AI infrastructure demand remains strong despite a tripling in data-center GPU supply, with Oracle saying all GPU capacity up for renewal was re-contracted or resold at prices 20% above previous agreements. Most of the equipment had been operating for more than four years. Serenity said the pricing supports Nebius and IREN, benefits Nvidia and the broader neocloud computing sector, and counters Michael Burry's argument that GPUs rapidly lose value. Oracle supplied 850 megawatts of capacity, representing more than 300,000 GPUs and three times the prior quarter's volume, while utilization reached 97.9%. Demand is also broadening beyond traditional GPU manufacturers as Meta, Microsoft and OpenAI develop custom ASICs that require increasing amounts of HBM. Meta's MTIA 400 is reported to raise HBM3E capacity from 216GB to 288GB, while AMD's MI455X is expected to use 12-high HBM4 with 432GB. Samsung Electronics and SK Hynix are pursuing more customized HBM supply arrangements, while Micron plans to increase HBM output to approximately 100,000 wafers per month by year-end. Advanced package substrates are developing alongside the memory market, with Samsung Electro-Mechanics and LG Innotek targeting glass-substrate mass production in 2028 and LG Innotek targeting high-value AI accelerator and server FC-BGA production in 2027.