Nvidia CEO sees strong demand for Blackwell chips

Nvidia CEO Jensen Huang said the company is seeing strong demand for its Blackwell chips during an event in Hsinchu, Taiwan.

He noted that Nvidia relies on TSMC to produce the wafers needed for these chips but did not disclose specific order numbers.

Huang added that Nvidia’s success depends on TSMC’s support, marking his fourth trip to Taiwan in 2025.

TSMC CEO C.C. Wei confirmed Nvidia had requested wafers but said details were confidential.

Huang said memory makers SK Hynix, Samsung, and Micron have provided advanced chip samples and increased capacity to support Nvidia.

SK Hynix recently said it sold out all chip production for 2026 and plans to boost investments, while Samsung is in talks to supply its next-generation memory chips.

He added there are currently no talks to sell Blackwell chips to China due to US export restrictions.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

CoWoS, or Chip‑on‑Wafer‑on‑Substrate, is TSMC’s process that connects multiple compute chips and memory on one substrate targeting 90–95k wafers per month by end‑2026 vs 70k at end‑2025 1. Packaging is the pinch, so supply should stay tight through 2025 2. Relief is hoped in 2026 even with wafer output rising 2.
NVIDIA may take ~66% of CoWoS in 2026, and Blackwell’s die is ~2x H100, expanding CoWoS‑L area per device (a large‑format version of CoWoS used for multi‑die packages) 1 3.
High Bandwidth Memory (HBM) adds a chokepoint alongside CoWoS, as SK Hynix said 2026 output is sold out, forecasts put HBM near 2x by 2025 3.

GB200 NVL72 is Nvidia’s rack‑scale system built on the Grace Blackwell “GB200” platform. It draws 120 kW, versus ~12 kW for CPU racks and ~40 kW for air‑cooled H100 racks 4.
Colocation providers (data center landlords serving multiple customers) and hyperscale operators (the largest cloud platforms) use liquid‑to‑air conversion systems, heat exchangers that link a rack’s liquid loop to a facility’s air cooling, shown by AMAX (a server and liquid‑cooling systems integrator) 5.
Cold plates (chip‑level heat spreaders), cooling distribution units (floor‑level modules that circulate coolant), and rack manifolds (in‑rack plumbing that routes coolant) see demand 1 6. Morgan Stanley pegs 2025 at ~30,000 GB200 NVL72 racks 1 6. The modular MGX architecture, Nvidia’s server reference design that lets operators mix and match components, lets operators scale upgrades in steps thereby helping tier‑2 data centers (regional colocation providers and enterprise‑run facilities) compete 7.