Where could an AI deployment get delayed? Trace documented manufacturing and cooling dependencies, then ask what capacity is actually ready.
Sources reviewed: · Statements below retain the dates and scope of their sources.
These are dependencies to investigate, not a ranking of current shortages. Reference designs describe how a site can be built; they do not show a site's operating capacity, signed orders or supplier revenue.
Check foundry disclosure and qualified packaging capacity
Separate component shipments from completed racks
Verify site readiness and commissioning
Conceptual deployment dependencies. Arrows do not assert supplier contracts or measured capacity constraints.
NVIDIA names TSMC and Samsung as wafer foundries and discloses use of CoWoS packaging technology.
Source limitations: The filing does not allocate CoWoS volumes or establish a current shortage at a particular supplier.
What to check next: Look for later capacity, qualification and delivery disclosures for the relevant product.
Source published:
NVIDIA · FY2026 Form 10-K ↗
Micron reported Q1 2026 volume shipments of 36GB 12-high HBM4 designed for NVIDIA Vera Rubin.
Source limitations: Component shipments do not establish complete-system availability or NVIDIA purchase shares.
What to check next: Compare memory shipment evidence with the system builder's delivery and operator's deployment evidence.
Source published:
Micron · HBM4 volume production for Vera Rubin ↗
Vertiv and NVIDIA released a 7MW GB200 NVL72 reference architecture, combining liquid and air cooling for racks up to 132kW.
Source limitations: 7MW describes the design, not installed customer capacity. It is not evidence of a purchase order.
What to check next: Check the operator's named site, power availability, cooling installation and commissioning milestones.
Source published:
Vertiv · GB200 NVL72 power and cooling reference architecture ↗
Schneider and NVIDIA described GB300 NVL72 infrastructure and controls designs, including Motivair cooling technology and a design for up to 142kW per rack.
Source limitations: A reference design is not a completed deployment or a disclosed NVIDIA supply contract.
What to check next: Seek a specific customer deployment announcement before assigning installed capacity or revenue to this design.
Source published:
Schneider Electric · GB300 power and liquid cooling designs ↗
