Market Signal

The finding: NVIDIA's quarterly revenue doubled and its stock returned 15.5% over twelve months. The memory cohort returned a 412% median over the same window. The Q2 filings, read against the cohort tape, explain why — and point at the one layer that has not been paid yet.

What actually got paid

Median one-year return by layer, to the 26 August close, across the desk's tracked buildout cohorts:

Layer Median 1y Median off-high ------:---: NAND flash memory +468.3% −27.9% Memory (broad) +412.2% −30.1% Semiconductor test & burn-in +203.5% −19.9% Nearline storage +197.7% −23.3% Optical supply chain +168.1% −38.2% HBM memory +168.1% −28.6% Advanced packaging equipment +155.0% −35.1% Data-center switch / NIC +142.7% −25.7% NVDA (single name) +15.5% −11.4% Transformers & grid interconnect +56.5% −18.4% Power distribution / busways +46.4% −18.8% Racks & mechanical infrastructure +48.0% −16.2% Data-center thermal management +32.4% −15.9% Fab construction & engineering +23.4% −10.1% UPS & energy storage for DCs +14.1% −26.4% Grid infrastructure & transmission +12.2% −13.6% Renewable hyperscaler PPAs +2.9% −16.2% Nuclear energy for data centers −1.2% −16.2% Nuclear fuel cycle −10.9% −36.5%

Two clean populations. The semiconductor supply chain got paid — memory, packaging, test, optics, equipment. The physical and electrical layer did not. NVIDIA itself sits with the second group.

The filings say the constraint just moved to the group that wasn't paid

Three disclosures, all from the 10-Q and the Q2 call on 26 August:

• Memory is the margin. Gross margin guided from 75.0% to a 71–72% Q4 trough and 72–73% in FY2028, cause named as memory pricing running past the company's own expectations. On the guided $108.0B quarter that is roughly $3.8B of gross profit per quarter, paid to suppliers. Supply and capacity commitments went $119B → $279B in one quarter, attributed primarily to memory, plus a multiyear SK hynix partnership. The tape had already voted: on the session NVDA fell 1.59%, SIMO +8.8%, WDC +4.0%, STX +3.0%, SK hynix +2.4%, Samsung +1.8%, SNDK +1.3%, MU +0.6%.
• The bottleneck is now land, power and shell. Stated in the filing, with $105B of credit support behind an OpenAI affiliate's leases on 4.25 GW in Ohio and an option on 3.8 GW more. NVIDIA revenue per gigawatt: ~$18B Hopper → ~$25B Blackwell → ~$40B Vera Rubin, against a total data-center cost management put near $60B/GW. The non-NVIDIA share of a gigawatt is still roughly $20B, and neoclouds go from ~3 GW to 8 GW installed this year.
• NVIDIA is now the credit behind its fastest-growing customers. ACIE — the neoclouds and AI natives — grew 138% YoY to $40.3B, ~45% of Data Center and heading for half, backed by $108.5B of guarantees. Every tracked neocloud closed red on 26 August, 29–53% off highs.

So the sub-lanes that have returned 2.9% to 14.1% over a year — renewable hyperscaler contracts, grid transmission, UPS and data-centre storage — are the ones a $5T buyer just called its binding constraint and put $105B behind. Stated precisely: a consolidated electricity cohort built for this piece (60 names, spanning generation through power semiconductors) has a median one-year return of +28.7%, because it also contains power-semiconductor and optical crossovers that did run — BE +336.9% 1y, VICR +282.9% 1y, TSEM +268.7% 1y, COHR +221.4% 1y. The laggards are the pure generation and grid names inside it: PRIM −33.3% 1y, VST −28.2% 1y, TLN −19.1% 1y, CEG −11.7% 1y, PEG −11.3% 1y. The independent power producers are the single worst-performing group in the whole buildout.

The specific anomaly: TE Connectivity

Screening that layer on fundamentals rather than drawdown, one name separates. All figures are trailing twelve months through the June 2026 quarter, from Massive statements cross-checked against EDGAR:

P/E TTM rev YoY TTM op income YoY Op margin (now vs 2y ago) 1y price Off high ------:---:---:---:---:---: TEL 20.1x +16.5% +29.5% 19.7% vs 17.8% −0.9% −18.8% APH 40.4x +54.2% +86.0% 27.0% vs 22.4% +46.8% −9.6% VRT 59.7x +26.2% +39.2% 18.9% vs 17.2% +106.2% −30.6% PWR 70.6x +26.3% +36.3% 6.2% vs 5.7% +60.6% −21.8% GEV 27.4x +20.7% +205.4% 4.3% vs 1.7% +52.3% −20.3% ETN 42.7x +15.0% +1.2% 16.1% vs 18.2% +19.1% −12.2% HUBB 28.0x +10.6% +2.7% 18.8% vs 20.2% +6.8% −16.4% CMI 29.5x +2.9% −6.0% 11.2% vs 12.3% +45.1% −21.6% MU 21.2x +167.0% +675.6% 65.6% vs 22.6% +706.5% −25.2%

TE Connectivity grew operating income 29.5% with margin expanding two points, and the stock is flat over twelve months at half its direct competitor's multiple.

Why it is cheap, and it is a real reason. The FY2025 10-K puts Transportation Solutions at 54% of net sales — automotive, commercial transportation, industrial equipment. The market prices TE as an auto supplier, and auto has been weak. That is the bear case and it is legitimate.

What the same filing shows underneath it. The mix is moving: Transportation Solutions went 60% → 60% → 54% across the three disclosed years while Industrial Solutions went 40% → 40% → 46%. Industrial is described in the filing as the segment supplying products that "connect and distribute power, data, and signals" — the layer NVIDIA just named.

Sized precisely, because the first pass overstated it. Industrial is not all buildout. Multiplying the disclosed end-market splits through: digital data networks is 28% of Industrial and therefore 12.9% of the company; energy — utilities, grid, generation, transmission, distribution — is 17% of Industrial and 7.8% of the company. Combined buildout-exposed revenue is about 20.7%, against automotive at 40.5%. This is a mix shift, not a pure play, and the correct claim is narrower than "the growing half is the buildout half": a fifth of the company sells into the layer NVIDIA named, that fifth is taking share, and the multiple still prices the two-fifths that is automotive.

Even narrowed, the disconnect stands. Operating income grew 29.5% with margin expanding, the June quarter set records on sales and adjusted EPS with guidance for another double-digit quarter, and the stock did not move. Full treatment in the TE Connectivity deep dive of the same date.

What the same screen says NOT to buy

Two consensus "electrification" names fail on their own numbers. ETN carries 42.7x with operating income up 1.2% and margin contracting 18.2% → 16.1%. CMI — the data-center-genset narrative — has TTM revenue up 2.9% and operating income down 6.0%, margin contracting. The story and the statements disagree, and the statements are current.

MU is optically the cheapest thing here and that is the trap, not the opportunity. 21.2x on 167% revenue growth looks free until the 65.6% operating margin is read as what it is: a cyclical peak, up from 22.6% two years ago. A trough multiple on peak memory margins is the oldest setup in the sector.

One data-integrity note. EDGAR XBRL reported HUBB diluted EPS of $4.19, which implied 112.8x. The Massive quarterly series gives TTM EPS of $16.89 and 28.0x — a 4x error that would have made a soft business look like a distressed one. The figure above is the reconciled one. HUBB is still not a candidate: operating income +2.7% with margin contracting.

Who owns this, and who doesn't

The corpus tracks 13F mirrors and congressional disclosure books. Read against them, the layer split shows up in who bought what.

TEL is owned by nobody the desk tracks. It appears in no 13F mirror and in no congressional book. It surfaced in two members' filing-page captures only as embedded issuer reference data, and both members' actual ledgers are clear of it.

Rep. Josh Gottheimer bought the layer that paid. His disclosed 2026 purchases include SNDK (+3066.6% 1y), MU (+706.5% 1y), LITE (+653.5% 1y), FN, AMD and ADI — the memory and optical supply chain, almost exactly the top of the layer table above. He sold FN on 2026-07-13; FN is now 41.5% below its high while LITE, the same layer, is 13.5% off.

Speaker Emerita Pelosi's disclosed household trades bought the platforms and the electricity, and skipped the supply chain. The book holds GOOGL, AMZN, NVDA, AAPL, AVGO, INTC, TEM, UBER — and two power names: VST, bought January 2025 (disclosed range $500K–1M) and again January 2026, and BE, bought twice in July 2026 (disclosed ranges $1M–5M and $500K–1M). Those July Bloom Energy purchases are the most recent new position in the book. There is no memory, optical or packaging exposure in it at all.

Two disclosure trails, two different bets: one on the semiconductor supply chain that had already re-rated, one on compute platforms plus the electricity layer that has not. On the twelve-month scoreboard the supply-chain trail is far ahead — but VST is −28.2% and BE is +336.9%, so the power bet is not one trade, it is a dispersion.

These are public disclosures read as a research input, not as a recommendation, and a disclosed range is a range, not a position size.

Desk Call