AI borrowing is pushing long yields toward equity-breaking levels even as component shortages spread beyond GPUs. China’s tech engine is outrunning domestic demand, while memory pricing stays tight but faces a 2026 supply test.
AI demand is still strong, but the way it is being financed is becoming a market constraint. Hyperscalers are issuing debt just as Treasury supply and weak foreign demand push long yields higher. Earnings are buffering equities for now; a 10-year yield near 5% would turn the funding boom into a valuation test.
HBM, NAND and AI-server memory remain tight, with contract floors, customer requests above supply and spot prices rising. The caveat is that AI-specific shortages coexist with price-sensitive legacy demand and new capacity expected in 2026. The cycle can stay constructive, but allocation discipline and price trends matter more than a blanket memory call.
Fabrinet’s results and lengthening MLCC lead times show AI infrastructure demand spreading into optical transport, custom silicon, networking and power components. Yet Fabrinet produced only $4 million of free cash flow despite record earnings. The durable beneficiaries will be the suppliers that convert backlogs into cash without overbuilding inventory or capacity.
China’s high-tech manufacturing push is working faster than its domestic economy. AI and semiconductor activity remain strong, but property, retail and consumer hardware are weak; Baidu has logged another revenue decline and Xiaomi’s profits fell. Policy support may lift activity, but a narrow tech boom is not yet a broad earnings recovery.
The author disclosed adding EnerSys to the portfolio, arguing that its data-center and defense power exposure trades well below electrical-equipment peers.
AltayCap disclosed long positions in Katakura and Inui Logistics while describing TOC as the favorite among inexpensive Japanese real-estate names.
Semicon_player explicitly disclosed a Cerebras long; separate evidence in the same window points to paid OpenAI inference as the commercialization checkpoint.
Serenity said NAND, MLCC and InP constraints do not disappear when related shares fall, but the captured call is a watch on Sandisk and AXT rather than a disclosed buy.
Vineland’s planning approval reduces local land-use risk, while the remaining test is conversion of installed capacity into customer-accepted, billable GPU service.
Airbnb’s customer-service automation lowered cost per booking, yet the cited valuation was considered too high to chase without a pullback.
Shkreli said he held a very small Applied Optoelectronics short and intended to increase it, directly opposing the bullish capacity-expansion thesis elsewhere on X.
Accounting_ds disclosed trimming AirJoule from 13,000 to 4,000 shares to raise cash and reduce concentration in a frothy market while retaining a smaller position.
Jim Cramer treated strength in ExlService and MakeMyTrip as an opportunity to reduce exposure because AI-displacement concerns and valuation could persist despite sound current operations.