Two Sides of One AI Trade, Opposite VerdictsThe Hyperscaler Side Just Cleared the Capex Bar The desk had one question for Microsoft, Amazon, and Alphabet this quarter. Whether the trillion-plus dollars they are spending to build AI infrastructure was turning into revenue instead of running through the P&L. The answers came in three consecutive sessions. Microsoft printed Azure at 43% for the quarter, above the 39 to 40% guide and above the 36% desk bogey. Q1 FY27 guidance moved to 45% constant currency. Commercial RPO reached $627 billion. Copilot paid seats crossed 30 million. The stock added roughly $450 billion in a single day, the largest one-day gain by any stock ever recorded. Amazon followed with AWS at 28%, the fastest pace in fifteen quarters. Operating margin at AWS reached 37.7%. The custom-chip franchise crossed a $20 billion revenue run rate. Shares rallied 15.63% Friday. Alphabet added 7.12% Friday, folding into the cloud winners column with the other two. All three raised or maintained aggressive capex guidance for 2026 and beyond. The Chip Side Just Fell Behind It The chip complex went the other way, hard. The Philadelphia Semiconductor Index closed the week in bear-market territory, more than 20% off its record high. Global memory took the brunt. In Korea, SK Hynix fell 14.65% in a single session. Samsung dropped 13%. Kioxia in Japan lost 18%. Sandisk fell 14%. Western Digital 7%. Micron and AMD both fell 8%. Dell, the largest server buyer of that memory, dropped 8%. Intel closed down 6%. The Kospi is 30% off its recent high inside just 25 trading days. The reason is not demand. JPMorgan estimates AI infrastructure spending will approach $870 billion in 2026, up 77% year over year. The hyperscalers alone account for around $750 billion of that. The demand is there. The question the tape is asking is who captures the margin. The Custom Silicon Wedge The reason the demand story does not translate into revenue for the traditional chipmakers is that the hyperscalers are increasingly building around them. Amazon’s Trainium, Graviton, and Nitro franchise crossed a $20 billion revenue run rate this quarter. That is silicon Amazon designed, contracted to a foundry, and put into AWS instead of buying from Nvidia and AMD. Microsoft has Cobalt and Maia. Alphabet runs the seventh generation of its TPU line. Every hyperscaler-designed chip that displaces a name-brand purchase changes who captures the AI-capex dollar. The $870 billion still gets spent. It just gets spent on foundries, utility bills, and internal silicon teams instead of on Nvidia’s price list. Our view: the market is discounting a future in which the AI trade is more about who owns the workload than who makes the chip. The Moonshot Signal The sell-off had a second trigger. Chinese startup Moonshot released an open-weight AI model last week performing on par with leading systems from OpenAI and Anthropic. The signal to the market was specific: the pricing power of proprietary frontier models is contestable. In this tape, that reads through to chip demand in a particular way. If open-weight models keep the cost per inference falling, the same revenue takes fewer chips to produce. If frontier models get commoditized, the premium Nvidia has been charging compresses. The chip complex traded that reading through the week. Worth Watching Monday opens with a live tariff schedule for the first time on a full trading session. Watch whether the 30-year Treasury holds 5.25%, the level it printed Friday and its highest since 2007, into the new regime. The July non-farm payrolls report lands Friday Aug 7. Nvidia reports August 27, the last major AI verdict of the summer. Stock closed roughly flat this week despite the chip complex getting routed. The tape has kept Nvidia in a category by itself. Whether that category holds through late August is the deciding question. Two Mag 7 names remain outside this week’s vote: Nvidia and Tesla. Everything else has already voted. | 
