The core thesis is that AI stocks are being priced as if compute can scale indefinitely, but the bottleneck is now in the machinery: ASML machine output, TSMC packaging, and NVDA next-gen rack deployment. The notes frame the recent pullbacks in ASML, TSM, LRCX, KLAC, and VRT as dip-buy opportunities if AI demand keeps outrunning chip-supply constraints. The sharpest risk is that real-world limits — CoWoS sold out into 2027, EUV capacity, and Kyber timing — could push the next leg of AI infrastructure growth later than the market expects. At the same time, the video argues these same bottlenecks can widen margins and entrench market power for the suppliers already inside the stack.
TSM reported July 16 revenue above $40 billion, up 34% YoY, with EPS up 77% YoY; both beat expectations and management raised full-year guidance.
TSM's CEO said CoWoS is at max capacity and sold out into 2027, with some customers waiting more than a year for chips to clear packaging.
ASML said it will build around 65 EUV machines this year and around 85 next year; each machine ships in hundreds of crates, then needs months of assembly, calibration, testing, and tuning before full-volume output.
EUV tools must run in vacuum clean rooms with essentially zero dust because air absorbs EUV light and vibration or particles can ruin chips; it can take up to 2 years from delivery to full production.
SemiAnalysis reported earlier this month that NVDA's Kyber rack has been delayed by more than a year; Kyber is the next-gen server rack with 576 GPUs and about 600 kW per rack.
NVDA denied the Kyber delay and only said its roadmap remains intact; Vera Rubin systems are in full production, but if the Kyber delay is real, Rubin Ultra could slip to 2028.
The Strait of Hormuz has been closed to shipping since the Iran war began earlier this year, and Trump reinstated the naval blockade of Iran's ports this past month, raising supply-chain risk for TSMC.
Taiwan imports over 90% of its energy, keeps less than a month of gas in reserves, and a third of the world's helium ships through the Strait of Hormuz, all inputs TSMC needs at full output.
TSM's 2nm shift from FinFET to gate-all-around is its first fundamental transistor change in over a decade and only its second in almost 40 years; 2nm chips are about 15% faster at the same power or 30% lower power at the same speed, but wafers cost around 50% more than current 3nm.
Until 2nm fabs ramp over the next few quarters, they drag TSM margins down; the notes put TSM gross margins at 67% and stock performance at almost -15% over the last month.
ASML is expanding production capacity by about 30% per year while demand is still growing faster, so shortages let ASML set the price; ASML stock is down 10% over the last month.
LRCX sells etch, deposition, and wafer-cleaning machines used by TSMC, Intel, MU, Samsung, and SK Hynix; if those fabs expand slower, LRCX feels it first, and the stock is down 25% over the last month.
KLAC has over 50% share of the overall semiconductor process-control and inspection market and over 80% in optical wafer inspection, with AMAT at 10%; KLAC's advanced-packaging inspection business is expected to hit about $1 billion this year, up more than 50% YoY.
VRT is positioned around NVDA's power architecture for Kyber and Rubin Ultra, with an 800-volt DC portfolio rolling out before those systems; if Kyber slips, VRT would feel it too, and the stock could be down almost 20% over the last month.
The closing argument is that ASML can only ship dozens of chip-making machines a year, TSMC packaging is full into 2027, and NVDA's next-gen racks may be delayed, creating a three-sided bottleneck that the market is only starting to price in.