No lawsuit filing today, so I'll read the mispriced signal in Nadella's warning and the SK Hynix follow-through.

Satya Nadella's warning to enterprise customers this week names a specific failure mode: companies bolting AI agents onto workflows without redesigning the underlying process, then blaming the model when output quality drops. That is not a caution about capability, it is a caution about deployment discipline, and it lands four days after Fidji Simo's exit left OpenAI's enterprise contract relationships (the Microsoft 365 Copilot pipeline chief among them) without a named owner. Nadella runs the distribution layer GPT-5.6 ships through. When the company that owns the interface between the model and 400 million paid Microsoft 365 seats says the failure is on the deployment side, that is a Redmond product-org read, not a model-capability critique, and it is the read that determines whether GPT-5.6's default-model status inside Copilot survives the next contract renewal cycle.

SK Hynix priced its $26.5 billion Nasdaq listing this week on the same premise Vincent Lai's Bond Connect reporting keeps surfacing from the other side of the strait: HBM3E memory, not GPU allocation, is the binding constraint on how fast Nvidia's Blackwell-generation racks actually ship. The IPO prospectus commits SK Hynix to new fabrication capacity, but fabs built in Icheon or Cheongju take 18 to 24 months to reach yield, not quarters. Until that capacity lands, every hyperscaler placing B200 NVL72 orders, Microsoft's Azure buildout for OpenAI included, is bottlenecked by Korean memory supply before it is bottlenecked by Taiwan Semiconductor's CoWoS packaging line in Hsinchu. That is the queue Nadella's enterprise customers are standing in without knowing it.