Artificial Intelligence

Industry reports from ZDNet Korea and DIGITIMES indicate Samsung is considering a dual-sourcing model for custom HBM base dies, pairing its own DRAM core dies with base dies manufactured at either Samsung Foundry or TSMC depending on customer requirements, with NVIDIA's NVHBM specification cited as the likely application starting at HBM4E. Samsung's official record confirms only that its internal 4nm foundry process is currently producing base dies for HBM4 and HBM4E, contributing measurably to Foundry earnings in Q2 2026. The TSMC arrangement remains unconfirmed by either company, with no process node, contract terms, production volumes, or qualification status established in the available evidence.

IREN's 2 GW Sweetwater Hub in Texas received conditional inclusion in ERCOT's Batch Zero interconnection study process on September 8, 2026, a procedural entry point that requires further verification steps before any final grid approval is granted. Both Sweetwater 1 and Sweetwater 2 remain under construction, with the first 300 MW phase of data center capacity targeted for delivery in Q4 2027 and Sweetwater 2's substation energization not expected until 2027. No public ERCOT records confirm completed verification, final classification, or actual power delivery at either site.

QumulusAI completed the customer handoff of NVIDIA B200 GPU capacity on September 8, 2026, under a three-year agreement carrying approximately $72.0 million in aggregate contracted fees, with the NVIDIA B300 portion of the same contract still in final deployment. The $72.0 million figure represents a multiyear committed value, not recognized revenue, and no contract-specific recognition schedule has been disclosed publicly. Key details including the exact GPU count, operating site, and customer identity remain undisclosed across company releases and SEC filings reviewed through September 16.