Nvidia Blackwell Problem Is Complicated – Scenarios and Implications

Super Micro Computers (SMCI) confirmed the Nvidia’s (NVDA) Blackwell delay rumors during its earnings call saying that no meaningful Blackwell volumes are likely until Q1 2025. Despite this confirmation, and despite the importance of Blackwell to Nvidia, Nvidia’s stock has held up rather well. On investigating the disconnect between the news and the stock reaction, Beyond The Hype has found that most investors and analysts do not see Blackwell delays as a big issue. The main thesis is on the lines that: a) Blackwell delay will only be for a short period of time (three months being the primary sentiment although the rumor was 3+ months), b) the unit and revenue impact will only be modest, and, c) revenue lost from Blackwell delay can be made up by substituting B100/B200 with H100/H200 initially and with B200A when it becomes available.

Given this narrative, it is understandable that the stock response has been relatively mild. But is this assessment true?

Before we get too far into assessing the impact Blackwell delays may have on Nvidia and the industry, first we need to understand the underlying product level issues and interactions. To recap from the previous Nvidia note, most of the leaks so far suggest one or both of the following problems:

1. There is a technical problem with the Blackwell chiplet and/or CoWoS-L packaging that requires a re-spin of the Blackwell chip.

2. There is a problem with the CoWoS-L ramp that delays volume Blackwell ramp to Q2 2025.

Since the leaks surfaced and as a result of them Nvidia is believed to have taken several actions:

Nvidia canceled the Blackwell B100 chip

Since B100 was an interim chip on the way to B200 and a delay makes B100 largely meaningless, the cancellation makes sense. The cancellation may cause Nvidia to discard a certain amount of work in progress, or WIP, and finished goods inventory of B100. So, expect Nvidia to take a small write down on this count. This will also mean that Nvidia will have to place new orders to make up for the loss of supply (most probably H200s) on TSMC (TSM). This is incremental demand for TSMC assuming TSMC has the capacity to entertain new short term orders on N4. 

Nvidia severely reduced the volumes for Blackwell B200 chip and introduced B200A chip

Backwell B200 is a 2-die solution and should be having the problem mentioned in 1 above and little wonder then that it is delayed and seeing reduced volumes. What is more telling is that Nvidia is introducing the B200A chip which is a single die Blackwell solution. The emergence of a single die solution is a tacit admission that there is some problem with the 2-die B200 solution. What is slightly puzzling is that the rumors say that B200 will be available in low volumes. This suggests that the problem could be a yield problem. Depending on exact nature of the problem, there are two potential scenarios for B200 shipments:

• In one scenario, let us call it Scenario 1, the yield problem is so acute that there will be no meaningful production of Blackwell B200. Maybe just samples. None of the rumors or analyst reactions seem to suggest this possibility.

• In the second scenario, let us call it Scenario 2, the yields are low but given the high Nvidia margins, Nvidia will be able to ship B200s to the extent it can absorb the margin hit. There has been no discussion of any margin impact in any of the rumors and analysis so far. Also, the market is completely missing that low yields mean that ramping B200 will be tough with capacity being tight across the board. With tight capacity, Nvidia will have to limit volumes - possibly in a very significant way.

Neither of these are good scenarios for Nvidia. In both scenarios, Nvidia will likely take a write down of inventory and will suffer from lower margins to the extent that B200 ramps.Here is where the single-die B200A becomes the savior. Note that Nvidia desperately needs a single-die variant because a) the company likely has a huge multi-billion dollar die bank of chiplets used to make B200s it needs to use (or take a potentially multi-billion dollar write-off). Nvidia also needs to use this die bank to alleviate the capacity crunch - otherwise the Company needs to build massive amounts of H100/H200s which could be a problem given TSMC is running at 100% utilization. 

In short, the B200A is absolutely needed given the circumstances.Otherwise, Nvidia would have to pretty much write off the entire Blackwell generation. Luckily, as designed, B200A has a more capable compute engine than H100/H200 so it is an improvement over what Nvidia is shipping today. The downside is that since B200A has far less compute power than B200 and supports far less HBM than B200 - let alone AMD competitive offerings (see image below). 

While it will take time for B200A specs and benchmarks to become available, the main purpose of the above table is to show that B200A is a minor upgrade over H200 but is no competition to AMD offerings in inference due to the already known MI300 performance and HBM capacity. Given the 2x memory footprint of MI3x5 compared to B200A, AMD inference solutions can fit large modes for less than half the capex of Nvidia equivalents. For capex constrained environments, this is a difficult proposition to beat.

B200A also poses several logistical problems to Nvidia. The entire Blackwell ramp was planned with infrastructure based on B200. Now, Nvidia will have to create new baseboards and systems with different amounts of HBM memory and IO configurations. This will take time to build, validate, and benchmark, and as a result B200A will not be available to customers until at least Q2 2025.

Given the weak competitive positioning and logistical issues, Nvidia would never have imagined that B200A would have to step in as a major volume SKU of the Blackwell generation. This indicates that B200 problems are severe. B200A is essentially a salvage operation. To the extent that analysts and investors are thinking of B200A as B200, it is not a good take. With B200A as the primary product, Blackwell is now increasingly looking like a half a generation progress than a full generation progress. It may take Nvidia until Blackwell Ultra to have a full generation upgrade over Hopper.