Nvidia is pulling custom silicon into its platform as AI delivery shifts from GPU demand to power, permits and financing. HBM scarcity remains real but China threatens its duration, while Warsh's hike signal collides with fiscal and Hormuz pressure on long yields.
Nvidia's $3.5 billion MediaTek convertible links custom XPUs to NVLink, networking and HBM, while MediaTek targets 15% of an $80 billion AI-chip market next year. X specialists see ecosystem control, but JPMorgan checks suggest MediaTek could take Google TPU units from Broadcom. The profit pool is widening, yet vendor share is becoming less stable.
AI capex remains strategically non-discretionary, but delivery and shareholder economics are splitting. Hyperscalers treat capacity as existential optionality; Nscale raised about $3 billion and Saudi deployments are scaling. Yet 60% of approved 2027 projects cited on CNBC had not broken ground, while high-beta AI portfolios are de-risking. Power, permits, financing structure and margins now determine which demand becomes revenue.
A Reddit post citing TrendForce says Samsung allocated roughly 70% of memory capacity through 2031 to Nvidia, Microsoft and Google. X specialists say HBM4's larger DRAM input and early yields are lifting spot prices and keeping substrates near full utilization. The disagreement is duration: CXMT's HBM3E may first divert wafers and tighten DRAM, but later scale could pressure incumbents.
Markets repriced September-hike odds after Warsh, while Bessent warned against tightening into a supply shock. David Solomon sees higher term premium as structural because fiscal spending, inflation and growth persist; renewed Hormuz attacks and slow Venezuelan output relief reinforce that pressure. A hike may flatten the curve without delivering durable relief at the long end.
Public preview titles disclose re-entry into a leveraged silver-miner ETF and a US ETF; the instruments and sizing were not public.
BUYING BACK OUR LEVERAGED SILVER MINER ETF
BUYING BACK OUR LEVERAGED SILVER MINER ETF
Myles said he would probably add after the pullback, citing a single-digit P/E and a Swiss mountain-railway monopoly; the disclosure is intent until a purchase is confirmed.
One analyst argues customers now prepay 45-55% of capex; a high-ranked skeptic warns that four-year forecasts in crowded stories often fail. Track cash funding and delivered capacity rather than treating projections as positions.
An investor disclosed an existing BE position as a counterweight to NBIS grid risk. The signal is portfolio construction around power bottlenecks, not a new purchase.
BlackPantherCap sold PNG.V and SIVE at losses, trimmed OUST and raised cash to 14% as yields and oil rose, while retaining higher-conviction names.
AltayCap has a tiny TTWO short, arguing that blockbuster launches have historically failed to convert into shareholder profits.
Public preview titles call for taking PayPal profits and report FIG and VZLA stopped out; no rationale or price levels were public.