Jahanara Nissar

NVDA stock traded up ~20% since early October when the CEO started a media/Street campaign to message that Blackwell demand is ‘insane’ and that ‘everything’s on track’. With the stock close to its all-time high, we detect investor nervousness given that the valuation is leveraged to Blackwell, a future product cycle that has had trouble getting off the starting block due to well-known heating issues. We will wait for an update from management regarding the progress with the fix it claimed to have accomplished.  

Having said that, the fact remains that NVDA is the only game in town. Going into the recent AMD AI event, we said we see little out there which makes us believe AMD could make a dent (link). At hyperscale players, we do not expect their internal silicon to make a dent for the next couple of years. An exception is Google’s TPU program. Even though Google Gemini could be a price spoiler, Google’s AI strategy is yet to emerge. Nevertheless, inference token cost seems to be coming down on its accord, without much pressure from Google, and with it, the pricing of Hopper GPUs.

NVDA management is tasked with having to convince investors that it has visibility for Blackwell demand at least into 1FH26. Colorful phrases may not be enough. Phrases such as ‘demand is outstripping supply’ may not mean much if supply is constrained due to yield issues. Are customers willing to plunk down ~$1bn-$2bn in upfront hardware cost for a 20K GPU datacenter without first gaining conviction that the problems associated with the initial delay in launch have been resolved? We do not think so.

Then there is the matter of demand for Hopper. At the previous earnings call management set the expectation that Hopper growth was to continue, even as Blackwell ramped in 4FQ25. Going into this call, there is some worry on the Street that Hopper growth is in question. Falling price of Hopper server rental and token cost are now being noted on the Street. We focused on this issue three months ago in our July quarter preview (link). Besides, could it be possible that the giant miss of print/guide at SMCI was in part related to falling demand for Hopper? We think so.

Going into the previous earnings event, with the stock trading at an all-time high, we suggested that investors may step aside and look for a better point of entry. After the earnings call, the stock dropped ~20% before finding its footing. We are cautious going into this earnings event as well – step aside and look for a better entry point.

As the Gen AI cycle matures, the initial excitement wears off, and hyperscale customers put more focus on ROI metrics, investors need to anticipate growth rates inflecting down. Even a year from now, consensus estimate for data center revenue growth in Oct 2025 is 50% y/y. Is that realistic? We do not think so.

–xx–

Blackwell’s lead customer – NVDA itself? According to AWS blog and media reports, the Ceiba project at AWS is a joint development effort with NVDA. The project is based on a purpose-built liquid-cooled GPU datacenter large enough to house 20K GPUs and is expected to be among the first large customers of Blackwell GB200. NVDA’s Blackwell revenue in the Jan quarter we suspect will largely be derived from AWS/Ceiba. We believe Foxconn is the ODM chosen for this project. We expect the 20K shipment to complete in 4FQ/1FQ.

The anchor customer of the data center may be NVDA itself! We believe the shell construction may be externally financed. The hardware infrastructure too is probably financed by third parties. As such, we believe AWS’s capex exposure and risk of underutilization could be minimal. We believe subsequent customers of Blackwell could hold off orders for Blackwell until NVDA delivers performance metrics from the AWS/Ceiba data center.

When does Blackwell order from MSFT kick in? Microsoft runs some of the largest LLMs in the nascent AI industry. We expect Microsoft to be a key customer of NVDA’s Blackwell, given the inferencing needs of GPT-4 class of models and the training/fine-tuning needs of Orion class of models. As such it is surprising to us that MSFT does not appear to be the lead customer of Blackwell. We wonder if MSFT is waiting for performance data from AWS/Ceiba before committing to Blackwell for its data centers.

Sovereign AI? Beyond AWS/Ceiba delivery in 4FQ/1FQ, and perhaps Microsoft in 1FH26, it is not clear to us where would Blackwell go. NVDA management often points to potential demand from sovereign customers – India, the Middle East etc. The timing of delivery is not clear to us. And are these even credible customers? Their business cases are less than obvious, the timing of orders tied to political dynamics, especially as the US government gets more stringent with export rules.

SMCI earnings disappointment – a view into Hopper demand? At its recent earnings call SMCI provided preliminary Sept quarter revenue at ~$6bn vs. the original guidance of $6bn-$7bn, a $0.5bn miss at midpoint. The company guided December quarter revenue to a $5.5bn-$6.1bn vs. consensus of $6.9bn, a $1.0+bn miss at midpoint.

Management explained their giant miss as due to non-availability of Blackwell. However, this explanation is inconsistent with management expectations set three months ago at the June quarter earnings call. At the June quarter event Management explicitly stated that it expected no revenue from Blackwell in the September quarter and very little in the December quarter. It seems to us that management’s explanation set forth at the September quarter earnings call lacks credibility.

Investors seem to take the view that the miss to both quarters was due to end customers re-directing orders away from a company that appeared to be on its way to getting delisted. We can understand customers trimming order pipeline in out-quarters, say in March/June quarters. But why would customers cancel firm orders for delivery in September and December, as the company appears to be well capitalized to handle working capital requirements. If the demand for Hopper is as strong as NVDA management says it is, why would data center customers cancel orders?

Our view: While there may be many factors that may have gone into the mix, it seems to us that softening in demand for Hopper could be a key reason for the SMCI miss, either due to lack of end demand from customers of data centers, or due to non-availability of liquid-cooled data center shell space. Reduced demand for Hopper at SMCI naturally implies reduced demand for Hopper and networking gear at NVDA.

With Blackwell expected by NVDA management to amount to ‘several billions’ in its Jan quarter, the overwhelming majority of the ~$37bn worth of Jan quarter revenue consensus expectation comes from Hopper. Any softening of Hopper revenue could have meaningful impact to NVDA’s Jan quarter guidance.

Demand centers in Asia: In the 6-month period ending in July 2024, NVDA reported the revenue exposure to Taiwan, Singapore and China geographies at 46%, just as large as the exposure to the US. However, Taiwan and Singapore do not have giant data centers. Are GPUs shipped to these two countries finding a home in Europe? If not where? With the US government clamping down on shipment to China via third nations, sales into Asia geography, especially Singapore (11% revenue exposure) could come under regulatory scrutiny.

Taiwan ODMs may absorb surplus H100: Media reports of H100 being made available to Tier 2 ODMs such as Asus, ASRock and Gigabyte implies demand shortfall in the US vs expanded supply, in our view. We are not clear who could be the end customers of these ODMs. We do not think they are supplying customers in the US or Europe. We are not aware of any meaningful data centers in Asia outside of China.

Outside of Foxconn, none of the Taiwan ODMs have expressed an expertise in liquid-cooling technology. Given the advantages posed by liquid cooled racks, going forward why would data centers invest in air-cooled GPU racks? It appears to us that Tier 2 ODMs may simply act as holders of channel inventory, rather than active suppliers into working data centers. As such, sale into Tier 2 ODMs could obfuscate true end demand.

Net/Net: We are cautious into print due to our view 1) demand for air-cooled H100 could be softening, 2) the uncertainty surrounding the performance of Blackwell could put the brakes on demand after the initial rush of orders and 3) limited availability of liquid-cooled data centers could crimp GPU demand.  

On a longer-term basis, as the Gen AI cycle matures, the initial excitement wears off, and hyperscale customers put more focus on ROI metrics, investors need to anticipate growth rates inflecting down. Even a year out into Oct 2025, consensus estimate for data center revenue growth stands at 50% y/y growth. Is that realistic? We do not think so, if not for any other reason but for the well-known issues with power/water considerations. We think growth expectations must reset for the stock to find a firmer footing.

KC Rajkumar