Looking forward

1. Wed Aug 26 · after close

   NVIDIA earnings

   catalyst

   Guidance on data-center and China. The whole SMH complex gaps with it. Do not add into the print.

2. Thu–Sat Aug 27–29

   Jackson Hole · financial innovation theme

   watch

   Not an AI print, but a liquidity print. Duration, the dollar, and risk appetite reprice the multiple on chips.

3. Ongoing

   Hyperscaler custom silicon + capex circularity

   bearish

   The bear case does not need a new headline. If Google/Amazon/Microsoft keep pulling training onto their own ASICs, NVDA's multiple compresses even with 'fine' numbers.

4. Next foundry print

   TSMC utilization / CoWoS capacity

   catalyst

   If the foundry is tight, the pick-and-shovel bid holds. If it is not, the AI build-out is digesting.

Vehicles, not edges

ETFs the desk actually uses

• VanEck Semiconductor SMH

  ETF

  560.92

  +0.00%5D -4.79%

  Cleanest one-ticket for the AI silicon stack — NVDA, TSM, AVGO, ASML. Use it when you want the trade without single-name gap risk into the NVIDIA print.

  Risk. Top-heavy. If NVDA gaps down, SMH gaps with it. This is not diversified tech; it is a chip bet.

  Confirm. Holds the 50-day after Aug 26 earnings. Relative volume > 1.2 on up days vs QQQ.

• iShares Semiconductor SOXX

  ETF

  519.67

  +0.00%5D -5.64%

  Same neighborhood as SMH, slightly different weights. Pair it only if you already live in semis and want a second vehicle, not a second thesis.

  Risk. Same factor as SMH. Owning both is not diversification.

  Confirm. Tracks SMH without leaking. If it lags hard, the tape is picking winners inside the stack — go to single names instead.

• Global X Robotics & AI BOTZ

  ETF

  36.38

  -0.00%5D -3.81%

  The 'robots and factory AI' sleeve, not the GPU melt-up. Useful when you think inference and automation outlast the training-capex spike.

  Risk. Can sit still while NVDA rips. International industrial exposure. Not a substitute for SMH.

  Confirm. Starts beating XLK on a 20-day basis. Until then it is a satellite, not a core.

• Global X Artificial Intelligence AIQ

  ETF

  62.99

  -0.00%5D -2.52%

  Broader AI basket (software + hardware). A way to own the theme without betting the farm on one foundry or one GPU vendor.

  Risk. Diluted. You will lag a NVIDIA melt-up and still eat a capex air-pocket.

  Confirm. 20-day relative vs QQQ turns positive and stays there after the NVIDIA print.

Watchlist — not a buy list

Early names worth watching

• Micron MU

  Early

  937.11

  -0.00%5D -1.34%

  HBM memory is the less-owned pick-and-shovel. Training clusters are memory-bound. Early relative to NVDA, still a real company with a cycle.

  Risk. Memory is a brutal cycle. If HBM pricing rolls over, MU does not get a 'theme' bid — it gets a cycle bid, which goes both ways.

  Confirm. HBM commentary and a 20-day relative vs SMH that is not leaking. Do not size it like NVDA.

• Constellation Energy CEG

  Early

  274.17

  -0.00%5D -1.60%

  The power behind the AI build. Hyperscalers need electrons, not just GPUs. A way to own the trade if you think the bottleneck has moved from silicon to the grid.

  Risk. It still trades like a utility on rate days. A Warsh shock hits the multiple. Nuclear headlines gap both ways.

  Confirm. Holds relative vs XLU on a risk-off day — that is the 'AI power is real' tell. Size as a satellite to SMH.

• Vertiv VRT

  Early

  261.00

  0.00%5D -9.08%

  Cooling and power infrastructure for the same campuses. Pick-and-shovel without a GPU multiple. Useful when you believe the build-out continues even if NVIDIA's print is messy.

  Risk. Orders can pause as fast as they ramped. Valuation already prices a long cycle. Not a defensive stock.

  Confirm. Backlog holds after Aug 26 and the stock is not a one-day wonder vs SMH.

• Astera Labs ALAB

  Early

  289.02

  +0.00%5D -10.46%

  Connectivity silicon inside the rack. Early, expensive, and levered to hyperscaler design-wins. A sliver, not a core.

  Risk. Customer concentration and a rich multiple. A capex air-pocket hits ALAB harder than TSM.

  Confirm. You can afford to be wrong. Design-win commentary, not a chart, is the confirm.

• Arm Holdings ARM

  Early

  249.34

  +0.00%5D -10.52%

  Royalty on every AI-adjacent CPU/NPU design. Less capex, more torque to unit growth. Early vs the GPU complex, still priced as a story.

  Risk. Multiple, customer concentration, and China. A royalty model does not save you from a de-rate.

  Confirm. Royalty growth holds while GPU names chop. That is when ARM earns a core slot.

• Palantir PLTR

  Early

  175.19

  -0.00%5D -2.13%

  Software expression of the same spend — platforms, not chips. Commercial + government. A different way to lose (or make) money than SMH.

  Risk. Valuation and narrative. It trades like a momentum name, not a software multiple. Size accordingly.

  Confirm. Commercial remaining deal value keeps compounding after the NVIDIA print. Price holding the 50-day on quiet days.

• IonQ IONQ

  Early

  43.36

  -0.00%5D -3.60%

  Trapped-ion quantum. This is a lottery ticket on a 2030s platform, not an AI-capex peer. Tiny, binary, and noisy — only as a sliver.

  Risk. Can go to zero. Dilution, technology risk, and no earnings ballast. Do not confuse a theme with a business.

  Confirm. You can afford to lose the entire line. Anything else is the wrong size.

• CrowdStrike CRWD

  Early

  201.63

  -0.00%5D -10.60%

  Cyber spend follows the AI attack surface. Quality compounder, not a chip. Useful when you want AI adjacency without GPU valuation.

  Risk. Competition, multiple, and a growth scare. Not a replacement for SMH.

  Confirm. Net-new ARR holds and the stock is not a momentum orphan vs XLK.

The names the theme still runs through

Core tape

• NVIDIA NVDA

  Core

  217.56

  +0.00%5D -3.44%

  Still the core of the training trade. The Aug 26 print is the near-term tell. Own it as a position you already have, not as a chase into the number.

  Risk. Valuation, circular financing with customers, custom silicon from hyperscalers, export controls. A miss on data-center guidance rewrites the whole complex.

  Confirm. Guidance holds, China is not a new shock, and the stock reclaims the 50-day on volume.

• Broadcom AVGO

  Core

  362.48

  -0.00%5D -13.24%

  Custom ASICs plus networking. The pick-and-shovel if hyperscalers build their own silicon instead of renting NVIDIA's.

  Risk. Customer concentration. If capex pauses, custom orders pause with it.

  Confirm. Holds relative vs NVDA on a down NVIDIA day — that is the 'custom silicon is real' tell.

• AMD AMD

  Core

  466.42

  -0.00%5D -3.43%

  The other GPU. Cheaper than NVDA on the tape, more to prove on MI-series share. A way to own the trade if NVIDIA is extended.

  Risk. Share losses, software stack, and TSMC capacity. It can underperform for quarters while the theme is still 'fine'.

  Confirm. Outperforms SMH for 20 days and does not give it back into NVDA earnings.

• TSMC TSM

  Core

  412.09

  +0.00%5D -4.27%

  The foundry every GPU and ASIC still needs. If you believe in AI silicon and do not want to pick a vendor, this is the boring expression.

  Risk. Taiwan geopolitical risk is not a headline — it is the risk. A scare hits TSM first and hardest.

  Confirm. Utilization commentary stays tight and the ADR holds the 50-day through Jackson Hole.