The AI Stack Inside ARK, Graded Layer by Layer

Nearly half of ARKW, and about a quarter of ARKK and ARKF, sits somewhere in the AI value chain by our mapping. Graded on one ruler, the stack has a shape: high at the silicon and platform layers, lowest in the capital-heavy middle where compute capacity is built, and out of reach entirely at the model layer.

The map

We grouped ARK's AI-related holdings into five layers. The mapping is ours, and broad platforms such as Meta and Amazon do much more than AI — we include them because AI spending and deployment now drive a large part of their scorecards.

Weights as of 11/04/2026.

The shape: a barbell with a dip in the middle

Silicon: one A, then a cluster just below the line

NVIDIA, at 87, is the only A-grade company in any ARK fund — our scorecard describes a moat that is finished and a creative destruction that is done. Below it, the challengers cluster tightly: AMD (62), whose scorecard notes 6GW commitments each from OpenAI and Meta; Broadcom (61), with thirteen straight quarters of AI semiconductor acceleration; Cerebras (57), an inference challenger whose demand is proven and delivery pending. TSMC and Teradyne fail G2 — mature suppliers whose revenue is already too large for a ten-fold ceiling.

The gap between NVIDIA and the next silicon name is 25 points. On this ruler, the AI chip market has one finished franchise and several strong contenders, not several franchises.

Capacity: where the score dips

CoreWeave is the only pure capacity provider in ARK's published funds, and its scorecard title states the tension directly: an A-grade business pulled to C by the balance sheet. As of the October scorecard, speed scored at the top of the range (FY2025 revenue +168%, a backlog above $100B), while debt above $35B — about 4.6 times trailing revenue, press-reported — and quarterly interest of roughly five times adjusted operating income drove heavy demerits. The model reads capacity built on borrowed money as speed bought with debt.

This is the dip in the barbell. Capacity is the most capital-intensive layer, and the Blitz Index's demerit section is built to price that capital.

Models: outside the point scale

OpenAI is held in all three funds at about 4% each, larger than most scored holdings. Our company index scores it as a range — 51 to 68, C to B — because inputs are press-reported; the index line describes the fastest revenue scale in history with unit economics moving the wrong way. The Lens excludes range scores from fund averages, so the model layer is visible in the weights and invisible in the averages.

Platforms: the highest layer

Meta (78) carries the deepest moat on our record; Palantir sits exactly on the 65 line. Alphabet — held as GOOG — scores 83 on our index. Amazon fails G2. Where AI is deployed into an existing distribution network, the moat section does most of the work, and the scores are the highest in the stack.

Data and infrastructure software: solid, mostly below the line

Seven names average 62.7. Snowflake (76) stands out, with growth re-accelerating on AI workloads; Datadog sits at 66 after twelve straight quarters of 24%+ growth, and Cloudflare at 65. The rest — Figma, CrowdStrike, Rubrik, GitLab — score between 52 and 62, mostly B grades whose scorecards point to demerits, deceleration or a second curve not yet in revenue rather than to weak businesses.

What the shape suggests

The Blitz Index rewards proven margin and lasting moats, and it charges for leverage. Applied to the AI stack, that produces a predictable pattern: the companies that own a chokepoint (NVIDIA) or a distribution network (Meta, Alphabet) score highest; the companies building capacity on borrowed money score lowest; and the companies whose economics are still private cannot yet be scored at all.

ARKW holds the most of four of the five layers (ARKF holds slightly more OpenAI). That is one reason its average (65.3) sits just above the line — its AI exposure spans both ends of the barbell.

What the components say: the dip is a demerit story, not a growth story

The layer averages above are totals. Breaking them into the scorecard’s four components — A for moat, B for speed, C for creative destruction, D for demerits — shows where the barbell actually comes from. No layer is dragged down by a weak moat: silicon averages 20.0 on the A component, platforms 20.5 (Meta 20, Alphabet 21), and data and infrastructure software 17.4. The moat section is the most even part of the stack.

The spread lives in B and D. Silicon’s B average of 39.0 is almost entirely NVIDIA’s 52; the other three silicon names score 32 to 36 on B, a tight cluster that matches their tight total scores (57 to 62). Data software averages 38.7 on B — nearly the same as silicon once NVIDIA is set aside — and 10.6 on C. In other words, the challengers at the silicon layer and the software layer are growing at roughly the same measured pace; what separates them is a few points of moat and a few points of demerits, not a different growth regime.

Capacity is the exception that proves the pattern. CoreWeave’s components are A 21, B 37 and C 10 — at or above the data-layer averages on all three. Its total of 53 is lower than every data name except GitLab for one reason only: D of −15, the joint-deepest demerit in the stack. The capacity dip is not a judgment that compute demand is weak or that CoreWeave is slow. It is the model charging, in a single column, for debt above $35B and quarterly interest of roughly five times adjusted operating income. Remove the D column and CoreWeave would sit at 68, above the data-layer average. That is worth stating plainly, because it is also the most likely way the score moves: growth is already scored near the top of its range, so the upside in the total comes from the balance sheet, not from faster revenue.

Platforms show the mirror image. Meta’s B of 53 and Alphabet’s B of 52 are the two highest speed scores in the stack, and Alphabet’s C of 15 is the highest creative-destruction score anywhere in it. Their D columns (−6 and −5) are ordinary. Palantir’s components are not published in the score record — history carries its total of 65 only — so no platform component average is computed here. The platform layer’s lead, on the evidence available, is a speed lead: distribution networks converting AI deployment into measured growth, with moats already in place.

How much of each fund the stack actually is

The totals row of the first table deserves a second look, because the three funds do not hold the stack in anything like the same proportion. ARKW holds 48.8% of its fund in these five layers — nearly double ARKK’s and ARKF’s 25.6% each. The gap is not in silicon (14.0% in ARKW against 10.0% and 4.8%) or in models (3.8% against 4.1% and 4.4%). It is in platforms, where ARKW holds 15.5% against ARKK’s 8.0%, and in data and infrastructure software, where ARKW holds 12.3% against ARKK’s 1.1% and ARKF’s 3.2%.

That last figure is the most lopsided in the table. ARKK’s AI exposure is silicon, platforms and models; software barely registers at 1.1%, even though the layer holds seven scored names. ARKF sits between the two, with 13.2% in platforms — its largest AI layer — and 3.2% in software. Read alongside the layer averages, the shape of each fund’s AI bet differs: ARKW owns the full barbell, ARKK owns its two ends minus the software middle, and ARKF’s AI exposure is concentrated in the highest-scoring layer, platforms, plus the unscored model layer.

Two cautions belong here. First, the model layer’s roughly 4% in each fund is the largest single-layer weight that contributes nothing to any fund average, because the Lens excludes range scores. A reader who looks only at averages will miss about a twenty-fifth of each fund. Second, capacity — the lowest-scored layer — is also the smallest in every fund (2.4% to 3.3%). The dip in the barbell is real in score terms and small in weight terms; it is a statement about how the model prices leverage, not a large position.

What to watch

Four things would change this picture, and each has a clear trigger. The widest uncertainty is OpenAI’s 51–68 range — seventeen points, the broadest in the stack. Narrowing it needs audited figures rather than press reports, and until that happens the model layer stays visible in weights and invisible in averages. The two G2 failures in silicon, TSMC and Teradyne, would need a market definition that clears ten times current revenue — a re-examination the judgment files, not the scorecard, would carry. CoreWeave’s path runs through D: deleveraging or interest coverage, since its growth components are already scored near the top. And every name here is re-examined after its next quarterly earnings report, when the October scorecards this article uses are replaced — the layer averages above are a snapshot of that vintage, not a permanent shape.

The boundary

The layer mapping is a simplification; Meta, Alphabet and Amazon run large non-AI businesses, and Palantir and Snowflake could reasonably sit in different layers. Scores are October 2026 scorecards and are re-examined after each earnings report. Nothing here says anything about AI stock prices, which our research finds are governed by entry multiples, not by business scores.

BlitzLab scores businesses against a scholar's standard — not investment advice. No price targets, no buy/sell recommendations. Holdings come from ARK Invest's daily public disclosures; BlitzLab is not affiliated with ARK Invest and this research is not endorsed by them. Company figures are as labeled on each scorecard.

Sources

1. BlitzLab, ARK Lens fund pages — https://blitzlab.app/ark/arkk/ · https://blitzlab.app/ark/arkw/ · https://blitzlab.app/ark/arkf/ — holdings as of 11/04/2026 — checked 2026-11-05
2. BlitzLab, "All Company Scores" (NVDA, AMD, AVGO, CBRS, META, PLTR, SNOW, DDOG, NET, FIG, CRWD, RBRK, GTLB, OpenAI range) — https://blitzlab.app/companies/ — checked 2026-11-05
3. BlitzLab, "CoreWeave — Blitz Index" scorecard — https://blitzlab.app/companies/crwv/ — Oct 2026 — checked 2026-11-05
4. BlitzLab, "Alphabet — Blitz Index 83 (A)" — https://blitzlab.app/companies/googl/ — checked 2026-11-05