AI Buildout Tool
The AI Buildout Constraint Scanner
A Baker-style model for ranking watts, wafers, memory, power, and Mag 7 AI exposure.
9 July 2026 · YK Research
The Market Is Still Sorting AI Stories From AI Constraints
Gavin Baker's useful contribution to the AI debate is that he starts with physical limits. AI demand is not limited only by software adoption. It runs into watts, wafers, HBM, advanced packaging, optical links, cooling, interconnects, and the time it takes to turn capex into revenue.
That changes the screen. A normal AI basket overweights the names with the loudest narrative. A constraint-led basket asks which company controls the scarce input, whether that scarcity is showing up in revenue growth, whether the business is good enough to keep the economics, whether the price already discounts it, and whether the tape confirms the thesis.
This is not a price target model. It is a research queue. The point is to force a clean first pass before doing the real underwriting.
The Model
The score combines the physical bottleneck view with the QVMR discipline: quality, value adjusted for quality, momentum, and risk. The weights are deliberately simple:
Final score
Growth score
Momentum score
Risk score
The Result: The First Queue Is TSM, NVDA, MU, ASML, KLAC
The top of the run is not a generic Mag 7 list. It is foundry, GPUs, memory, EUV, process control, wafer-fab equipment, optical interconnect, custom silicon, cooling, and grid equipment. That is exactly what a Baker-style lens should produce.
| Rank | Ticker | Score | Verdict | Bucket | C/G/M/Q/V/R | Revenue | Rev growth | Read |
|---|---|---|---|---|---|---|---|---|
| 1 | TSM | 75.5 | Underwrite | Wafers | 77/69/66/92/94/78 | $127.7B | 35% | Foundry plus advanced packaging. The cleanest bottleneck score in the run. |
| 2 | NVDA | 73.5 | Watch/add | GPU/system | 84/100/36/91/58/73 | $253.5B | 85% | Best growth and quality. Tape was not as strong as the fundamentals in this pull. |
| 3 | MU | 73.3 | Watch/add | HBM/memory | 71/96/100/78/71/56 | $90.3B | 346% | HBM and server DRAM leverage. Treat the growth score as cyclical, not permanent. |
| 4 | ASML | 67.3 | Watch/add | Equipment | 82/40/77/80/39/78 | $38.5B | 13% | EUV monopoly. Lower growth score, high scarcity score. |
| 5 | KLAC | 66.0 | Watch/add | Equipment | 82/33/93/81/34/66 | $13.1B | 12% | Process control and yield learning. A quality way to own node difficulty. |
| 6 | LRCX | 63.6 | Trade/watch | Equipment | 68/50/97/73/36/66 | $21.7B | 24% | Etch/deposition exposure to memory and foundry capacity. |
| 7 | AMAT | 62.8 | Trade/watch | Equipment | 68/39/100/70/41/69 | $29.0B | 11% | Broad wafer-fab equipment exposure. Less pure, but captures several AI capex lanes. |
| 8 | CRDO | 62.5 | Trade/watch | Optical/interconnect | 73/81/100/71/34/40 | $1.3B | 157% | High AI cluster leverage. Also real customer concentration and young-company risk. |
| 9 | AVGO | 62.0 | Trade/watch | Custom silicon | 50/83/42/88/48/72 | $75.5B | 48% | Custom AI silicon and networking with strong cash generation. |
| 10 | VRT | 61.8 | Trade/watch | Power/cooling | 70/63/96/56/44/58 | $10.8B | 30% | Liquid cooling and power infrastructure. Strong theme, valuation sensitive. |
| 11 | GOOGL | 61.4 | Trade/watch | Hyperscaler | 50/64/53/81/58/82 | $422.5B | 22% | TPU, cloud, search and buybacks. Best Mag 7 quality/value blend after NVDA. |
| 12 | GEV | 61.4 | Trade/watch | Grid/power | 76/56/81/50/47/75 | $39.4B | 16% | Grid and generation equipment. Direct read-through to the watts bottleneck. |
The model's strongest message is not "buy the top five tomorrow." It is that the AI buildout should be underwritten through the supply chain first. TSM is the purest bottleneck. NVDA still has the best growth and quality mix. MU has the most violent growth signal, but memory cyclicality must be haircut. ASML and KLAC have slower reported revenue growth, but they own critical process layers.
Mag 7 Are Included, but They Do Not All Rank the Same
Mag 7 exposure matters because these companies fund and monetize the buildout. But the model separates demand owners from bottleneck owners. NVDA is both. GOOGL has TPU optionality and reasonable quality-adjusted valuation. META and MSFT have elite businesses but scored weaker on recent price momentum. AMZN and AAPL are better businesses than their scanner rank implies, but their direct AI infrastructure leverage is lower. TSLA is a physics-and-autonomy option, not a clean current earnings compounder.
| Rank | Ticker | Score | Verdict | Bucket | C/G/M/Q/V/R | Revenue | Rev growth | Read |
|---|---|---|---|---|---|---|---|---|
| 1 | NVDA | 73.5 | Watch/add | GPU/system | 84/100/36/91/58/73 | $253.5B | 85% | The only Mag 7 name that is both demand owner and hard bottleneck owner. |
| 2 | GOOGL | 61.4 | Trade/watch | Hyperscaler | 50/64/53/81/58/82 | $422.5B | 22% | Demand owner with TPU optionality and reasonable quality-adjusted valuation. |
| 3 | META | 58.4 | Trade/watch | AI application | 50/70/18/87/73/77 | $215.0B | 33% | Elite cash engine. Capex ROI is the debate. |
| 4 | MSFT | 57.9 | Trade/watch | Hyperscaler/software | 55/51/10/89/69/85 | $318.3B | 18% | Highest durability, but weaker price momentum in this run. |
| 5 | AMZN | 53.5 | Prove-it | Hyperscaler | 50/59/32/62/65/77 | $742.8B | 17% | AWS matters, but FCF conversion has to keep improving. |
| 6 | AAPL | 52.6 | Prove-it | Edge AI/device | 38/50/47/75/58/78 | $451.4B | 17% | Great business, weaker direct AI infrastructure leverage. |
| 7 | TSLA | 31.3 | Prove-it | Autonomy/energy | 43/43/22/36/22/59 | $97.9B | 16% | Autonomy and robotics upside, but current margins and narrative risk drag the score down. |
How To Use It
Use the scanner monthly, then do the actual stock work only on the names that clear the queue. A score above 75 deserves underwriting. A score from 65 to 75 is a watchlist or add-on-pullback candidate. A score from 55 to 65 is a trade/watch name. Below 55, demand a specific variant view before capital goes in.
The workflow is simple: run the scanner, inspect the top names by bottleneck layer, compare the model to 13F behavior and earnings revisions, then build a real unit model. For semis, that means end demand to content per system to available capacity to yield to ASP/mix to gross margin to EPS revisions to stock reaction.
The revenue columns matter because bottleneck stories that do not become sales are just stories. The growth score pushes MU, NVDA, CRDO, AVGO, VRT, and GOOGL up because the theme is visible in reported numbers. It also warns you not to confuse TSM/ASML/KLAC's lower growth scores with weak businesses. Some toll roads monetize scarcity through durability and pricing, not explosive top-line acceleration every quarter.
What Breaks It
This framework fails if the AI buildout is already closer to overbuild than bottleneck. The kill signals are specific: hyperscaler capex guidance cuts, GPU cluster utilization falling, TSMC or CoWoS utilization weakening, HBM/DRAM contract pricing rolling over, optical orders slowing after the stocks rerate, and power/cooling backlogs failing to convert into revenue.
It also fails if the model overweights scarcity and underweights valuation. That is the classic infrastructure-cycle trap. A scarce asset can still be a bad stock if investors capitalize peak margins as if they are normal.
The fix is discipline: rerun the screen, reconcile the data to primary filings, then write the kill criteria before buying. If the stock needs perfect capex, perfect ASPs, and perfect multiple expansion, the score is lying.