Scry Fund

Market Framework

Capital Rotation and Flows

Narrative is a flow catalyst. AI capex is the current lab experiment.

24 May 2026 · YK Research

The Mispricing

“The market still says 'AI basket' when the real trade is capital rotation across bottlenecks. First GPUs, then HBM, then interconnect, then optical, then storage, then power. The edge is catching the next rotation before it becomes the consensus label.”
YK Research
2026 hyperscaler capex
~$700B
FINRA margin debt
$1.304T
Retail call-buying
2021 high
Trade type
Rotation

AI capex is now a physical-infrastructure story. The important flow question is not whether AI is real. It is which scarce input receives the next forced allocation from PMs, analysts, retail calls, and quant baskets. Once a bottleneck is named, capital can move violently because everyone can map it to tickers.

Theory of Edge

The market pays us for identifying flow/narrative mismatch before it is benchmarked, screened, and packaged into a clean basket.

  • Structural reason: PMs need liquid expressions for every new AI capex bottleneck.
  • Behavioral reason: investors anchor to the old label, then chase once price validates the new story.
  • Institutional reason: factor/basket allocators are slow to reclassify exposures like copper vs interconnect or telecom vs AI optical.
  • Tradeable edge: buy misclassified beneficiaries, avoid crowded names where the narrative is true but fully capitalized.

What This Is Not

This is not a blind long-AI thesis. The easy part is saying the capex wave is real. The hard part is separating genuine capex capture from keyword beta.

  • Not every AI supplier has pricing power.
  • Not every pumped stock is early; many are already crowded.
  • Not every power name is an AI winner; grid services and equipment are different from merchant power.
  • Not every server company deserves the same multiple; trust and execution matter.
Capital Rotates Toward the Next BottleneckThe mistake is treating AI as one factor. The trade is a sequence of scarcity discoveries.GPUHBMNetworkingOpticalStoragePowerYK Research framework: flows chase whichever input makes the next $1 of hyperscaler capex possible.

1. Real Capex

CNBC reported Microsoft, Amazon, Alphabet, and Meta could spend nearly $700B combined in 2026, up more than 60% from 2025 levels. Alphabet guided $175B-185B. Amazon pointed to roughly $200B, mostly AWS/data centers. This is real corporate spend, not just online noise.

2. Levered Retail Flow

CNBC/CBOE reported retail call buying in Mag 10 names at the heaviest 10-day clip since 2021. 52% of new positions were call-buying vs 17% call-selling. That is a flow accelerant, especially when dealers must hedge upside.

3. System Leverage

FINRA margin debt reached $1.304T in April 2026, the highest value in the checked dataset, +53.3% YoY. It does not prove AI-specific demand, but it explains why narrative rotations can become violent melt-ups.

Current Rotation Buckets

Storage / HDD

What moved

STX +183% YTD, WDC +158% YTD.

Flow read

The market discovered AI needs data storage: checkpoints, logs, retrieval systems, training datasets, warm/cold storage.

What breaks it

Already violent moves. Need prove pricing/backlog durability, not just narrative catch-up.

Optical / Scale-out

What moved

LITE +145%, CIEN +137%, COHR +94% YTD.

Flow read

Bandwidth between racks, clusters, and data centers is becoming a named bottleneck. Goldman framed optical networking as a 9x TAM unlock to $154B.

What breaks it

Telecom beta masquerading as AI datacom exposure. Need revenue-mix proof.

HBM / DRAM

What moved

MU +138% YTD.

Flow read

Memory finally trades as scarcity, not just cyclicality. HBM is the gating input for accelerators.

What breaks it

If HBM supply catches up or hyperscaler orders pause, memory becomes cyclical again fast.

Interconnect / Scale-up

What moved

MRVL +120%, ALAB +71%, CRDO +53% YTD.

Flow read

The interesting trade is misclassification. ALAB may not be a copper loser if the real exposure is retimer/switch/interconnect content.

What breaks it

Copper-vs-optical framing may still matter if topology shifts reduce content per rack. Product-level mapping required.

Servers

What moved

DELL +133%, HPE +57%, SMCI +15% YTD.

Flow read

AI server demand is real, but market prefers institutional trust and execution. DELL became the clean proxy.

What breaks it

Server margins are thinner than chip/networking margins. Keyword exposure is not enough.

Power / Grid

What moved

PWR +65%, GEV +53%; VST -5%, CEG -20% YTD.

Flow read

Power is real but fractured. Grid equipment/services map better to capex than generic power-producer baskets.

What breaks it

Rates, regulation, permitting, commodity exposure, and valuation can swamp the AI story.

How to Trade the Framework

  • Start with the capex line item: what physical input does the next dollar buy?
  • Map that input to public tickers by revenue exposure, not lazy sector labels.
  • Check whether price has already moved more than earnings revisions.
  • Look for analyst/basket lag: mixed ratings despite strong price action can mean the narrative is still being institutionalized.
  • Prefer catalyst windows: earnings, guide raises, customer wins, analyst days, backlog disclosure.

Best Research Lanes

  • ALAB: prove or disprove the copper-loser misclassification.
  • STX/WDC: separate real storage shortage from narrative melt-up.
  • LITE/COHR/CIEN: identify who has true AI datacom leverage vs telecom recovery.
  • PWR/GEV vs VST/CEG: power-capex capture vs commodity/rate exposure.
  • DELL/HPE/SMCI: understand which AI server revenue converts into durable margins.

What Breaks the Thesis

Fundamental breaks

  • Hyperscalers cut, defer, or slow capex.
  • Cloud AI revenue fails to justify spend.
  • FCF/debt concerns dominate the growth narrative.
  • NVDA guide implies digestion rather than shortage.
  • HBM, optical, or storage supply catches up faster than expected.

Flow breaks

  • Retail call-buying exhausts.
  • Dealer hedging reverses from accelerant to air pocket.
  • Margin debt rolls over and forces de-risking.
  • Correlations re-collapse into one broad AI risk factor.
  • Misclassified names get correctly bucketed after the easy rerating is done.
Sources: CNBC hyperscaler capex and CBOE retail options reporting; FINRA margin statistics; Goldman Sachs Research optical networking framing; Yahoo/yfinance price and recommendation snapshots, 22 May 2026.