Nvidia's Blackwell Back on Track. Servers Ship in December
[Exclusive] Microsoft, AWS and Oracle among first to receive the new GB200 servers after production delay
Good Evening from Taipei,
Nvidia will start delivering the first racks of its GB200 servers to major cloud-service providers in early December after chip-design & production problems were resolved. The new schedule means shipments will only be about one month later than originally planned, which is faster than feared after the issues came to light.
Microsoft is expected to get one of the largest allocations of Nvidia’s most-advanced AI chips, with Oracle, AWS and Meta among other top clients, my sources tell me. Nvidia hammers out the deals to sell Grace CPUs and Blackwell GPUs directly with the final customer, but the chips are sent straight to suppliers for system integration. (So, technically, it’s the SIs who deliver servers to clients.)
GB200-NVL36 servers were originally slated for delivery at the end of this month (there weren’t many orders for the NLV36, I am told), while GB200-NVL72 versions were scheduled for early November. Delays pushed that timeline back to January. Now, both incarnations will ship around the first week of December. (There’s a ‘72 server that’s really two ‘36 racks connected together.)
Production of Blackwell GPUs was upended a few months ago when Nvidia and TSMC discovered problems related to the CoWoS-L packaging which connects the processor to the high-bandwidth memory. SemiAnalysis has an in-depth rundown of what happened. A key challenge is warpage, a major topic at this month’s SEMICON Taiwan.
The chips were to be delayed three months or more, The Information reported on August 3. As a result, big shipments weren't expected until the first quarter, the news outlet wrote.
But rework of the design and manufacturing have put Nvidia and TSMC back on track. Now, module makers and server system-integrators are already receiving qualification samples, my sources tell me. The chips are fabbed with TSMC’s 4nm process. They’re then combined with HBM in the CoWoS packaging step, which is also done by TSMC but at different facilities.
From there, the chips head to module assembly. Foxconn Industrial Internet (FII) is set to be among the largest suppliers at this step. They’re then sent to server makers, chiefly Foxconn and Quanta, to be integrated into full server racks. Spokespeople for all three Taiwanese companies declined to comment when I spoke to them this week. A spokesperson for Nvidia also declined to comment.
Wistron’s Taipei-based affiliate Wiwynn and China’s Luxshare-ICT are among others hoping to build and sell GB200-based servers. However, since Nvidia decides the allocation of its chips, these assemblers need to win a major client (one which got such allocation) before they can get their hands on the GPUs.
Because of the relatively quick resolution by TSMC and Nvidia of the design & manufacturing problems, FII, Foxconn, and Quanta have been able to build and test systems which will be sent out as GB200-NVL36 (36x Blackwell GPUs apiece) or NVL72 (72x) servers.1
Mass production, from module to system assembly, will be in full swing with significant shipments by early December, sources tell me. Both Ariel boards (1x Grace, 1x Blackwell) and Bianca boards (1x Grace, 2x Blackwell) will be ready. SemiAnalysis explained the various configurations in depth back in July.
Demand remains strong, despite the hiccup, in part because the prior Grace-Hopper project was effectively cancelled — only a small number of enterprise customers received them, I am told.
Initially, assembly of the GB200 servers will be done entirely in Taiwan. Overseas factories, such as in Mexico and Texas, may be added to the production schedule in the first quarter of next year.
That’s all I have for now. Will follow up and share with you if I learn more.
Media, you know the deal: remember to source and link. Thanks
(Got a tip? Contact me. Discretion assured, as long as it checks out)
Yes, I know, there are different versions within these two groups. This is the TL;DR
Nvidia’s quick fix with TSMC on the CoWoS-L packaging keeps the GB200 servers mostly on track, which is critical as AI workloads grow. It’s interesting to see how system integrators like Foxconn and Quanta play a bigger role, and how Nvidia tightly controls chip allocation. Taiwan’s manufacturing still dominates, but with talks of expanding to Mexico and Texas, diversification is inevitable. Given the competition in AI infrastructure and limited chip supply, Nvidia’s control over distribution will be key to meeting demand.
Always appreciate the great insight you’re providing, Tim!