Nvidia: Hidden Risks Loom

At first glance, NVDA is the big winner in AI. AI needs specialized processors for training and inference. NVDA dominates this market with its GPUs and has a significant competitive barrier in its nearly universally adopted CUDA software rubric. Its revenue more than doubled in 2023 and its market cap nearly tripled. Consensus expects sales to rise 56% and EPS 79% in FY25. However, the sunny narrative masks its vulnerability to its increasingly concentrated customer base. AMZN, MSFT, GOOGL, and META account for as much as 2/3rds of the demand for NVDA’s AI-tuned GPUs and given the rise of foundation models on these platforms, we believe the concentration is likely to increase. Historically, these companies’ capex has been very lumpy and in an environment of shortage, double ordering is a real possibility, raising the threat of a temporary downturn in spending, or even a prolonged inventory glut in 2024. More concerning is research initiatives by each of these customers to design proprietary ASIC processors for their internal AI needs, following GOOGL’s lead (it is on its 5th generation self-designed TPU). While we are very skeptical that new products from NVDA rivals like AMD or INTC could displace the H100 and its successors, ASICs could give hyperscale platform operators much more control over the economics of their burgeoning foundation model franchises. This could crowd NVDA out of the biggest part of the future AI processing market. In this context, we believe the risks for investors now outweigh the potential rewards.

NVDA dominates AI processing. We believe that AI is prompting a profound shift in computing that will enhance almost all existing applications and enable wholly new ones. In turn, this is driving a new datacenter architecture that has thus far been centered on GPUs from NVDA. While rivals like AMD and INTC, and numerous startups aim to displace the incumbent, almost all AI engineers have been trained on NVDA’s CUDA software rubric, using it as the bridge between their models and the underlying hardware. This is an extremely powerful competitive barrier, as retraining engineers on an alternative rubric is both expensive and time consuming.

NVDA’s valuation depends on the dominance narrative. NVDA trades at 74x trailing earnings and 31x trailing sales, with a $1.4T market cap. Of course, the bull narrative anticipates 57% sales growth and 70% EPS growth for FY25, making the forward P/E a much more palatable 28.3x and the PEG ratio on 5-year expectations of 38% sales growth a paltry 0.76. On this basis, NVDA looks cheap, but any crack in the presumption of long-term dominance and sustained growth could bring a painful rerating.

AI evolving to a highly concentrated market. AI is coalescing around a “foundation model” paradigm under which applications and specialized AI models are built atop massive, pretrained, multi-modal, self-reinforced, and highly adaptable programs. While dozens of these are now available from both established players and startups, the costs and challenges of keeping these models at the state-of-art are already separating a few leaders from the also-rans. In the end, we believe that a handful of foundation models (i.e., MSFT/OpenAI, GOOGL, AMZN, META) will dominate, likely integrated with a hyperscale datacenter platform, itself a market that is already highly concentrated.

Customer concentration creates risks for NVDA. Analysts estimate that more than half of the demand for NVDA’s AI-tuned A100 GPU and an even higher percentage of the just introduced $30K+ flagship H100 comes from just 4 customers, AMZN, MSFT, GOOGL, and META. As the foundation model paradigm advances, we believe that this concentration will increase with time, giving these companies considerable future leverage. We believe that this creates risks for NVDA that investors may not fully appreciate, calling into question the dominance narrative.

Lumpy CAPEX will make demand less predictable. Historically, the biggest datacenter platforms have been inconsistent in their capital spending patterns, often showing big YoY spending growth in one quarter only to follow with flat or even declining spending in the next. With the majority of A100 and H100 demand coming from just four of them, the chances that spending spikes and troughs coincide are higher, yielding unpredictable fluctuations in demand and thus, both upside and downside surprises. A negative quarterly surprise caused by this could introduce doubt to the domination narrative and trigger a sell-off.

GPU shortages could yield an inventory glut. A100 GPUs have been in short supply for over a year. Historically, component shortages have prompted customers to order more than they need in anticipation of receiving less than they had ordered. However, once supply constraints ease, these customers have found themselves with excess inventory, resulting in a sharp reduction in orders until the excess is worked off. NVDA experienced an inventory glut in 2019, when crypto-miners began hoarding its graphics cards only to abruptly stop ordering as crypto-coin prices dropped, resulting in 4 quarters of declining YoY sales. We believe the risk of a similar predicament is exacerbated by the extreme concentration of customers for AI-tuned GPUs.

Huge platforms have the wherewithal and incentive to self-design ASIC processors. With the price of NVDA A100 GPUs at more than $10K and its new H100 at more than $30K, hyperscale platforms which employ 10’s of thousands of these chips have incentive to consider lower cost options. While investors have focused on rival GPU offerings from AMD, INTC, and others, we believe the real threat is from self-designed ASICs. GOOGL has been training and running their internal AIs on its proprietary TPU ASIC family, now on its 5th generation. AMZN, MSFT, and META have all expressed intention of following GOOGL’s lead with their own self-designed processors. We believe that these companies are likely to dominate future AI with their foundation models, potentially locking out NVDA from the largest part of the market.

What is it worth? NVDA is now a fast-growing and hugely profitable company. Consensus expects it to sustain better than 30% annual sales growth and improve its profitability over the next 5 years. If investor confidence in that narrative is shaken, we believe shares could drop by a third or more, in line with the 20-24x P/S ratios of fast-growing cybersecurity names like CRWD or NET. We believe the risk of a quarterly miss, soft guidance, or discouraging news flow at some time this year outweighs the potential upside from here.