Who's Actually Selling The Shovels? A Fundamentals-Only Map Of The AI Infrastructure.

Ask ten people whether the AI infrastructure buildout is in its early innings, its middle, or its late stage, and you'll get ten confident answers and zero shared evidence. That's because the question, as usually asked, is a forecasting question — "how much more capex is coming" — and forecasting isn't something this tool does or claims to do. DUEL has no view on future AI spending, no analyst estimates, and no macro model.

What DUEL does have is something narrower and more checkable: a standardized read of what the already-published financials of the companies building AI infrastructure say about growth, margin, capital efficiency, valuation math, and balance-sheet resilience, all sourced from the same place — SEC filings — and run through the same formulas for every company involved. So instead of asking "where are we in the cycle," this piece asks a version of that question DUEL can actually answer with data instead of opinion:

What do the numbers already on the books say about who is actually capturing the AI infrastructure buildout right now — and how much of that capture is already priced into DUEL's own valuation math?

Part 1 — Building the Basket

"Picks and shovels" is the natural framing for AI infrastructure investing: instead of betting on which AI application wins, buy the companies supplying the compute, networking, power, and storage that every AI workload needs regardless of who wins. We built a basket across four supply layers, plus a fifth group that sits on the other side of the transaction — the hyperscale buyers — as a control group rather than a set of "beneficiaries":

• Compute: NVDA, AMD, AVGO, MU

• Networking / interconnect: ANET, CIEN

• Power & cooling: VRT, ETN, MOD, GEV

• Memory & storage: MU (shared with Compute), WDC

• Buyers (control group, not scored as "beneficiaries"): MSFT, GOOGL, AMZN, META

One name is conspicuously absent. TSM was excluded entirely. As a Taiwan-domiciled foreign private issuer, TSMC files a 20-F annual report with the SEC, not a 10-K, and does not file 10-Qs. DUEL's entire model is built exclusively on 10-K/10-Q data, by design — so a company that structurally doesn't file those forms is outside what this tool can measure, no matter how central it is to the actual value chain. We'd rather flag that gap plainly than quietly patch it with a number from somewhere else.

That leaves 11 unique "shovel sellers" and 4 buyers — 15 companies total, each run through DUEL's Battle Report, DCF Report, and Resilience Report exactly as any other duel on the site. All duels behind this piece were generated between 26 July and 13 August 2026. This is a snapshot of that window, not a permanent ranking — the methodology stays valid long after any individual number goes stale.

Part 2 — Growth and Margin: Where the Capital Is Actually Landing

No single number here reads like "the cycle is over." Nine of eleven names are still growing revenue in double digits, and the two networking names — ANET (27.2%) and CIEN (9.5%) — sit at opposite ends of that range within the same layer, which is itself informative: interconnect demand is not lifting every networking name equally. NVDA's 100% figure is an outlier by an order of magnitude against its own compute-layer peers — a reminder that "AI compute demand" and "NVIDIA's revenue growth" are not the same variable, even though headlines often collapse them into one.

The margin picture is where the basket splits into two visibly different businesses. Compute and networking names (NVDA, AVGO, ANET) post FCF margins in the 42–48% range — software-like cash conversion despite being hardware companies, largely because the actual manufacturing is outsourced. Power and cooling names (VRT, ETN, MOD, GEV) post FCF margins between 2% and 19% — capital-intensive, lower-margin businesses by nature, closer to traditional industrials than to chip design. WDC is the one name with negative revenue growth in the entire basket (-20.3%), reflecting a cyclical trough in NAND/HDD pricing that has nothing to do with AI-specific demand — a reminder that "storage" as a category is exposed to more than one demand cycle at once.

Part 3 — How Much Growth Is Already in the Price

This is the part that actually engages with the "cycle stage" question — indirectly, and on DUEL's own terms. DUEL's DCF engine builds a fair value and a 3-year price projection from each company's own filings alone: no market price as an input, no analyst growth estimates, just revenue CAGR, FCF, ROIC, and Sloan Ratio run through a fixed formula (full methodology in our earlier [DCF confidence-band piece]). The resulting 3-year modeled upside is not a price target — it's a measure of how much of a company's own historical trajectory, projected forward under the model's fixed rules, is already reflected in the model's own fair-value estimate versus how much room the math still shows.

NVIDIA — by a wide margin the fastest-growing, highest-ROIC name in the entire basket — lands 12th out of 15 on modeled upside. That is not a coincidence of the model; it is the direct arithmetic consequence of NVDA's own enterprise value already being large enough, relative to its own projected cash flows, that compounding a fixed 3-year growth path forward produces a smaller percentage move than doing the same exercise on a smaller base. This is the same pattern our first piece in this series found in a single NVDA-vs-AMD duel back in early August; running it across an 11-company infrastructure basket instead of one pair shows it isn't an artifact of that specific comparison.

Two validity notes worth flagging plainly. First, MU's fair value ($38.29), 3-year price ($43.06), and every input behind them appear identically whether MU is compared against AMD or against WDC — direct, in-the-wild confirmation that DUEL's per-company DCF outputs don't depend on who the opponent is, exactly as the model is designed to behave. Second, GEV's modeled upside (20.7%) sits near the top of this table, but should be read alongside Part 4 below before drawing any conclusion from it in isolation.

Part 4 — Who Actually Has the Balance-Sheet Armor

This is where the "beneficiary" story gets more complicated than a single growth chart suggests. Two of the four power-and-cooling names — MOD (31) and GEV (11) — land in DUEL's LOW resilience tier, the same category the model reserves for companies whose balance-sheet cushion looks thin relative to their cash-flow demands, regardless of how strong the underlying growth narrative is. GEV in particular is worth a data-quality caveat rather than a verdict: several of its raw balance-sheet components (cash, total debt, operating income, equity) were unavailable in the filing period pulled, most plausibly because GE Vernova's SEC filing history is still short following its 2024 spin-off from GE. Its LOW tier should be treated as provisional until fuller post-spin-off filings accumulate — a company with a thin filing history will mechanically look thinner on components the model can't yet see, independent of its actual financial condition.

ETN is a cleaner case: its Debt-vs-Cash consistency check triggered a genuine RISK flag — debt-to-equity above 1.0x with cash covering essentially none of total debt — a straightforward leverage signal on an already-mature industrial balance sheet, not an accounting-quality concern. AVGO's single CAUTION flag is the same story at a smaller scale.

Meanwhile, the highest resilience scores in the basket don't belong to the highest-growth or lowest-upside names — they belong to AMD, ANET, and VRT, three companies with meaningfully different growth profiles and margin structures, unified only by clean, contradiction-free balance sheets across all three of DUEL's internal-consistency checks.

Part 5 — The Scorecard: Upside vs. Resilience

Putting Parts 3 and 4 on the same page — modeled 3-year upside on one axis, resilience tier on the other — sorts the infrastructure names into four groups that no single table above shows on its own. The split point on each axis is the median calculated across the 11 supply-side companies only (17.65% upside, FRI 75); the four hyperscale buyers are deliberately left out of that calculation and shown separately in Part 6, so their much larger, more diversified balance sheets don't quietly drag the supplier median around.

Priced attractively and structurally sound (above-median upside, above-median resilience): AMD, ANET, CIEN

Modeled upside, thinner cushion (above-median upside, below-median resilience): WDC, GEV, MOD

Already priced in, but rock-solid (below-median upside, above-median resilience): AVGO, MU, VRT

Already priced in, comparatively thinner cushion (below-median upside, below-median resilience): NVDA, ETN

Only two names land in that last, least-favorable quadrant — and NVIDIA is one of them, next to ETN, a mature industrial that this same analysis flagged with a genuine leverage RISK note in Part 4. That's the real surprise here: the name every headline treats as the obvious "AI winner" sits in the quadrant this basket's own math associates with the least remaining modeled room and the least resilience cushion — not because anything is wrong with the business, but because both its growth and its enterprise value have already run far enough that the model's own arithmetic has less distance left to cover, on both axes, than for nine of the other ten suppliers in the chain it sits at the center of.

Part 6 — The Control Group: What the Buyers Look Like

All four hyperscale buyers land in DUEL's HIGH resilience tier, unlike their suppliers, where resilience ranged from 11 to 100. That's expected — these are four of the largest, most diversified cash generators in the entire market, buying AI infrastructure as one line item among many rather than depending on it the way a pure-play supplier does. What's less expected: their modeled upside (16.2%–20.6%) sits squarely inside the same range as the supplier basket, not obviously higher or lower. Being the customer in this transaction doesn't structurally produce a different valuation-math signature than being the supplier — at least not one this snapshot can detect.

Part 7 — What This Actually Answers, and What It Doesn't

To be precise about the limits of this exercise, since overclaiming here would undercut the whole point: this piece does not, and cannot, tell you whether AI infrastructure spending has further to run, is plateauing, or has peaked. That's a question about the future, and every number above describes the past and present only.

What the data does show, on its own terms: trailing growth rates across this basket are still high and still dispersed — nothing here resembles a sector where growth has synchronized and flattened, which is typically what a "late and maturing" cycle looks like in the fundamentals rather than in the headlines. It also shows that resilience is genuinely uneven beneath a growth story that gets told as if it applies equally to everyone in it, and that DUEL's own no-market-price DCF math is already assigning less remaining modeled room to the most-discussed name in the space than to most of the rest of the chain around it — which is a statement about today's starting point for that specific company, not a prediction about tomorrow.

Limitations

TSM excluded entirely, and for a structural reason, not a data-quality one: as a foreign private issuer, it files a 20-F rather than a 10-K, which places it outside what DUEL's SEC-EDGAR-only pipeline is built to read, regardless of its real-world importance to the compute layer.

Snapshot, not a live ranking. Every underlying duel was generated between 26 July and 13 August 2026; the specific numbers above will drift with each new quarterly filing, though the comparative method they came from will not.

No market price, on purpose — and that cuts both ways. DUEL's DCF engine deliberately ignores current share price and analyst estimates, which is exactly what makes "how much growth is already assumed by the company's own trajectory" a clean, checkable question. It also means "modeled upside" here is not a market-implied expectation and should not be read as one.

Basket construction is a judgment call, not a census. Eleven infrastructure names and four buyers is a reasonable cross-section of the AI value chain, not an exhaustive one; a different reasonable basket could shift which names land in which scorecard quadrant.

GEV's resilience score should be read with a data-completeness caveat, flagged directly in Part 4 — several raw balance-sheet inputs were unavailable given the company's short filing history since its 2024 spin-off, and its LOW tier may partly reflect that gap rather than a fully resolved picture of its financial condition.

Fifteen companies is a diagnostic snapshot, not a statistical sample — read the patterns above as descriptive, not as evidence generalizable beyond this specific basket.

Bottom Line

Across eleven AI infrastructure suppliers and four of their biggest buyers, growth is still broadly strong and still unevenly distributed — not a picture of a plateaued sector. Resilience, on the other hand, does not track the growth story at all: some of the most-hyped growth names carry the thinnest balance-sheet cushions in the basket, and some of the least-discussed names carry the strongest. And the one figure that most directly engages the "how much is already priced in" question — DUEL's own opponent-independent, market-price-free modeled upside — currently assigns the least remaining room to the name most people would call the obvious center of this entire cycle. None of that is a forecast. It's what the filings already say, run through the same formula for every company in the room.

Every number in this piece traces back to a duel you can pull yourself and re-verify at duelstocks.

Further Reading — Same Series, Same Method

This piece builds directly on methodology established across six earlier deep-dives in this series:

• Absolute vs. Relative Valuation: A Tale of Two Questions

• Beyond the Pair: Ranking a Full Sector With Paired Comparisons

• A Point Estimate Isn't a Forecast: Building a Confidence Band for DCF

• How Fragile Is a Relative Stock Score?

• How Many Independent Signals? A Principal Component Analysis

• The Flag Nobody Talks About: A Base-Rate Analysis of Internal Contradictions

References

Damodaran, A. (2012). Investment Valuation: Tools and Techniques for Determining the Value of Any Asset (3rd ed.). Wiley.

U.S. Securities and Exchange Commission — Foreign Private Issuer reporting requirements (Form 20-F vs. Form 10-K), sec.gov.

A https://duelstocks.com fundamentals deep-dive. Not investment advice, and — unlike most AI-cycle commentary — not a forecast either. All figures come from public SEC EDGAR filings (10-K/10-Q) as processed by DUEL's Battle, DCF, and Resilience (STR) reports.