Nvidia CEO says company has enough chips to meet demand

Nvidia CEO Jensen Huang said the company has enough Blackwell chips to meet rising demand, addressing concerns about product availability.

Huang clarified that when he previously described Blackwell chips as “sold out,” he meant that customers are using existing inventory at full capacity.

He added that Nvidia has managed its supply chain to ensure a steady supply of chips.

Nvidia reported strong Q3 results, and said it expects to surpass its long-term US$500 billion sales target for new chips and systems.

Huang said Nvidia is capturing a larger share of data center spending due to added product capabilities, and noted that the upcoming Vera Rubin generation is expected to deliver about US$35 billion in revenue to Nvidia out of the roughly US$55 billion spent per gigawatt of computing power.

He noted that Nvidia’s China business remains limited by US export restrictions, and the company expects no data center chip sales in the country for now.

Huang said Nvidia is in discussions with US and Chinese officials about potentially resuming sales in China.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Blackwell supply claims meet real deployment limits

Huang said there will be “enough supply”, yet supplier signals flag delays for GB200 NVL72 racks (Nvidia’s rack-scale AI system), with Super Micro (server manufacturer) cutting FY25 guidance and Vertiv (data center power and cooling vendor) guiding slower growth 1. The 132 kW thermal load sets a new high for server power, and networking plus power vendors see volume starting mid-2025 1.
System integration is the main constraint 1. GB200 NVL72 needs direct liquid cooling, high-speed networking, and power management integrated circuits (ICs) 1. Semtech (networking and analog semiconductor company) pulled Q1 guidance due to rack design changes, and reports have Nvidia pausing the NVL36x2 dual-rack variant (a two-rack variant of the GB200 design), which may push volume to the second half of 2025 1.
B200 GPUs in standard HGX (Nvidia’s standard server reference design) servers can take some near-term demand 1. The 30X inference gain Nvidia cites comes from the full NVL72 rack, so buyers seeking peak capability may still wait despite Huang’s pledge 1.

Non-GPU spend near US$20B per gigawatt on Rubin builds

For the Vera Rubin platform, Nvidia expects about US$35 billion of roughly US$55 billion per gigawatt of AI capacity (a measure of data center electrical capacity). That leaves about US$20 billion for power, cooling, networking, and integration.
Builders, integrators (firms that assemble and deploy multi-vendor hardware and software), and investors can size opportunities by tracking hyperscale buildouts. Data center builds cost US$8 million to US$12 million per megawatt 2, while cooling uses 30% to 55% of energy 3. GB200 NVL72 racks can cost up to US$3 million, with 35,000 units estimated for 2025 shipments 1. That scale favors power delivery and liquid cooling vendors, including Vertiv, despite near-term Blackwell timing questions 1.