AI Rush.
Picks-and-shovels watchlist. Fifty public companies catching hyperscaler capex, sorted by layer, scored.
clt_AIGuy · Updated 2026-05-04 · Live, public
The Road to $16K
Eight $2K tranches → $16K total. Each stop names what to buy and why. Stop at any tier and own a real position, not a half-baked one.
Click any tier to expand. Per-buy cap is $250 — keeps positions small and DCAs each name across multiple tiers. AI Chips allocations are concrete (deeply scored); the other 11 layers carry preliminary thesis-based allocations — pure-play per layer, weighted by how directly each captures hyperscaler capex. Weights will tighten as each layer's quant scoring lands.
- AI Chips this tierconcreteWhy this tierStart with the two highest-quality chip names PLUS a starter in each of the seven highest-conviction layers. Even if you stop at $2K you own a real diversified AI-buildout basket across 9 names — not just two chip lottery tickets.
Final basket at $16K — 24 names, $580 avg position:
AI Chips $2,000 (12.5%): TSM $700 · AVGO $550 · MRVL $450 · AMD $200 · NVDA $100 (conditional).
Other 11 layers $14,000 (87.5%): Power CEG/VST/GEV $2,400 · Networking ANET/GLW/COHR $2,050 · Semicap ASML/LRCX/KLAC $1,900 · Memory MU $1,650 · Cooling VRT/MOD $1,650 · DC REITs EQIX/DLR $1,300 · Power Dist ETN $950 · Generators CAT/CMI $800 · Server DELL $700 · Commodities FCX $600 · DC MEP $0 (LGN insufficient price history).
Backtesting
The Road to $16K basket → tracked over time. The 24-name allocation from the roadmap above, deployed at 2026-05-04 prices. Every few days we re-run with current prices and log the return — so the framework is measured, not just claimed.
Why this exists. The framework is only honest if we measure it. Each row in the run history table below is a real price pull, recorded forever. If the roadmap picks don't beat the equal-weight $200×50 baseline over time, we go back to the drawing board.
Recession watch. AI compute names have huge upside but also lead the way down if the macro turns. These four indicators have flagged every recession in the past 50 years — usually months before it shows up in equities. Three green, one orange right now.
A diffusion index of new orders, production, employment, supplier deliveries, and inventories from manufacturing purchasing managers. Updated monthly on the first business day. For this portfolio specifically: chip cycles are tied to manufacturing demand — when PMI rolls over, memory pricing and equipment orders follow.
Watch for. Drop below 50 (contraction starts), then below 47 (real recession risk). Trend matters more than any single print — three consecutive monthly declines is the signal.
Difference between 10-year and 2-year Treasury yields. Inversion (negative) preceded 7 of the last 7 recessions with a 12–24 month lead time. The recession itself usually arrives during re-steepening from inversion, not at the deepest point of inversion. We were inverted for ~24 months and have just barely re-steepened — historically that's the highest-risk window.
Watch for. Drop back below 0 (bond market repricing). Especially watch direction — sharp re-steepening (0 → 1.0+ in a few weeks) is the late-cycle 'rate cuts because something broke' signal.
3-month moving average of unemployment minus its 12-month low. Triggers at 0.50, which has flagged every US recession since 1970 with zero false positives. We had a brief trip to ~0.5 in 2024 that didn't lead to recession — partial false positive driven by labor-supply shocks — but the rule is still the cleanest real-time labor signal.
Watch for. Any sustained move above 0.30 (orange). 0.50 is the canonical recession-has-already-started line.
Option-adjusted spread on US junk bonds vs. Treasuries. Updates daily — your fastest moving indicator. At 2.75% credit is very loose, which is why AI capex deals are getting funded so easily.
Watch for. Sustained move above 4.0% (orange) which historically front-runs equity drawdowns by a few weeks. Above 5.5% means real funding stress — AI infra names will reprice hard. The thing to watch is a sudden 100bp+ widening over a few weeks — that's 'something broke' pricing.
Roadmap basket from /blog/ai-rush — 24 named positions across all 12 layers (DC MEP excluded; LGN insufficient price history). Per-buy cap $250 across the 8 tranches; the totals here reflect the sum at $16K.
Why this exists. The framework is only honest if we measure it. Each backtest run pulls live prices and logs portfolio value. Conditional names (AMD, NVDA) are included assuming both triggered — adjust mentally if not.
| Run date | Portfolio value | Δ vs prior | Δ vs entry | Return % |
|---|---|---|---|---|
| 2026-05-11 | $17,456.75 | +$441.70 | +$1456.75 | +9.10% |
| 2026-05-08 | $17,015.05 | +$568.62 | +$1015.05 | +6.34% |
| 2026-05-07 | $16,446.43 | -$574.18 | +$446.43 | +2.79% |
| 2026-05-06 | $17,020.61 | +$510.43 | +$1020.61 | +6.38% |
| 2026-05-05 | $16,510.18 | +$510.18 | +$510.18 | +3.19% |
| 2026-05-04entry | $16,000.00 | — | — | — |
| Ticker | Layer | Entry $ | Run $ | Cost | Value | Return % |
|---|---|---|---|---|---|---|
| MU | Memory | $542.21 | $795.33 | $1,650.00 | $2,420.27 | +46.68% |
| GLW | Networking | $158.26 | $207.39 | $550.00 | $720.74 | +31.04% |
| AMDIF | AI Chips | $360.54 | $458.79 | $200.00 | $254.50 | +27.25% |
| DELL | Server | $210.17 | $247.04 | $700.00 | $822.80 | +17.54% |
| LRCX | Semicap | $256.72 | $296.05 | $600.00 | $691.92 | +15.32% |
| COHR | Networking | $329.50 | $379.69 | $400.00 | $460.93 | +15.23% |
| FCX | Commodities | $56.55 | $64.37 | $600.00 | $682.97 | +13.83% |
| VRT | Cooling | $328.31 | $367.92 | $1,450.00 | $1,624.94 | +12.06% |
| NVDAIF | AI Chips | $198.45 | $219.44 | $100.00 | $110.58 | +10.58% |
| ASML | Semicap | $1427.02 | $1565.81 | $900.00 | $987.53 | +9.73% |
| KLAC | Semicap | $1726.26 | $1845.19 | $400.00 | $427.56 | +6.89% |
| CMI | Generators | $657.44 | $702.66 | $400.00 | $427.51 | +6.88% |
| MOD | Cooling | $266.83 | $284.80 | $200.00 | $213.47 | +6.73% |
| CAT | Generators | $889.67 | $926.79 | $400.00 | $416.69 | +4.17% |
| MRVL | AI Chips | $164.95 | $170.84 | $450.00 | $466.07 | +3.57% |
| TSM | AI Chips | $397.67 | $404.54 | $700.00 | $712.09 | +1.73% |
| AVGO | AI Chips | $421.28 | $428.43 | $550.00 | $559.33 | +1.70% |
| GEV | Power | $1062.95 | $1073.08 | $650.00 | $656.19 | +0.95% |
| EQIX | DC REITs | $1085.03 | $1086.22 | $700.00 | $700.77 | +0.11% |
| ETN | Power Dist | $425.55 | $419.00 | $950.00 | $935.38 | -1.54% |
| VST | Power | $155.28 | $152.05 | $900.00 | $881.28 | -2.08% |
| DLR | DC REITs | $200.70 | $196.24 | $600.00 | $586.67 | -2.22% |
| CEG | Power | $307.81 | $299.69 | $850.00 | $827.58 | -2.64% |
| ANET | Networking | $172.70 | $136.43 | $1,100.00 | $868.98 | -21.00% |
The Premise
: $400B/yr on AI infrastructure. This page = public watchlist of 50 companies catching that spend, sorted by layer, scored against my framework.
- Why it matters. That spend drove ~92% of US GDP growth last year. Without it, economy was flat.
- Where money goes. , custom chips, , networking, cooling, switchgear, transformers, electricity, fiber, land. 12 layers.
- Why I sort by layer. Same dollar, different speeds and margins per layer. Lump them = overweight one expensive corner.
- What's here today. All 50 names through the framework (quant). AI Chips has full deep cards; other 11 layers are compact tables you can expand.
The Stack
12 layers, 50 companies. Click any layer to jump.
- 01AI ChipsFlagship7 names
- 02Memory & HBM / Storage4 names
- 03Networking & Fiber8 names
- 04Semiconductor Equipment3 names
- 05Server & Hardware4 names
- 06Cooling & HVAC6 names
- 07Power Distribution3 names
- 08Generators (Backup)3 names
- 09Power & Energy (Utilities)5 names
- 10DC REITs & Construction4 names
- 11DC Construction / MEP0 names
- 12Commodities (Copper)1 names
Layer · AI Chips
Most-talked-about layer. Most expensive. These companies design or fabricate the that turn electricity into AI computation.
- Three roles, everything else is a flavor. NVIDIA designs Blackwell. Broadcom designs custom chips for hyperscalers. TSMC etches the silicon. AMD, MRVL, ARM, TSEM are riffs on those three.
- The risk. Every name trades at a multiple that already assumes capex grows for years. Hyperscalers slow even one quarter → this layer takes the biggest re-rate first.
Sole leading-edge foundry. NVIDIA, AMD, AVGO all etch silicon here. Piotroski 9/9 (perfect). EV/EBIT 21 — fair for the only fab on the leading edge. Highest-conviction chip name.
- $2.09T
- 33.5
- 21
- 46%
- Stage 2
- 9 / 9
- 3.06
- 58%
- Rev growth
- +35% YoY
Custom AI chips for hyperscalers (Google TPU silicon, Meta MTIA design parts). Dual play: AI silicon + 30%-margin VMware software. Piotroski 8/9. EV/EBIT 23 — expensive, but quality is real. Stage 2.
- $1.95T
- 81
- 23
- PEG
- 0.92
- 20%
- Stage 2
- 8 / 9
- 45%
- Rev growth
- +30% YoY
Networking silicon + custom chips for hyperscalers. AI revenue ramping fast (data center is 60% of business now). Piotroski 8. EV/EBIT 27. Stage 2.
- $144B
- 47
- 27
- PEG
- 1.93
- 17%
- Stage 2
- 8 / 9
- 19%
- Rev growth
- +22% YoY
Number two in AI chips, miles behind NVIDIA. MI300/MI350 selling but not leading. P/E 122 = absurd; doesn't earn the premium NVIDIA earns. Pass — wait for pullback or earnings reset.
- $560B
- 122
- 29
- PEG
- 1.07
- 7%
- Stage 2
- 7 / 9
- 17%
- Rev growth
- +34% YoY
Chip IP licensor. Every smartphone + most edge AI runs ARM ISA. Royalty model = high margins. But P/E 282 + EV/EBIT 99 = priced for perfection × 5. Pass.
- $216B
- 282
- 99
- PEG
- 2.54
- 16%
- Stage 2
- 6 / 9
- 15%
- Rev growth
- +26% YoY
Specialty foundry: niche analog/RF chips. Small cap = volatile. P/E 100 + EV/EBIT 39 = expensive for the cap size. Piotroski 6. Pass for now.
- $25.6B
- 100
- 39
- PEG
- 5.24
- 8%
- Stage 2
- 6 / 9
- 16%
- Rev growth
- +14% YoY
Dominant AI chip company. = software moat — switching costs multi-year. But: framework flags failed hard gate (Piotroski 4) + EV/EBIT 33. World-class business; price reflects years of perfect execution.
- $4.81T
- 41
- 34
- 76%
- Stage 2
- 4 / 9
- 63
- $103B
- Rev growth
- +58% YoY
Layer · Memory & HBM / Storage
Layer · Networking & Fiber
8 names · Click any row for the snapshot detail.
Layer · Semiconductor Equipment
Layer · Server & Hardware
Layer · Cooling & HVAC
6 names · Click any row for the snapshot detail.
Layer · Power Distribution
Layer · Generators (Backup)
Layer · Power & Energy (Utilities)
Layer · DC REITs & Construction
Layer · DC Construction / MEP
No scored names yet for this layer.
Layer · Commodities (Copper)
1 names · Click any row for the snapshot detail.
| Ticker | Company | Verdict | Quant | |||
|---|---|---|---|---|---|---|
| FCX | Freeport-McMoRan | Pass | 2 | 6/9 | 14.51 | 22/53 |
Honest Limits
- Quant only for non-flagship layers. AI Chips has full deep cards. The other 11 layers show quant-derived preliminary verdicts. Full qualitative (catalyst, management, business clarity) coming layer by layer.
- Snapshot frozen at 2026-05-04. , prices, scores all dated. Re-run the script to refresh.
- Backtest = dumb baseline. Equal-weight 50 names is the simplest possible strategy. If the framework can't beat it, the framework has a problem. Smarter variants (gate-pass only, top-quartile, score-weighted) come in follow-ups.
- Framework punishes valuation hard. NVDA gate-fails on 4 + high . ARM + TSEM scored Pass on valuation. Reasonable investors can still own them — framework just won't pretend the price is.
- Roadmap is illustrative. Tier amounts assume a single $16K deployment over 8 tranches. Real position sizing depends on your account size, risk tolerance, tax situation. Talk to a real advisor.
- Public list only. Private players (CoreWeave, Lambda, Anthropic, OpenAI, xAI) — can't buy them as public shares.
- Skipped from snapshot: FPS (no price (insufficient history)), LGN (no price (insufficient history)).
How This Is Scored
Every name runs through the same 7-step framework: hard gates first, then a 100-point rubric.
- Hard gates — any failure = no buy. , , positive , cap ≥ $300M, $-vol ≥ $5M, NYSE/NASDAQ listing.
- Rubric — 100 points / 7 sections. Business (10), financial health (20), valuation (20), quality / (15), management (10), (15), (10).
- Sources. Joseph Piotroski (F-Score), Edward Altman (Z-Score), Tobias Carlisle (Acquirer's Multiple), Stan Weinstein (stage analysis), Aswath Damodaran (valuation).
- Data pipeline. yfinance for fundamentals + prices. Backtest =
backtest.py+emit_ts.py.
Sources. yfinance for fundamentals + price history; SEC EDGAR for filings; AI Compute Top 50 as the universe; Piotroski (2000), Altman (1968), Carlisle (Acquirer's Multiple), Weinstein (stage analysis), Damodaran (valuation).
Not investment advice. Research. Predictions are guesses with math attached. Your money, your responsibility.