Our AI infrastructure map moved in five places today. here’s what changed, and why it matters for investors: #1#: The packaging center of gravity may be shifting. Morgan Stanley estimates Nvidia’s share of TSMC CoWoS allocation falls from 53.4% to 45.6% in 2027, while AMD nearly doubles to 19.8%. Google also reportedly booked Intel packaging capacity for 3M+ TPUs rather than waiting in the TSMC queue. Advanced packaging is becoming a more contested, multi-supplier market. #2#: NAND crossed into critical. Samsung starts mass production from its new storage line in Q4, while expected price increases on the tape are now 35 to 40%. Storage has been the quiet part of this AI cycle. That may be changing quickly. #3#: SK hynix has begun HBM4 mass shipments to Nvidia for its next platform. We now have a live price test using Korean customs data that should resolve within 30 days: - Does HBM pricing accelerate through the generation transition? If not, the tightness story is weaker than consensus assumes. #4#: Optical demand visibility keeps extending. $AAOI and Fabrinet are both describing longer order horizons, verified against their own transcripts. Important distinction: a supplier saying “we expect” is not the same as booked backlog. But the directional signal across the chain remains remarkably consistent. #5#: The contrarian one: co-packaged optics eased from tight to balanced. Everyone watches for bottlenecks getting worse. But loosening matters too. A constraint moving from scarce to available can change margins, bargaining power, capex timing and which suppliers actually capture the economics. It gets far less attention than tightening.