NVIDIA

Nvidia formally announced the acquisition of Hugging Face, a large language model platform, for $12.9 billion. Nvidia intends to support and scale up Hugging Face to continue to serve LLMs, including open-source models.

Why it matters: Financially, we view the deal as immaterial to Nvidia’s future results. Strategically, we think the deal improves Nvidia’s position within the open-source community, not only to serve its own model, Nemotron, but to expand the usage of open-source models more broadly. Like Nvidia, we see a world where open- and closed-source models co-exist. We anticipate that open-source will be used to cost-efficiently handle heavy, monotonous workloads, while closed-source (like, say, Claude) will offer premium tokens for mission-critical workloads and knowledge. It seems unlikely to us that any LLM builders will move off Hugging Face now that it is owned by a large company (perhaps with a conflict of interest as Nvidia has an LLM), but the risk is plausible if Nvidia were to unthinkingly tilt the scales toward itself.

The bottom line: We maintain our $310 per share fair value estimate for wide-moat Nvidia. It's hard to gauge the stock move related to this news, since shares remain in the afterglow of Nvidia's stellar earnings report last week. Shares remain undervalued as the durability of artificial intelligence growth appears to be underestimated. Overall, Nvidia remains focused on building AI systems and expanding the AI ecosystem. We think this is a wise offensive move but also defensive. If a world exists where Anthropic and OpenAI both shift to using more in-house chips, Nvidia may emerge with more competitive models. It seems unlikely to us that any LLM builders will move off Hugging Face under Nvidia's ownership, even though Nvidia may have a mild conflict of interest by building Nemotron. We don’t expect Nvidia to tilt the scale toward itself in any meaningful way.

Per Nvidia, more than 18 million developers, researchers, and creators use Hugging Face to share more than 3 million models, 500,000 datasets and 1 million applications. More than 200,000 companies use the platform to discover, evaluate, customize, and deploy AI.

Our fair value estimate is $310 per share. Our fair value estimate implies a fiscal 2027 (ending January 2027 or effectively calendar 2026) and fiscal 2028 price/adjusted earnings multiple of 33 times and 19 times, respectively.

Nvidia’s data center business has achieved historic growth from $3 billion in fiscal 2020 to $194 billion in fiscal 2026 and we estimate it will be $385 billion in fiscal 2027, representing 99% annual growth. We were amazed with Nvidia’s forecast for 70%-plus growth in fiscal 2028 (or effectively calendar 2027), but we think it’s achievable and we model $672 billion of data center revenue the following year. This estimate also does not include sales into China, as the Chinese government is dissuading its local champions to use Nvidia gear, and we no longer model revenue from China either.

In the medium term, we model 15%, 15%, and 8% in data center revenue growth in fiscal 2029, 2030, and 2031, respectively, to over $960 billion in fiscal 2031 (which is effectively calendar 2030) and $1.01 trillion of total revenue when including Nvidia’s edge computing segment. The main driver of this tremendous growth is an ongoing increase in capital expenditures in data centers at leading cloud computing, enterprise, and sovereign government customers. We think Nvidia is earning about $40 billion per gigawatt of AI data centers being built out today. We agree with Nvidia’s estimate that its market opportunity per GW could reach $80 billion-$100 billion by 2030, as we think over 100 GW of AI data centers might be built out globally by 2030.

We think it is reasonable that Nvidia may face an inventory correction or a pause in AI demand at some point in the medium term thereafter, so we model a flat revenue year in fiscal 2032. We anticipate average annual DC growth in the 10% range thereafter as AI matures. In the long run, we think that cloud computing revenue at the hyperscalers can grow at a low-teens rate (if not mid-teens), capital expenditures as a percentage of revenue remain at consistent levels at these hyperscalers, and thus we model Nvidia’s revenue growth to be on par with these cloud computing growth rates.

In terms of total revenue (including edge products), we model Nvidia growing at CAGRs of 36% and 21% over the next five and 10 years, respectively.

Nvidia’s massive DC growth has been gross margin-accretive, as we think Nvidia should achieve mid-70% gross margins in fiscal 2027. In the long run, we anticipate modest gross margin deterioration in the decade ahead, as we lower it to the high-60% range a decade from now. Still, we are optimistic about Nvidia’s ability to retain its pricing power in DC products, thanks to the high switching costs associated with the Cuda platform and Nvidia’s excellent suite of interconnectivity products.

These high gross margins translate to stellar GAAP operating margins. Looking ahead, we think GAAP operating margins will hover in the high 50% range to the mid-60% range in each year of our 10-year forecast period, depending on the pace of R&D growth.

We assign Nvidia a wide economic moat rating. We believe Nvidia benefits from intangible assets around its graphics processing units and its networking and interconnectivity gear. Nvidia also maintains strong pricing power via high customer switching costs around its proprietary software, Cuda, for AI tools, which enables developers to use Nvidia’s GPUs to build AI models.

Nvidia was an early leader and designer of GPUs, which were originally developed to offload graphic processing tasks on PCs and gaming consoles, but are not critical components in AI. We attribute at least a portion of Nvidia’s AI leadership to intangible assets associated with GPU design. GPUs perform parallel processing, in contrast to the serial processing of 0s and 1s performed by central processing units used to run the software and applications on PCs, smartphones, and many other types of devices (like Intel and Apple processors).

Parallel processing does not need to run in a linear order. This was originally useful when displaying images on a PC, as GPUs would run simple processing (such as displaying a pixel) but on many more cores to display realistic images in PC games. In AI, these favorable characteristics of GPUs (many cores, simple calculations, not necessarily in order) became useful for matrix multiplication used in large language models. Effectively, these GPUs calculate the tens of thousands of scores and weights used to determine the next token to be provided in an AI query, known as inference.

Nvidia’s GPU was present at the dawn of the AI era, not only because of chip design expertise, but also because its GPUs could be programmed via its proprietary software platform, Cuda. Since 2012, Nvidia has made shrewd moves to build and expand Cuda for AI, creating and hosting a variety of libraries, compilers, frameworks, and development tools that allowed AI professionals to build their models. This prescient work has given Nvidia a technological advantage in AI over its peers.

If Nvidia’s GPU design expertise and Cuda were not enough, we also think Nvidia has carved out a wide moat via networking and interconnectivity product expertise. While many analysts think of AI as running on GPUs, the more important aspect is having clusters of GPUs “talk” to one another to process larger and larger workloads, especially in AI training. With this in mind, Nvidia’s interconnections of GPUs via NVLink, and its high-end networking gear in InfiniBand and Ethernet, enable Nvidia’s GPUs to work relatively seamlessly with one another. Even if an external vendor were to develop an AI accelerator on par with Nvidia, as well as catch up to Cuda and develop a strong software platform, Nvidia may still have an advantage over rivals by connecting GPUs and building AI rack solutions more seamlessly than peers.

Looking at the competitive landscape, AMD is a well-capitalized chipmaker with GPU expertise, although we view the company as being in a position of weakness on the software front. Perhaps the biggest threat might be from in-house chip solutions from hyperscalers, such as Google’s tensor processing units, or TPUs, and Amazon’s Trainium chips. It’s possible that each of these in-house chips might perform specific workloads better than a general AI GPU from Nvidia or others. If competitors can develop gear that is, say, half as performant as Nvidia but at a quarter of the price, customers can throw more of these solutions at their AI problems and perhaps match, if not exceed, Nvidia’s performance.

However, we believe that cloud computing companies will have to offer their enterprise customers a full menu of GPUs and accelerators so that they can run AI workloads. Enterprises are typically loath to be locked into a single vendor and might not put 100% of their AI fortunes into a single in-house chip.

We also anticipate that AI GPU flexibility will be a requirement for AI customers over time. In-house chips might run certain AI tasks even better than Nvidia, but as techniques change in the AI industry (and they seem to be changing weekly), Nvidia’s GPU programmability offers flexibility that in-house chips can’t match right away.

Further, we now view Nvidia as an “AI systems” company. The firm not only sells AI GPUs and networking gear, but CPUs, storage solutions, software, and open source models that can enable any business to pick and choose components to deploy AI. We think Nvidia’s ultimate goal is to expand AI use cases and the AI ecosystem globally (as evidenced by all of Nvidia’s investments in partners and startups). Even if large tech companies develop enough in-house expertise to reduce their reliance on Nvidia, there are a host of smaller cloud vendors (neoclouds), governments, and enterprises that will remain reliant on Nvidia’s full stack. For all of these reasons, we foresee Nvidia remaining at the head of the pack in AI for quite some time.

The AI infrastructure opportunity is massive, and Nvidia foresees $3 trillion-$4 trillion of annual AI infrastructure spending by 2030.

Nvidia’s data center GPUs and Cuda software platform have established the company as the dominant vendor for AI model training and inference.

Nvidia is expanding nicely within AI, not just supplying industry-leading GPUs but also moving into networking, software, and services to tie these GPUs into even more-powerful clusters.

Nvidia’s customers are a handful of the largest Tech companies in the world, and they all have an incentive to eventually diversify away from Nvidia to some extent.

AI infrastructure spending has been impressive but revenue and use cases are less certain, perhaps providing doubts that there is a good return on investment on AI that might lead to a spending downturn at some point in the future.

Geopolitics have entered the AI space, most notably limiting Nvidia's AI opportunities in China.