A major global bank, HSBC, has revised its forecast for NVIDIA's CY2025 datacenter GPU revenue due to concerns about a slower ramp-up of its Blackwell architecture GPUs and the impact of new US export controls on AI chips. This comes amid reports of overheating and glitching issues with NVIDIA's new Blackwell AI servers, further complicating the picture for the chipmaker.
HSBC Reduces NVIDIA's Revenue Outlook
HSBC analyst Frank Lee expects NVIDIA to face "increasing pressure" to deliver a stronger ramp-up of its GB300/B300 platform in the second half of FY26 (roughly aligning with CY2025) to offset a potentially weaker first half. Lee now forecasts NVIDIA will sell approximately 35,000 NVL72-equivalent AI server racks in CY2025, down from a previous estimate of 41,500. This has led to a reduction in the full-year datacenter revenue estimate to $236 billion, down from $253 billion. Further, the bank has lowered its NVIDIA target price from $195 to $185.
HSBC's bear-case scenario estimates sales of between 20,000 and 25,000 NVL racks and lower EPS, although still above consensus estimates.
US Export Controls on AI Chips
The Biden administration has tightened export controls on AI chips, which has significant implications for NVIDIA. The restrictions include:
- No sales restrictions for 18 key allies and partners.
- Cloud giants can buy unlimited GPUs but must keep at least half of their computational power in the US.
- No restrictions on orders with a cumulative power equivalent to 1,700 advanced GPUs.
- "Universal Verified End User" (UVEU) status allows 7% computational power outside the US for specific entities.
- "National Verified End User" (VEU) status unlocks a purchasing power of 320,000 GPUs for eligible entities over two years.
- Non-VEU entities have a 50,000 GPU per country buying power.
- Governments meeting criteria can purchase up to 100,000 GPUs per country.
NVIDIA has criticized these new export controls, stating that they will stifle innovation and harm the US economy.
Blackwell AI Server Issues
NVIDIA's first shipments of Blackwell GB200 AI servers are reportedly experiencing overheating and glitching issues, with the key problem lying in the "way chips connect." This has led to major customers, including Microsoft, Amazon, Google, and Meta, cutting down orders. There are also claims that the issues could be related to TSMC's advanced packaging technology.
The problem has led some customers to switch to older "Hopper" generation GPUs until the issues are resolved.
