Conditions monitor · not a forecast

Cassandra

Is the AI build-out overheating? A transparent, real-data read of the AI / semiconductor cycle across value, fundamentals, supply chain and the macro / Fed backdrop — a composite, not a single trend line. High = stretched and fragile.

63.4

elevated

5 of 6 panels live · weights valuation 22% · fundamentals 18% · supplychain 14% · macro 22% · crowding 12% · adoption 12%

CAPE 41.58 — 99.1th percentile (richer only at the 2000 peak)Equity risk premium -2.05pp — equities priced below TreasuriesHyperscaler capex +64% YoY ($358B) — build-out acceleratingHY credit spreads 2.66% — 3.1th percentile (peak complacency)Market breadth at an extreme — equal-weight at the 1.3th percentile (narrowest in years)NVDA = 40.8% of the semis complex — single-name dominanceSemis fundamentals strong — PEG 0.91, margins +9.5pp (cushions valuation)GPU H100 rents near multi-year lows (-10%/90d)China open-weight (HF) download share 64.2%

01 / VALUATION · BROAD MARKET

CAPE & risk premium

89.1/100

Cyclically-adjusted earnings — immune to single-year EPS, ranked vs 155 years. A negative risk premium means equities yield less than Treasuries.

Shiller CAPE41.58 · 99.1ᵖᶜᵗ

Equity risk premium-2.05pp vs 10y 4.46%

02 / FUNDAMENTALS · SEMICONDUCTORS

Earnings quality

15.0/100

The counterweight: do profits justify the price? PEG<1 = growth more than pays for the multiple; expanding margins cushion rich valuations. Low score = healthy.

Median PEG0.91

Operating margin Δ+9.5pp

Earnings growth91% YoY · ROE 37%

name	fwd P/E	PEG	Δmgn
NVDA	16.4	0.65	+16.5
AMD	42.1	1.29	+3.6
AVGO	20.2	0.75	+9.5
TSM	23.8	1.44	+9.6
ASML	40.2	2.83	+0.7
SMCI	11.2	0.91	+2.9
MU	10.0	0.36	+45.6

03 / SUPPLY CHAIN · BUILD-OUT

Capex & GPU rents

74.8/100

MSFT · GOOGL · AMZN · META capex (the AI build-out) and H100 rental prices. A parabolic ramp builds capacity ahead of demand; depressed rents signal oversupply.

Hyperscaler capex$358B · +64.5% YoY

H100 rent$2.73/gpu-hr · -10%/90d

04 / MACRO & FED

Liquidity & risk appetite

50.7/100

Is the cycle being fuelled or squeezed? Low VIX & tight spreads = complacency; an inverted curve and tightening conditions are what end build-outs. FRED + CBOE, decades deep.

VIX19.8 · 61.8ᵖᶜᵗ

HY credit spread2.66% · 3.1ᵖᶜᵗ

Yield curve 10y−2y+0.27

Financial conditions-0.51 · fed funds 3.63%

05 / CROWDING & POSITIONING

Breadth & concentration

99.2/100

How narrow is the trade? A low equal-weight / cap-weight ratio means a handful of mega-caps hold up the index — thin breadth, the "who's left to buy" fragility. Orthogonal to VIX.

Equal-wt / cap-wt0.2816 · 1.3ᵖᶜᵗ

Semis vs market0.8799 · 100.0ᵖᶜᵗ

NVDA share of complex40.8%

06 / AI DEMAND · CHINA vs US

Open-model adoption

sub-score: building history

Open-weight Hugging Face download share — a real but partial demand proxy (excludes closed APIs). China 64.2% (30d) / 41.6% (cumulative).

China 64.2%US/West 35.8%

Qwen/Qwen3-0.6B	27.7M
Qwen/Qwen3-4B	16.3M
openai-community/gpt2	13.4M
Qwen/Qwen3-8B	12.9M
Qwen/Qwen2.5-7B-Instruct	12.7M

What this is — and what it isn't

Cassandra is a transparent conditions monitor, not a crash predictor. It blends six real-data categories — valuation (Shiller CAPE & the Fed-model risk premium), fundamentals (semis margins, growth, PEG), supply chain (hyperscaler capex & GPU rents), macro / Fed (VIX, credit spreads, the yield curve, financial conditions), crowding (equal- vs cap-weight breadth, semis concentration) and AI demand — into one gauge. Each sub-score is ranked against real history or scored by a transparent, documented rule; weights are a judgement (valuation 22% · fundamentals 18% · supplychain 14% · macro 22% · crowding 12% · adoption 12%), not fitted. A category without enough data is shown as "building history" and excluded — never given a fabricated score. The point is the balance: rich valuations can be offset by strong fundamentals, or amplified by a tightening macro.

Adoption caveat: the China-vs-US panel is open-weight Hugging Face download share, a real but partial proxy — it captures open-model self-host adoption (skewed toward open-weight Chinese labs) and excludes closed APIs (GPT, Claude, Gemini) that dominate actual inference. CAPE is a broad-market GAAP measure; the semis fundamentals are current snapshots with a multi-quarter margin trend.