SEC filings, earnings releases, investor presentations, and credible reporting) dated through early June 2026 and are subject to change; market caps and valuations move daily and should be re-verified before any decision. “10X” scenarios are deliberately aggressive thought-experiments, not forecasts - the base rate of a public company 10X-ing in 24 months is extremely low. Do your own due diligence and consider consulting a licensed advisor.

Companion tools - Explore the interactive 115-company dashboard (sortable scores, the supply-chain map, rankings, watchlist, and risk controls) or download the full Excel scoring model.

Executive Summary

The AI buildout entered 2026 as the largest private capital-deployment cycle in history. At Q1-2026 (calendar) earnings in late April, the four largest US hyperscalers - Microsoft, Amazon, Alphabet, and Meta - collectively guided to roughly $725 billion of 2026 capex, up ~77% from ~$410B in 2025, with Big Tech capex now seen topping $1 trillion in 2027 (Tom’s Hardware; CNBC). Every major spender raised guidance; the only debate is payback timing. NVIDIA’s Q1 FY2027 (reported May 20, 2026) confirmed the demand with $75.2B data-center revenue, +92% YoY (Intellectia).

Three binding constraints now define the cycle and, therefore, where the money pools into pricing power:

1. Memory. A ~15-year-worst shortage. DRAM contract prices rose ~58-63% QoQ and HBM is sold out industry-wide; memory is now ~30% of hyperscaler data-center spend, up ~4x from 2023 (Tom’s Hardware). Microsoft attributed ~$25B of its 2026 budget purely to memory/component inflation.
2. Power. US data-center demand is forecast to roughly triple by 2030 (~106 GW by 2035, BloombergNEF). Gas-turbine slots are sold out to ~2029; transformer lead times run 128-144 weeks. Power is the hardest physical bottleneck.
3. Advanced packaging & leading-edge silicon. TSMC is roughly quadrupling CoWoS capacity (~35k → ~127-130k wafers/month by end-2026), with NVIDIA holding >60% of it.

The structural conclusion: the most durable economics sit not at the most famous name (NVIDIA) but at the chokepoints - CoWoS/foundry (TSMC), HBM (SK Hynix, Micron), EUV (ASML), liquid cooling and electrical gear (Vertiv, Eaton), gas turbines (GE Vernova), optical interconnect (Coherent, Lumentum, Celestica), and custom silicon (Broadcom, Marvell). The biggest percentage upside lives in mid-caps levered to a single tightening bottleneck, while the megacaps are quality compounders that mathematically cannot 10X in 24 months.

Highest-conviction risk-adjusted names (full reasoning in the rankings and watchlist sections): Broadcom (AVGO), TSMC (TSM), and Vertiv (VRT) among large/mega-caps; Celestica (CLS), Coherent (COHR), GE Vernova (GEV), Credo (CRDO), and - from this edition’s expanded coverage - Advantest (ATEYY) and Kulicke & Soffa (KLIC) in test/packaging and Siemens Energy (ENR) and ABB in power/electrical, among the higher-beta breakouts; and a clearly-labeled speculative basket (Oklo, NuScale, Nebius, CoreWeave, plus new adds Aehr Test (AEHR) and the AI-datacenter hosts IREN/APLD/WULF/CORZ) for highest-risk/highest-upside capital. The single biggest portfolio risk is that all of these names are correlated to one variable - the durability of hyperscaler capex - so position sizing and the “what invalidates the thesis” triggers matter more than stock selection.

What’s new in this edition (June 4, 2026 coverage review): a second-pass sweep added 17 AI-supply-chain companies that were missing from the prior cut - across energy/electrical (NextEra, Siemens Energy, ABB, Emerson), semis & equipment (GlobalFoundries, MACOM, Lattice, Aehr Test, Kulicke & Soffa), AI-datacenter hosts (Applied Digital, IREN, Core Scientific, TeraWulf), and software (Dynatrace, Elastic, UiPath, and - for completeness, with a caution - the contracting C3.ai). The universe is now ~115 scored names, and a dedicated Top 10X Candidates section isolates and pressure-tests the highest-probability multibaggers.

Interactive scoring dashboard

Sort and filter all 115 companies across the 11 scoring dimensions, and explore the supply-chain map, rankings, watchlist, and risk controls without leaving the page.

Open the dashboard full screen · Download the Excel model

Methodology and OSINT Sources

Approach. Eight parallel research streams swept the full value chain (Big Tech capex; semiconductors; foundry/equipment/EDA; memory/storage; data-center physical infrastructure; networking/optical; energy/utilities; AI software/edge/neoclouds). Each company was assessed on a fixed 15-field schema (ticker, layer, AI exposure, revenue trend, margins, balance sheet, customers, moat, valuation risk, execution risk, dependency risk, catalysts, downside risks, and OSINT evidence). Numbers were pulled from primary filings wherever possible and corroborated across multiple outlets.

Source hierarchy used (most-to-least weighted):

• Primary filings - SEC 8-K/10-Q/10-K/6-K (earnings releases, CFO commentary, balance sheets). These anchor all load-bearing financials (e.g., NVIDIA, AMD, Broadcom, Marvell, Micron, TSMC, Vertiv, GE Vernova, CoreWeave filings).
• Company IR & official press releases - capex guidance, deal terms, partnership announcements (OpenAI/Stargate, Anthropic-Amazon/Google, NVIDIA optics investments).
• Earnings-call transcripts - Motley Fool, AlphaStreet, company webcasts.
• Credible trade & financial press - CNBC, Fortune, FT-sourced compilations, Utility Dive, POWER Magazine, DataCenterDynamics, Tom’s Hardware, DIGITIMES, LightCounting.
• Market-data aggregators - companiesmarketcap.com, stockanalysis.com, macrotrends.net for caps/multiples (snapshots; these move daily).

Scoring model. Each scored company receives 1-10 ratings on eleven dimensions (defined in the scoring section). Scores are directional (10 = most attractive) and reflect judgment applied to the OSINT evidence, not a mechanical formula. A composite is provided but rankings also use qualitative overlays. Data-quality flags are noted throughout where sources diverged (e.g., CoreWeave/Nebius market-cap methodology, HBM share splits, Broadcom AI-revenue quarter labeling).

Known limitations. Valuations are early-June 2026 snapshots. Private companies (OpenAI, Anthropic, Databricks, xAI, Crusoe) are not investable directly and appear only as demand drivers. Forward statements (10X conditions) are scenario analysis. Export-control, FERC co-location rulings, and the memory cycle are live variables that can move fast.

Big Tech AI Spending: Where the Money Goes

Every major spender raised 2026 guidance at Q1-2026 earnings (reported late April 2026). The aggregate is now ~$725B for the Big 4 and ~$680-720B+ including Oracle, nearly double 2025.

AI-lab demand underneath the capex (the “why”):

• OpenAI / Stargate: ~7 GW planned, >$400B committed within three years. Oracle ~$300B (5 yrs); NVIDIA LOI ≥10 GW of systems + up to $100B investment; AMD 6 GW (~$90B); Broadcom 10 GW of custom accelerators; AWS $38B/7 yrs. OpenAI raised ~$110-122B in early 2026 at an ~$800B+ valuation (NVIDIA; OpenAI).
• Anthropic: revenue run-rate surpassed $30B in 2026 (from ~$9B end-2025). Amazon investing up to $25B + up to 5 GW Trainium (Project Rainier); Google up to $40B + up to 1 million TPUs; Google+Broadcom multi-GW next-gen TPU from 2027 (Anthropic; DataCenterDynamics).

The flow in one sentence: Big Tech capex → (a) merchant GPUs/custom ASICs (NVIDIA, AMD, Broadcom, Marvell) → fabbed by TSMC on ASML tools, packaged in CoWoS, fed by HBM (SK Hynix/Micron/Samsung) → racked and connected by optical/networking (Arista, Coherent, Celestica) → housed in data centers cooled and powered by Vertiv/Eaton/GE Vernova drawing on nuclear/gas (Constellation/Vistra/Talen) → consumed by the software/agent layer (Palantir, ServiceNow, Dynatrace, Elastic) and neoclouds / AI-datacenter hosts (CoreWeave, Nebius, Oracle OCI, and the build-to-suit hosts IREN, Applied Digital, Core Scientific, TeraWulf).

AI Supply Chain Layer Map

flowchart TD
    SPEND["💰 BIG TECH AI CAPEX 2026<br/>~$725B Big 4 · ~$720B+ incl. Oracle<br/>MSFT $190B · AMZN $200B · GOOGL $185B<br/>META $135B · ORCL $50B · TSLA $25B"]

    SPEND --> CLOUD["☁️ CLOUD / NEOCLOUD BUILDOUT"]
    SPEND --> LABS["🧠 AI LABS (private demand)<br/>OpenAI · Anthropic · xAI"]

    CLOUD --> HYP["Hyperscalers & AI Clouds<br/>MSFT · AMZN · GOOGL · META<br/>ORCL · CRWV · NBIS<br/>DC Hosts: APLD · IREN · CORZ · WULF"]
    LABS --> HYP

    HYP --> CHIPS["🔲 CHIPS & ACCELERATORS"]
    CHIPS --> MERCH["Merchant GPU<br/>NVDA · AMD"]
    CHIPS --> ASIC["Custom ASIC / XPU<br/>AVGO · MRVL"]
    CHIPS --> CONN["Connectivity / Control Silicon<br/>ALAB · CRDO · LSCC"]

    MERCH --> FAB["🏭 FOUNDRY & PACKAGING"]
    ASIC --> FAB
    CONN --> FAB
    FAB --> FOUNDRY["Foundry / CoWoS<br/>TSM · INTC(foundry) · GFS"]
    FAB --> EQUIP["Equip / EDA / Test / Packaging<br/>ASML · AMAT · LRCX · KLAC<br/>TOELY · CDNS · SNPS · KLIC · AEHR"]

    FOUNDRY --> MEM["💾 MEMORY / HBM"]
    MEM --> HBM["HBM4 / DRAM / NAND<br/>SK Hynix · MU · Samsung"]
    MEM --> STORAGE["Storage for AI data lakes<br/>WDC · STX · SNDK · PSTG"]

    FOUNDRY --> NET["🔗 NETWORKING & OPTICAL"]
    NET --> SWITCH["Ethernet / Switching<br/>ANET · CSCO · CLS"]
    NET --> OPTICS["Optics 800G→1.6T / CPO<br/>COHR · LITE · FN · CIEN · MTSI"]
    NET --> INTERCO["Connectors / Fiber<br/>APH · GLW"]

    HYP --> POWER["⚡ POWER · COOLING · ELECTRICAL"]
    POWER --> COOL["Cooling / Thermal<br/>VRT · NVT · MOD · CARR · JCI · ASPN"]
    POWER --> ELEC["Electrical / UPS / Switchgear<br/>ETN · SBGSY · PWR · CMI · GNRC<br/>ABB · EMR"]

    POWER --> ENERGY["🔌 ENERGY & UTILITIES"]
    ENERGY --> NUKE["Nuclear / Restarts / SMR<br/>CEG · VST · TLN · CCJ · SMR · OKLO"]
    ENERGY --> GASGRID["Gas Turbines / Grid / Fuel Cells<br/>GEV · ENR · NEE · NRG · BE · D · SO"]

    HBM --> CHIPS
    HBM --> APP["🤖 AI SOFTWARE & APPLICATION LAYER"]
    SWITCH --> APP
    COOL --> APP
    NUKE --> APP

    APP --> SW["Software / Security / Data<br/>PLTR · NOW · CRWD · DDOG · DT<br/>SNOW · MDB · NET · ESTC · PATH"]
    APP --> EDGE["Edge AI & Robotics<br/>TSLA(Optimus) · SYM · ROK · MBLY"]

    SW --> END["👥 END CUSTOMERS<br/>Enterprises · Government · Consumers"]
    EDGE --> END
    END -.->|"Revenue funds next capex cycle"| SPEND

The dashed loop is the central bull (and bubble) question: end-customer AI revenue must eventually grow into the capex to keep the cycle self-funding.

Company-Level Research by Layer

All financials are the latest reported quarter as of June 4, 2026. Market caps are early-June snapshots (move daily). “Conc. risk” = customer-concentration risk (H/M/L). Valuation multiples are trailing unless noted.

Layer 1 - Semiconductors & Accelerators

NVIDIA (NVDA) - the anchor, but math caps the upside. Record $81.6B revenue, “sold out” Blackwell, networking +199% YoY, Vera Rubin into full production fall 2026 targeting $1T cumulative Blackwell+Rubin orders by 2027 (SEC CFO commentary). At ~$5.2T and P/E ~34 it is paradoxically the least valuation-stretched megacap in the group - but a 10X would require a ~$52T market cap, which is not a serious 24-month scenario. NVDA is a quality compounder and the sector’s demand barometer, not a 10X candidate.

Broadcom (AVGO) - the best large-cap risk/reward in silicon. AI semi revenue +143% to $10.8B with Q3 guided to ~$16B (>200% YoY), six core custom customers now including Google, Meta, Anthropic, and OpenAI (the ~$10B order), and a reiterated >$100B AI revenue ambition on a $73B backlog (SEC 8-K; Fool transcript). Custom ASIC is growing faster than merchant GPU as hyperscalers seek NVIDIA alternatives - AVGO and MRVL are the structural share-gainers. The VMware software slip (+9%) is the watch item.

Layer 2 - Foundry, Equipment & EDA (the chokepoint)

TSMC (TSM) - the indispensable, and oddly cheap, chokepoint. Everything in Layer 1 is fabbed here. Q1 revenue +40.6% YoY, HPC 61% of sales, and AI-accelerator long-term revenue CAGR raised to 56-59%, with CoWoS capacity going from ~35k to ~127-130k wafers/month by end-2026 and NVIDIA locking >60% of it (MacroMicro; FinancialContent). At a forward P/E of ~25 it is among the lowest-valuation-risk names in the entire chain; the offsetting risk is Taiwan geopolitics, a genuine tail risk that justifies the discount.

Layer 3 - Memory & Storage (the new #1 cost line)

The memory supercycle is the cleanest fundamental story in the report. Worst shortage in ~15 years; DRAM up ~58-63% QoQ, HBM consuming ~23% of DRAM wafer output and sold out across SK Hynix, Micron, and Samsung; HBM TAM ~$35B (2025) → ~$54.6B (2026) → ~$100B (2028) (Fortune; Futurum). SK Hynix (72% operating margin, ~7x forward earnings) and Micron (only US HBM, ~11x forward) screen as the best growth-versus-valuation combinations in the entire chain - the key risk is that memory is cyclical and “this too shall pass” if AI capex digests. Data-quality flag: HBM4 supplier-share splits vary by source; Micron’s filed FQ2 GM is 74.4% (some secondary outlets cited a stale ~45%).

Layer 4 - Data Center Infrastructure (power, cooling, electrical)

The binding physical constraint is electrical, and it is quantified. Power-transformer lead times run 128 weeks, generator step-up units 144 weeks, with some high-capacity units quoted at 3-5 years and a modeled ~30% transformer supply deficit (POWER Magazine). Meanwhile GB300 racks now run 120-130kW and Rubin Ultra targets ~600kW, forcing liquid cooling to the default. Vertiv (VRT) is the purest beneficiary - a $15B backlog up 109%, ~2.9x book-to-bill, investment-grade and S&P 500-added (SEC 8-K) - but at ~80x trailing earnings the multiple already prices hyper-growth. Eaton offers the same theme with a diversification cushion at ~41x.

Layer 5 - Networking & Optical Interconnect

The $4B NVIDIA optics bet (March 2, 2026: $2B into Coherent, $2B into Lumentum) is the single most important catalyst in this layer - it locks EML/transceiver capacity, de-risks two balance sheets, and signals optics as AI’s next bottleneck (NVIDIA). The structural shifts to watch are Ethernet overtaking InfiniBand in AI back-ends (helps Arista/Celestica) and the 800G→1.6T transceiver ramp (helps Coherent/Lumentum/Fabrinet). Celestica is the highest-growth ODM but carries the most extreme single-customer concentration (~36%).

Layer 6 - Energy & Utilities (the hardest bottleneck)

Power is where demand most clearly exceeds near-term supply. US data-center demand is forecast to ~triple by 2030 / ~106 GW by 2035 (Utility Dive). GE Vernova is the cleanest “arms dealer” - it sells turbines to every IPP/utility building gas, with pricing power from sold-out 2029 slots and a $150B+ backlog (GE Vernova). The named hyperscaler PPA map is now dense (Microsoft→Constellation, Amazon→Talen/Vistra, Meta→Vistra/Oklo, Oracle→Bloom). The swing regulatory variable is FERC’s treatment of behind-the-meter co-location (it rejected Talen’s original structure). Valuations have re-rated hard - CEG/VST/GEV at ~25-30x are the “reasonable” growth multiples; TLN (~77x), CCJ (~95x), and the pre-revenue SMRs (Oklo, NuScale) are where the binary, highest-risk upside lives.

Layers 7-9 - AI Software/Security/Data, Edge/Robotics, Neoclouds

Software splits into the richly-valued operational-AI leader (Palantir, re-accelerating to +85% but at a P/S ~45 that already fell ~26% YTD on multiple compression) and the more reasonably-valued agentic-automation compounders (ServiceNow at a ~29x forward P/E with Now Assist ACV tracking 50% above target is the standout quality name; MongoDB at ~12.5x P/S is the cheapest). Neoclouds show the most explosive headline growth and the most extreme risk: CoreWeave’s $99.4B backlog is real, but so is its ~$25B debt and Microsoft 67%-of-2025-revenue concentration plus circular NVIDIA/OpenAI financing (CNBC; 247WallSt). Nebius is better-capitalized ($9.3B cash, NVIDIA-backed) but trades at a P/S of ~42-67. Data-quality flag: CoreWeave (~$60-65B) and Nebius (~$36.5-70B) market caps vary widely by share-count methodology and date.

10X Winner Evaluation - Scoring Model

How to read this. Each company is scored 1-10 on eleven dimensions, where 10 is always the most attractive (so for “customer-concentration risk,” 10 = most diversified / lowest risk; for “valuation attractiveness,” 10 = cheapest relative to growth). Scores are judgment applied to the OSINT evidence, not a formula. The Composite (sum, max 110) deliberately rewards quality and therefore skews toward mega-caps that cannot 10X - so it is a measure of conviction/quality, not upside. For genuine 10X potential, read the 10X column on its own; it is low for almost everyone because a public-company 10X in 24 months is rare and usually requires a sub-$30B cap or pre-revenue optionality.

Legend: AIX = AI revenue exposure · GRW = growth acceleration · OPL = operating leverage · TAM = market-size expansion · MOAT = competitive moat · CONC = customer diversification (10 = lowest concentration risk) · VAL = valuation attractiveness · BS = balance-sheet strength · BOT = supply-chain bottleneck importance · MEX = probability of multiple expansion · 10X = probability of 10X in 24 months.

Note how the highest Composite names (NVDA, TSM, SK Hynix) score lowest on 10X - they are the quality anchors. The highest 10X scores (Oklo 7, NuScale 6, Nebius/CoreWeave/Credo/Lumentum/Celestica/Bloom 5) are concentrated in small/mid-caps and pre-revenue optionality, and they pair high upside with low balance-sheet/concentration scores. That tension is the entire point of the exercise.

Expanded & Overlooked Universe (the nodes people miss)

Beyond the headline names, the chain runs through layers the market under-prices: who assembles the GPUs into systems, who tests the chips, who makes the substrates and power silicon, who enriches the fuel, who pours the concrete, and who sells the AI software. These are added below with the same schema and then scored.

Layer A - AI Servers / ODM-OEM / EMS (thin margins, heavy concentration)

Layer B - Semi Materials, Test & Substrates (the hidden bottlenecks)

Layer C - Power Semis, Connectors & Data-Center REITs (the 800V transition + the buildings)

Layer D - Enrichment, Switchgear & DC Construction (fuel, gear, concrete)

Layer E - AI Software, Ad-Tech & Cyber (the application edge)

Scoring Model - Expanded Universe

Same legend (10 = most attractive; CONC 10 = least concentrated). Sorted by Composite.

The standout from the overlooked set is Advantest (78) - an HBM/AI-test near-duopoly with ~47% operating margins that screens with the third-highest composite tier in the entire ~115-name universe, behind only NVDA/TSM/SK Hynix. Other high-quality overlooked names: Siemens Energy (74) - the highest-scoring of the June coverage-review additions, a GE Vernova-class turbine/grid play with a record €154B backlog - plus Teradyne, Kulicke & Soffa (72) (TCB packaging), Besi (hybrid bonding), EMCOR (DC construction), ABB (71) (DC electrification), and Sterling Infrastructure (DC site work). Salesforce screens well on a cheap multiple but scores a 1 on 10X (too large). The highest 10X-optionality names in this set are Aehr Test (6) - a pure-play AI-burn-in newcomer - alongside Navitas (6), and a cluster at 5: the AI-datacenter hosts IREN, Applied Digital, TeraWulf, Core Scientific (4), plus Centrus, Innodata, and Sterling - each paired with real concentration, leverage, or execution risk.

Coverage-review note (June 4, 2026): the names above in bold - Siemens Energy, ABB, Emerson, NextEra, GlobalFoundries, MACOM, Lattice, Aehr Test, Kulicke & Soffa, Applied Digital, IREN, Core Scientific, TeraWulf, Dynatrace, Elastic, UiPath, and C3.ai - were added in a second-pass sweep for AI-supply-chain companies missing from the original cut. They are scored on the same 11-dimension model and slotted into the layer tables above. Two carry important flags: C3.ai (45) is included for completeness but its revenue is contracting (-52.5% YoY) - an AI-branded name that is not currently an AI beneficiary - and Core Scientific remains independent after its CoreWeave all-stock buyout was rejected by shareholders on Oct 30, 2025.

Top 10X Candidates - Highest-Probability Multibaggers

This is the section to read if you only read one. It isolates the names that screen highest on the 10X column of both scoring models - i.e., the highest probability of a 10X return in 24 months - and pressure-tests each against the one fact that dominates this question: the math. A 10X means a company’s market value rises ~10-fold. From a $1.7T name (TSMC) that implies ~$17T; from a $5.2T name (NVIDIA), ~$52T. Those are not 24-month scenarios. A literal 10x in 24 months is realistically confined to sub-~$15B caps with a binary catalyst or pre-revenue optionality - which is exactly where the model’s high 10X scores cluster, and exactly where balance-sheet and concentration scores are weakest. High upside and high fragility are the same names. Size every one of these small.

Tier A - Where a literal 10X is physically conceivable (sub-~$15B, binary)

These are small enough that a favorable resolution of a single catalyst could re-rate them ~10x. Each is a barbell: the same binary that creates the upside can halve the equity.

Reading Tier A. Oklo is the textbook 10X-or-bust: the INL Aurora criticality milestone (~mid-2026) is the single most-watched binary in the report, and a “nuclear-neocloud” build-own-operate re-rating from ~$13B toward $100B+ is the bull dream - paired with a -70% downside if licensing slips or it has to dilute. The SMR/enrichment trio (Oklo, NuScale, Centrus) is a thematic bet that the AI-power gap forces advanced nuclear; if SMR timelines slip (the base-rate outcome historically), all three de-rate together. Aehr Test is the cleanest new 10X-optionality name - a genuine bottleneck (burn-in for high-power AI ASICs) on a $3.5B cap, gated only by whether its lead-hyperscaler win becomes a repeat order in FY27. The three AI-datacenter hosts (APLD, WULF, CORZ) carry enormous contracted backlogs relative to their caps, but all three are GAAP-lossmaking, debt-financed, and concentrated on one or two counterparties (CoreWeave for APLD/CORZ; Google-backed Fluidstack for WULF) - they are leveraged calls on those single relationships staying intact.

Tier B - High-beta multibaggers (strong 10X optionality, but caps too large for a literal 24-month 10X)

These score 5 on 10X because of explosive growth and bottleneck importance, but at their current size a literal 10x would require $200B-$800B+ valuations. The realistic prize is a 2-4x, with a true 10x only on a multi-year horizon.

Honest takeaway: the report’s single highest composite scorers (NVDA/TSM/SK Hynix at 85) score 1-2 on 10X - they are the quality anchors that compound, not multiply. The genuine 10X tickets are the Tier A names, and they are tickets, not investments: a diversified, explicitly speculative basket sized so a total loss on any single name is survivable. If you cannot tolerate -70% on a position, it does not belong in this section.

Rankings by Category

These rankings weight the scoring model with qualitative overlays (catalyst timing, risk asymmetry, liquidity). Across ~115 names studied.

1. Best risk-adjusted winner (quality × durability × reasonable valuation)

1. TSMC (TSM) - indispensable chokepoint, +40% growth, forward P/E ~25, the rare “cheap” megacap. Only real risk is geopolitical.
2. Broadcom (AVGO) - custom-ASIC secular share gainer, AI +143%, forward P/E ~31, six hyperscaler programs.
3. SK Hynix - HBM #1, 72% op margin, ~7x forward earnings; the cleanest fundamental + valuation combo. (Honorable: Advantest, Micron.)

2. Most likely large-cap compounder (>$200B, durable double-digit growth)

1. NVIDIA (NVDA) - still the platform; ~34x earnings is undemanding for +85% growth, but size caps the multiple.
2. Broadcom (AVGO) · 3. TSMC (TSM). (Honorable: Eaton, ServiceNow, GE Vernova, Arista.)

3. Most likely mid-cap breakout ($10-60B, inflecting)

1. Celestica (CLS) - AI ODM, HPS +63%, margin-mix re-rating, 1.6T/CPO program 2027.
2. Coherent (COHR) / Lumentum (LITE) - 1.6T optics + the $4B NVIDIA capacity lock-up.
3. Sterling Infrastructure (STRL) - DC site-work pure-play, +92% revenue, 23% margins. (Honorable: Credo, nVent, Advanced Energy, Powell, Vertiv.)

4. Highest-risk / highest-upside candidate (binary, speculative - size small)

1. Oklo (OKLO) - pre-revenue SMR with 12 GW Switch + Meta deals; INL criticality (targeted mid-2026, now slipping) is the catalyst; could be worth multiples or a fraction.
2. Nebius (NBIS) - +684% growth, NVIDIA-backed, but P/S ~42-67 and customer-concentrated.
3. Navitas (NVTS) - NVIDIA-blessed 800V GaN partner on <$35M revenue.
4. Aehr Test (AEHR) - pure-play AI-processor burn-in; a lead-hyperscaler ASIC win + record $41M order on a ~$3.5B cap; lumpy but a genuine bottleneck-optionality name.
5. AI-datacenter hosts - IREN, Applied Digital (APLD), TeraWulf (WULF), Core Scientific (CORZ) - explosive contracted backlogs (Microsoft/NVIDIA, CoreWeave, Google-backed Fluidstack) but heavy leverage and single-counterparty concentration; size as a basket. (Honorable: NuScale, CoreWeave, Innodata, Bloom.)

5. Best picks-and-shovels beneficiary (sells to every AI builder regardless of who wins)

1. TSMC (TSM) - everyone’s foundry. 2. Advantest (ATEYY) - everyone’s tester. 3. ASML - everyone’s lithography. (Honorable: Amphenol, KLA, Besi, Ibiden, Kulicke & Soffa (KLIC) for HBM/AI advanced-packaging TCB, and Aehr Test (AEHR) for AI burn-in - you don’t have to pick the winning chip.)

6. Best data-center infrastructure beneficiary (power/cooling/electrical/build)

1. Vertiv (VRT) - purest liquid-cooling + power play, $15B backlog +109%.
2. Eaton (ETN) - electrical gear with diversification cushion, DC orders +240%.
3. GE Vernova (GEV) - turbines (cross-listed with energy). (Honorable: ABB - record DC-driven electrification orders +24% - nVent, Sterling, EMCOR, Comfort Systems, Equinix/Digital Realty.)

7. Best AI power / energy beneficiary

1. GE Vernova (GEV) - the arms dealer; sold-out turbine slots to 2029, $150B+ backlog, pricing power.
2. Constellation (CEG) - largest nuclear fleet, Microsoft + Meta PPAs, TMI restart 2027.
3. Vistra (VST) - nuclear + gas in ERCOT/PJM, AWS + Meta deals. (Honorable: Siemens Energy (ENR) - GEV’s only global turbine/grid peer, €154B backlog - and NextEra (NEE) - largest US renewables + 12 GW FPL large-load pipeline. Speculative upside: Oklo, Centrus, Bloom.)

8. Best AI software beneficiary

1. ServiceNow (NOW) - agentic-AI ACV tracking 50% above target, ~29x forward P/E, fortress moat.
2. Palantir (PLTR) - fastest growth (+85%) and best operational-AI franchise, but valuation is the entire risk.
3. AppLovin (APP) - ~85% EBITDA AI ad engine (SEC overhang). (Quality observability: Dynatrace (DT) - Davis AI, ARR +18%, GAAP-profitable - and Elastic (ESTC) - vector search for GenAI/RAG. Value angle: UiPath (PATH) at ~3.4x P/S, now GAAP-profitable; Salesforce at ~13x forward; MongoDB at ~12.5x P/S. Avoid: C3.ai - AI-branded but revenue -52.5% YoY.)

Investment Output - Ranked Watchlist

A focused watchlist with bull/base/bear, the specific conditions for a 10X, why the market may misprice it, the evidence that would invalidate the thesis, and what to monitor. Tiered by risk. Not advice; scenario analysis only.

Tier 1 - Core / Risk-Adjusted (own-the-chokepoint)

TSMC (TSM) - Foundry & packaging monopoly

• Bull: AI accelerator revenue compounds ~57% CAGR; CoWoS doubling absorbed at premium pricing; 2nm ramp lifts margins; multiple re-rates toward 30x → stock doubles+ in 24 mo.
• Base: +25-35% earnings growth, multiple steady ~25x; ~30-50% total return.
• Bear: AI digestion + China cyclicality dents utilization; FX (strong NT$) erodes USD margin; ~flat.
• What must happen to 10X: Not realistic in 24 mo from ~$1.7T (would need ~$17T). 10X is a 7-10 year story at best.
• Why mispriced: Trades at a discount to slower-growing US megacaps purely on Taiwan geopolitical risk - arguably the cheapest critical asset in tech.
• Invalidates thesis: Sustained CoWoS order cancellations; a credible 2nm yield miss; a Taiwan-Strait shock.
• Monitor: Monthly revenue; CoWoS capacity (toward 130k wpm); HPC % of revenue; 2nm yield commentary; NVIDIA/AMD order signals.

Broadcom (AVGO) - Custom ASIC + networking

• Bull: AI semi revenue compounds toward the >$100B ambition; OpenAI/Anthropic programs ramp; networking attach grows; >$3T cap.
• Base: AI semi >$60B in 2027, software stabilizes; ~40-60% upside.
• Bear: A custom program slips/cancels; VMware drag; AI digestion; ~flat to down 20%.
• 10X: No (from $2.28T).
• Why mispriced: Market still debates merchant-GPU vs ASIC; AVGO wins either way on networking, and custom is growing faster than the debate assumes.
• Invalidates: A hyperscaler in-sources ASIC design away from AVGO; AI semi guide cut.
• Monitor: AI semi revenue trajectory vs the ~$16B Q3 bar; number of custom customers; networking %; backlog.

Advantest (ATEYY) - HBM/AI test (overlooked)

• Bull: Test intensity per AI die keeps rising; HBM4/HBM4E + SoC test drive repeated guidance raises; ~$13B+ revenue; stock doubles.
• Base: +20-30% as guidance implies; modest re-rate.
• Bear: Memory capex air-pocket; margins have plateaued at record highs; -30%.
• 10X: Unlikely from ~$119B.
• Why mispriced: A near-duopoly bottleneck that trades below the multiples of less-critical AI names; under-followed by US investors (ADR).
• Invalidates: Memory-test order cuts; SoC tester market growth <20%.
• Monitor: SoC/HBM test bookings; guidance revisions; memory-maker capex; share vs Teradyne.

Tier 2 - Growth / Mid-Cap Breakout

Vertiv (VRT) - Liquid cooling + power

• Bull: Liquid cooling becomes mandatory through Rubin (600kW racks); backlog converts at rising margins; ~2x+.
• Base: ~30% organic growth sustained; ~30-50% upside.
• Bear: AI capex digestion + ~80x multiple compresses hard; -40%+.
• 10X: Not from ~$121B in 24 mo; a multi-year case at best.
• Why mispriced: Book-to-bill ~2.9x signals demand the Street under-models; the cooling-content-per-rack step-up with Rubin is underappreciated.
• Invalidates: Book-to-bill <1; order cancellations; margin reversal.
• Monitor: Quarterly book-to-bill, backlog, organic growth, adj. operating margin, hyperscaler capex.

Celestica (CLS) - AI ODM

• Bull: HPS mix drives margin expansion; 1.6T CPO switch program (2027) + Google TPU systems scale; ~2-3x.
• Base: +30-40% as guidance rises; modest re-rate.
• Bear: Top customer (~36%) cuts orders; thin EMS margins reassert; -40%.
• 10X: Possible only over a multi-year horizon if it becomes the default AI-switch ODM; not a 24-mo base case.
• Why mispriced: Still carries a legacy “low-margin EMS” multiple while transforming into a differentiated ODM/co-design winner.
• Invalidates: Loss of a top-2 program; HPS margin stalls.
• Monitor: HPS revenue/margin, customer concentration, CPO program milestones, FY guidance raises.

Coherent (COHR) / Lumentum (LITE) - 1.6T optics

• Bull: 1.6T ramps faster than supply; NVIDIA capacity lock-up + CPO supply; demand booked to 2028; ~2x.
• Base: +40-70% on datacom growth and margin mix.
• Bear: EML/component shortages cap shipments; CPO disrupts pluggables; LITE convert dilution; -30%+.
• 10X: COHR/LITE not from current caps in 24 mo; high-beta doubles are the realistic upside.
• Why mispriced: Optics is the “next bottleneck” the market is only beginning to price; NVIDIA’s $4B vote of confidence is recent.
• Invalidates: Datacenter transceiver growth stalls; NVIDIA shifts optics sourcing.
• Monitor: Datacenter revenue mix, 1.6T ramp, gross margin, (LITE) convertible cleanup.

Sterling Infrastructure (STRL) - DC site development

• Bull: Hyperscaler multi-year site programs keep E-Infra compounding +30%+ at 23% margins; ~2x.
• Base: +20-30% with backlog conversion.
• Bear: A hyperscaler pause hits the 4-customers-=-35% concentration; margins mean-revert; -40%.
• 10X: A multi-year stretch case, not 24 mo.
• Why mispriced: Viewed as a cyclical contractor; actually earns software-like margins on scarce site-work capacity.
• Invalidates: Backlog decline; E-Infra margin compression below ~20%.
• Monitor: E-Infra backlog/margin, customer concentration, book-to-bill, hyperscaler capex.

Kulicke & Soffa (KLIC) - HBM/AI advanced-packaging (overlooked, new)

• Bull: Fluxless TCB becomes the default bonding step for HBM/AI stacks; TCB revenue tops $100M and compounds; earnings normalize off the cyclical trough; net-cash B/S funds buybacks; ~2x.
• Base: +25-35% as the June-qtr +28% guide and TCB ramp play out; forward P/E ~25 re-rates modestly.
• Bear: Equipment cyclicality + BESI/ASMPT competition in advanced packaging cap the TCB ramp; -30%.
• 10X: Not from ~$5.5B in 24 mo; a multi-year story if TCB wins the bonding standard.
• Why mispriced: Trailing P/E ~100 optically scary, but it’s a cyclical-trough artifact - forward ~25x understates a genuine HBM/AI packaging inflection with a fortress (net cash ~$448M) balance sheet.
• Invalidates: TCB revenue undershoots the >$100M / +70% guide; advanced-packaging share loss to BESI.
• Monitor: TCB revenue/bookings, gross-margin mix, advanced-packaging order trends, HBM-maker capex.

Tier 3 - Speculative / Highest-Risk-Highest-Upside (size small; binary)

Aehr Test (AEHR) - AI-processor burn-in (new)

• Bull: The lead-hyperscaler Sonoma win for next-gen AI ASICs ramps into repeat FY27 production orders; test-/burn-in-intensity per high-power die keeps rising; record $41M order is the first of a series; from ~$3.5B a re-rate toward multiples of that as a recognized bottleneck - a genuine 10X-optionality name.
• Base: Lumpy, order-driven; volatile and range-bound until the FY27 ramp is visible.
• Bear: Revenue is tiny and lumpy (-44% trailing), ~70x sales, and concentrated; an order air-pocket or a customer in-sourcing burn-in → -50%+.
• What must happen to 10X: Hyperscaler AI-ASIC burn-in becomes a standard, recurring production step and Aehr holds its near-monopoly → revenue steps up several-fold with high incrementals on a zero-debt base.
• Why mispriced (or not): A real bottleneck (reliability burn-in for 100s-of-watt AI accelerators) that is early, lumpy, and under-followed - the market can’t yet decide if it’s a niche tester or the “ASML of burn-in.”
• Invalidates: The hyperscaler win fails to convert to follow-on orders; SoC/HBM test growth <20%; a dilutive raise.
• Monitor: Bookings/backlog ($51M effective), Sonoma follow-on orders, customer count, gross-margin recovery on volume.

AI-datacenter host basket - IREN, Applied Digital (APLD), TeraWulf (WULF), Core Scientific (CORZ) - neocloud hosts (new)

• Bull: Enormous contracted backlogs convert on schedule - IREN (Microsoft $9.7B + NVIDIA $3.5B → ~$3.7B ARR), APLD (CoreWeave ~$11B + $5B + $7.5B), WULF (Google-backed Fluidstack ~$9.5B), CORZ (CoreWeave >$10B, 80-85% GM) - and the equities re-rate as the contracts de-risk; individual 2-4x, with the smallest (CORZ ~$9.2B) carrying the most 10X-optionality.
• Base: Backlogs ramp but GAAP losses, dilution, and rate-sensitivity persist; high volatility, mixed outcomes across the basket.
• Bear: A single anchor (CoreWeave for APLD/CORZ; Fluidstack/Google for WULF; Microsoft for IREN) renegotiates or slips, and refinancing tightens → -50-70% on the affected name; the shared-CoreWeave exposure can hit APLD and CORZ together.
• What must happen to 10X: Only the smaller names (CORZ, and on a stretch APLD/WULF ~$9-13B) can approach a literal 10x - requires full backlog conversion, a refinance, and a second large tenant to break single-counterparty risk.
• Why mispriced: Backlogs dwarf current caps, but the market (correctly) discounts single-tenant credit + leverage; the debate is whether these are “contracted annuities” or “levered bets on one customer.”
• Invalidates: Any anchor-tenant renegotiation/cancellation; a failed refinancing; CoreWeave credit stress; GPU-pricing collapse.
• Monitor: Contracted-revenue conversion to recognized revenue, MW energized vs plan, debt maturities/refi terms, tenant diversification, ultimate-counterparty (CoreWeave) health.

Oklo (OKLO) - Advanced SMR

• Bull: INL Aurora reaches criticality (targeted mid-2026, now slipping) and operations (2027-28); Switch (12 GW) + Meta deals convert; build-own-operate recurring model re-rates it as the “nuclear neocloud”; a genuine 10X candidate if first-of-kind works.
• Base: Milestones slip but optionality persists; volatile, range-bound to +50%.
• Bear: Licensing/construction delays, dilution, technology risk; -70%+.
• What must happen to 10X: Aurora online on schedule + at least one large PPA converts to construction + a financing path that doesn’t massively dilute → from ~$13B toward $100B+ as the market capitalizes the 12 GW pipeline.
• Why mispriced (or not): Pure optionality on first-of-kind nuclear; the market either capitalizes the pipeline or treats it as pre-revenue vapor - a true barbell.
• Invalidates: Criticality slips materially; a PPA cancellation; a safety/licensing setback; a dilutive raise.
• Monitor: INL criticality milestone, NRC licensing steps, PPA-to-construction conversions, cash runway/dilution.

Nebius (NBIS) - Neocloud

• Bull: Microsoft/Meta contracts ramp to $7-9B ARR; ClickHouse/Toloka stakes crystallize; ~2-3x.
• Base: Hits ARR targets but multiple compresses from ~50x P/S; volatile flat.
• Bear: Capex/financing strain, GPU price decay, customer concentration; -60%.
• 10X: Only if it becomes a structurally profitable top-3 neocloud and the multiple holds - low probability, non-trivial payoff.
• Why mispriced: Better-capitalized than CoreWeave ($9.3B cash, NVIDIA-backed) yet lumped with it.
• Invalidates: A major customer renegotiation; failure to fund the $20-25B capex; ARR miss.
• Monitor: ARR vs $7-9B exit target, customer mix, cash/financing, GPU utilization/pricing.

Centrus (LEU) - HALEU enrichment

• Bull: Becomes the domestic HALEU monopoly as SMRs scale; DOE + commercial offtake; ~2-3x+.
• Base: Contract-driven, lumpy; +30-60%.
• Bear: SMR timelines slip, demand stays prospective, appropriations risk; -50%.
• 10X: Plausible over multi-year if the SMR fleet materializes and Centrus is the sole US fuel source - not a clean 24-mo case.
• Why mispriced: A genuine national-security monopoly whose demand is real but back-end-loaded.
• Invalidates: $900M DOE order fails to finalize; SMR developers slip; cheaper fuel routes emerge.
• Monitor: DOE contract finalization, HALEU production volumes, SMR developer milestones, SWU pricing.

(Other speculative names to watch with the same discipline: NuScale, CoreWeave, Navitas, Bloom, Innodata, Talen.)

Risk Controls

Every name in this report is exposed, to varying degrees, to the same master variable: the durability of hyperscaler capex. Diversification across layers reduces single-stock risk but not the systemic AI-capex risk. The specific risks to manage:

1. AI capex bubble / digestion risk. The bull case requires end-customer AI revenue to eventually justify ~$725B → $1T+ annual capex (the dashed loop in the map). Meta and Amazon stocks fell on Q1 prints despite beats, purely on payback-timing fear. A single quarter of hyperscaler “optimization/digestion” language would hit the highest-multiple infrastructure names (VRT ~80x, MOD ~136x, ALAB ~216x) first and hardest. Control: size infrastructure/optics/ODM positions smaller; watch capex guidance and book-to-bill as the leading indicator.

2. Margin compression. ODM/EMS (Quanta ~4.8% GM, Foxconn ~2-3% op) make money on volume, not margin - any pricing pressure or memory-cost spike compresses them fast. Even Vertiv/Eaton face price-cost/tariff risk. Control: prefer high-gross-margin chokepoints (TSMC, Advantest, Arista, SK Hynix) over thin-margin assemblers for core positions.

3. Customer concentration. Severe in: Wiwynn (~95% A/R in 3, >50% Meta), Symbotic (~85% Walmart), Credo (~34% one customer), Celestica (~36%), CoreWeave (Microsoft 67%), Innodata (56%), Fabrinet (~50% top-3), Arista (MSFT+META ~48%), and now several coverage-review adds: Applied Digital and Core Scientific (effectively single-tenant on CoreWeave), TeraWulf (single Fluidstack/Google anchor), IREN (Microsoft + NVIDIA), and Aehr Test (extreme, lumpy). A new wrinkle is shared-counterparty risk: APLD and CORZ both depend on CoreWeave - itself a levered, Microsoft-concentrated credit - so a single CoreWeave stumble would hit multiple “diversified” host positions at once. Control: treat these as higher-beta; map the ultimate counterparty, not just the direct one; a single customer’s capex shift is an idiosyncratic shock.

4. Export controls. ASML China fell 36%→19% in one quarter; AMAT guided a $600-710M FY hit; Lam China heading 43%→<30%. The proposed MATCH Act is the escalation risk. NVIDIA’s China DC-compute is effectively zeroed. Control: model equipment names ex-China; treat any China re-opening as upside option, not base case.

5. Supply shortages (the double-edged sword). Memory (worst in 15 yrs), CoWoS, HBM, transformers (128-144 wk lead times), optics components, and AI-server parts (Wiwynn warns into 2027-28) are all constrained. Shortages = pricing power for suppliers but revenue caps and delivery slips for buyers/ODMs. Control: distinguish shortage beneficiaries (memory, test, packaging) from shortage victims (ODMs waiting on parts, hyperscalers waiting on power).

6. Power constraints. The hardest physical limit - gas turbines sold out to ~2029, multi-year grid queues, FERC co-location rulings (it rejected Talen’s original structure) are a binary regulatory variable. Control: favor power names with signed PPAs and real earnings (CEG/VST/GEV) over pre-revenue SMRs for core exposure.

7. Valuation compression. Much of the chain is priced for perfection: ARM (~160-490x), ALAB (~216x), MOD (~136x), Palantir (P/S ~45), PANW (P/S ~23), Nebius (P/S ~42-67). Multiple compression can erase returns even if fundamentals deliver - Palantir fell ~26% YTD after an +85% growth quarter; Salesforce -33% YTD despite a beat. Control: separate “great business” from “great stock at this price”; demand a valuation cushion (TSMC, Micron, SK Hynix, TEL, Hubbell, Salesforce) for larger positions.

8. Competition from hyperscaler in-house chips. Every hyperscaler is building custom silicon (Trainium, TPU, Maia, MTIA), which is simultaneously a tailwind for AVGO/MRVL/Marvell/Celestica and a long-term threat to NVIDIA’s share and to merchant economics. Control: own both sides (NVDA + AVGO) rather than betting the merchant-vs-custom outcome.

9. Interest-rate & macro risk. The buildout is debt-financed at the margin - CoreWeave (~$25B), Oracle (~$175B total), the AI-datacenter hosts (Applied Digital ~$2.7B, Core Scientific $3.3B bond, IREN $3.7B converts, TeraWulf converts), Amphenol, REITs, and IPPs all carry rate sensitivity; higher-for-longer raises financing costs and compresses long-duration growth multiples. The host model is especially fragile: it funds multi-year, single-tenant buildouts with near-term debt. Control: stress-test the levered names (CoreWeave, Oracle, Nebius, APLD, CORZ, IREN, WULF) for refinancing; favor net-cash balance sheets (Arista, NVIDIA, ServiceNow, Powell, SK Hynix, Aehr, Kulicke & Soffa, UiPath, Dynatrace).

10. Regulatory risk. AI governance/antitrust, FERC co-location, export controls, securities probes (AppLovin SEC, Super Micro DOJ/SEC + export indictment), and nuclear licensing (NRC) are all live. Control: discount names with active legal overhangs (SMCI, APP) and binary regulatory dependencies (SMRs, co-located IPPs).

Portfolio-construction takeaway. A defensible structure is a barbell: a core of high-margin, net-cash chokepoints bought with a valuation cushion (TSMC, Broadcom, SK Hynix/Micron, Advantest, Arista, ServiceNow), plus a small, explicitly speculative sleeve of binary 10X-optionality names (the Tier A set in the Top 10X section - Oklo, NuScale, Navitas, Aehr Test, Centrus, and the AI-datacenter hosts APLD/WULF/CORZ - plus Nebius) sized so that a total loss on any one is survivable. Avoid concentrating in the highest-multiple mid-caps as if they were low-risk compounders - they are leveraged bets on uninterrupted capex.

Appendix A - The One-Screen Monitoring Dashboard

The leading indicators that confirm or break the whole thesis, in priority order:

1. Hyperscaler capex guidance (MSFT, AMZN, GOOGL, META, ORCL) each quarter - the master variable. Watch for the word “optimize/digest.”
2. NVIDIA data-center revenue & guidance - the demand barometer for the entire chain.
3. Book-to-bill at the infrastructure names (Vertiv ~2.9x, nVent, Eaton orders, GE Vernova backlog) - first to roll over in a digestion.
4. Memory pricing (DRAM/HBM contract prices, HBM4 sold-out status) - confirms the supercycle for SK Hynix/Micron/Samsung.
5. CoWoS capacity (TSMC toward ~130k wpm) and substrate/ABF pricing (Ibiden) - packaging bottleneck health.
6. Optics datacenter mix & 1.6T ramp (Coherent/Lumentum/Fabrinet) - the “next bottleneck.”
7. Power deal flow (new hyperscaler PPAs; FERC co-location rulings; GE Vernova turbine pricing/slots) - the hardest constraint.
8. Neocloud financing & customer mix (CoreWeave/Nebius debt, OpenAI funding) - the circular-financing stress point.
9. Software AI-ARR inflections (ServiceNow Now Assist, Salesforce Agentforce, Palo Alto NGS ARR) - does the application layer monetize the capex?
10. Export-control headlines (MATCH Act, China rules) - equipment & NVIDIA tail risk.

Appendix B - Consolidated Data-Quality Flags

• Market caps for CoreWeave (~$60-65B) and Nebius (~$36.5-70B) vary widely by share-count methodology and date - re-verify before sizing.
• Micron FQ2 gross margin is 74.4% (SEC filing); several secondary outlets cited a stale ~45%. Revenue is $23.86B (+196%), not the ~$8.7B some stale sources show.
• HBM4 supplier-share splits (SK Hynix ~54-70% / Samsung ~25-30% / Micron ~18%) vary by source (Counterpoint vs allocation reports).
• Broadcom AI revenue appears as both $8.4B (Q1) and $10.8B (Q2) across the reporting window - both are correct for their respective quarters.
• Sandisk market cap ~$271B (148M shares × ~$1,783); some aggregators show a stale ~$126B. (Confirmed real in the June verification - the memory/storage supercycle re-rated SNDK, STX, and WDC far above historical ranges.)
• Navitas has SPAC-related (D. Boral ARC) filing activity in EDGAR - verify share structure before relying on its market cap.
• Confluent (CFLT) was acquired by IBM (~$11B, closed March 17, 2026) and is no longer investable.
• Core Scientific (CORZ) is independent. CoreWeave’s all-stock acquisition (signed July 2025, ~$9B) was rejected by Core Scientific shareholders on Oct 30, 2025 and the merger terminated; the two remain commercial (not corporate) partners. Do not treat CORZ as acquired.
• C3.ai (AI) revenue is contracting, not growing: Q4 FY26 revenue -52.5% YoY, full-year FY26 -35.7% (per the June 3 SEC release). It is included for coverage completeness and flagged as an AI-branded name that is not currently an AI beneficiary; FY27 guidance implies further shrinkage. Some secondary outlets mislabeled the full-year decline as -52.5% - use the SEC figures.
• AI-datacenter host market caps (APLD ~$12.8B, IREN ~$22B, CORZ ~$11B, WULF ~$12.8B) are intraday snapshots and move sharply; all four are GAAP-lossmaking (losses dominated by non-cash impairments/warrant marks) and debt-financed - re-verify caps, cash, and debt maturities before sizing.
• Elastic (ESTC) “GAAP profitability” (FY26 net income ~$367.8M) is inflated by a ~$435M one-time deferred-tax benefit; on an operating basis it was roughly breakeven-to-slightly-negative - do not characterize it as cleanly GAAP-operating-profitable.
• Aehr Test (AEHR), Kulicke & Soffa (KLIC), Lattice (LSCC) carry trailing P/Es distorted by cyclical-trough or loss-making earnings (AEHR ~70x sales and unprofitable; KLIC trailing ~100x vs forward ~25x; LSCC trailing n/m vs forward ~67x) - use forward multiples and note the lumpiness, especially for AEHR.
• Siemens Energy (ENR.DE) trades primarily in Frankfurt (US access via SMNEY/SMEGF OTC ADRs, less liquid); figures are in EUR (~€154B backlog) converted at ~1.08-1.10 USD. NextEra (NEE) carries a pending, material ~$66.8B acquisition of Dominion - model it as a live variable.
• Foreign-listed names (Foxconn 2317.TW, Quanta 2382.TW, Wiwynn 6669.TW, Ibiden 4062.T, Advantest 6857.T, SK Hynix 000660.KS, Samsung 005930.KS, Tokyo Electron 8035.T, Siemens Energy ENR.DE) trade primarily on home exchanges; US access is via ADR/OTC (often less liquid) - confirm the tradable line and FX exposure.

Appendix C - Sourcing

This report synthesizes ~290+ public sources gathered via live web research on June 4, 2026 (including a second-pass coverage review that added 17 companies), prioritizing SEC filings (8-K/10-Q/10-K/6-K), company IR/press releases, and earnings-call transcripts, corroborated by credible trade and financial press (CNBC, Fortune, Utility Dive, POWER Magazine, DataCenterDynamics, Tom’s Hardware, DIGITIMES, LightCounting, Futurum) and market-data aggregators (companiesmarketcap.com, stockanalysis.com, macrotrends.net) for caps and multiples. Representative primary and corroborating links are embedded inline throughout. Where a figure rested on a single secondary source or sources diverged, it is flagged inline or in Appendix B.

Appendix D - Verification & Fact-Check Log (June 4, 2026)

A full verification pass was run on all 115 scored companies using twelve parallel sector sweeps. For each company, the market cap, latest-quarter revenue growth, margin, and headline deal/customer claim were re-checked against primary sources (SEC 8-K/10-Q/6-K, company IR releases) and the market-data aggregators used elsewhere in this report (companiesmarketcap.com, stockanalysis.com). The 1-10 scores were left unchanged - corrections apply to descriptive figures, not the judgment ratings.

Headline result: the report’s qualitative facts and growth rates proved largely accurate - roughly 80 of 115 names needed no change. Several caps that looked implausibly high were confirmed real in this 2026 scenario: the memory/storage supercycle (Sandisk ~$271B, Seagate ~$208B, Western Digital ~$183B) and the AI re-rating of Dell (~$285B), HPE (~$73B), Lumentum (~$73B), Ciena (~$87B), and Bloom (~$83B). Core Scientific’s independence (CoreWeave buyout rejected by shareholders Oct 30 2025), C3.ai’s revenue contraction (~-50% YoY), and Navitas’s bubble valuation on <$35M revenue were all confirmed.

Material corrections applied across the report, dashboard, and spreadsheet:

Flags / limitations. Market caps are intraday snapshots and varied a few percent across aggregators (more for volatile post-earnings names like CRWD, CRM, PLTR, STRL) - treat all caps as approximate. ARM’s royalty-growth framing and Arista’s customer-concentration figure (~35% FY24 vs the ~48% shown) are period-dependent and were softened rather than hard-set. The OKLO 10X score (7) was not lowered despite the criticality slip - the optionality stands, but the catalyst timing now carries more risk. Full per-name results are retained in VERIFICATION-FINDINGS-resume-here.md.

Prepared June 4, 2026. Figures are point-in-time and will change. This is research and analysis for educational purposes - not investment advice, not a recommendation, and not a solicitation. Markets involve risk of loss; “10X” scenarios are low-probability thought experiments. Verify all figures independently and consult a licensed financial advisor before making investment decisions.